Integrating Physics-Based Modeling With Machine

Learning: A Survey
arXiv:2003.04919v4 [physics.comp-ph] 14 Jul 2020

JARED WILLARD∗ and XIAOWEI JIA∗ , University of Minnesota

SHAOMING XU, University of Minnesota
MICHAEL STEINBACH, University of Minnesota
VIPIN KUMAR, University of Minnesota
There is a growing consensus that solutions to complex science and engineering problems require novel
methodologies that are able to integrate traditional physics-based modeling approaches with state-of-the-art
machine learning (ML) techniques. This paper provides a structured overview of such techniques. Application
areas for which these approaches have been applied are summarized, then classes of methodologies used to
construct physics-guided ML models and hybrid physics-ML frameworks are described. We then provide a
taxonomy of these existing techniques, which uncovers knowledge gaps and potential crossovers of methods
between disciplines that can serve as ideas for future research.
Additional Key Words and Phrases: physics-guided, neural networks, deep learning, physics-informed, theory-
guided, hybrid, knowledge integration
ACM Reference Format:
Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar. 2020. Integrating Physics-Based
Modeling With Machine Learning: A Survey. 1, 1 (July 2020), 34 pages. https://doi.org/10.1145/1122445.1122456

1 INTRODUCTION
Machine learning (ML) models, which have already found tremendous success in commercial
applications, are beginning to play an important role in advancing scientific discovery in domains
traditionally dominated by mechanistic (e.g. first principle) models [22, 112, 117, 126, 127, 138, 214,
261]. The use of ML models is particularly promising in scientific problems involving processes that
are not completely understood, or where it is computationally infeasible to run mechanistic models
at desired resolutions in space and time. However, the application of even the state-of-the-art black
box ML models has often met with limited success in scientific domains due to their large data
requirements, inability to produce physically consistent results, and their lack of generalizability
to out-of-sample scenarios [126]. Given that neither an ML-only nor a scientific knowledge-only
approach can be considered sufficient for complex scientific and engineering applications, the
research community is beginning to explore the continuum between mechanistic and ML models,
where both scientific knowledge and data are integrated in a synergistic manner [2, 14, 126, 204].
This work was supported by NSF grant #1934721 and by DARPA award W911NF-18-1-0027.
∗ Bothauthors contributed equally to this research.

Authors’ addresses: Jared Willard, [email protected]; Xiaowei Jia, [email protected], University of Minnesota, Minneapo-
lis, Minnesota, 55455; Shaoming Xu, [email protected], University of Minnesota, Minneapolis, Minnesota, 55455; Michael
Steinbach, [email protected], University of Minnesota, Minneapolis, Minnesota, 55455; Vipin Kumar, [email protected],
University of Minnesota, Minneapolis, Minnesota, 55455.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from [email protected].
© 2020 Association for Computing Machinery.
XXXX-XXXX/2020/7-ART $15.00
https://doi.org/10.1145/1122445.1122456

, Vol. 1, No. 1, Article . Publication date: July 2020.

2 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

This paradigm is fundamentally different from mainstream practices in the ML community for
making use of domain-specific knowledge, albeit in subservient roles, e.g., feature engineering or
post-processing. In contrast to these practices that can only work with simpler forms of heuristics
and constraints, this survey is focused on those approaches that explore a deeper coupling of ML
methods with scientific knowledge.
Even though the idea of integrating scientific principles and ML models has picked up momentum
just in the last few years [126], there is already a vast amount of work on this topic. This work
is being pursued in diverse disciplines (e.g., earth systems [214], climate science [75, 135, 186],
turbulence modeling [28, 179, 273], material discovery [40, 202, 226], quantum chemistry [222, 228],
biological sciences [285], and hydrology [275]), and it is being performed in the context of diverse
objectives specific to these applications. Early results in isolated and relatively simple scenarios
have been promising, and the expectations are rising for this paradigm to accelerate scientific
discovery and help address some of the biggest challenges that are facing humanity in terms of
climate [75], health [259], and food security [118].
The goal of this survey is to bring these exciting developments to the ML community, to make
them aware of the progress that has been made, and the gaps and opportunities that exist for
advancing research in this promising direction. We hope that this survey will also be valuable for
scientists who are interested in exploring the use of ML to enhance modeling in their respective
disciplines. Please note that work on this topic has been referred to by other names, such as
"physics-guided ML," "physics-informed ML," or "physics-aware AI," although it covers many
scientific disciplines. In this survey, we also use the terms "physics-guided" or "physics," which
should be more generally or interpreted as science or scientific knowledge.
The focus of this survey is on approaches that integrate mechanistic modeling with ML, using
primarily ideas from physics and other scientific disciplines. This distinguishes our survey from
other works that focus on more general knowledge integration into machine learning [1, 256] and
other works covering physics integration into ML in specific domains (e.g., cyber-physical systems
[204], chemistry [184]). This survey creates a novel taxonomy specific to science-based knowledge
and covers a wide array of both methodologies and categories.
We organize the paper as follows. Section 2 describes different objectives that are being used
through combinations of scientific knowledge and machine learning. Section 3 discusses novel ML
methods and architectures that researchers are developing to achieve these objectives. Section 4
discusses opportunities for future research, and Section 5 contains concluding remarks. Table 1
categorizes the work surveyed in this paper by objectives and approaches.

2 OBJECTIVES OF PHYSICS-ML INTEGRATION

This section provides a brief overview of a diverse set of objectives where deep couplings of ML and
scientific modeling paradigms are being pursued in the context of various applications. Though we
enumerate these distinct categories in the following subsections, there are numerous cross-cutting
themes. One is the use of ML as a surrogate model where the goal is to accurately reproduce
the behavior of a mechanistic model at substantially reduced computational cost. Reduction in
computational cost is a known characteristic of ML in comparison to the often high cost of
numerical simulations, and ML-based surrogate models are relevant for many objectives. Also,
though uncertainty quantification is its own subsection, it is relevant for most of the other objectives.

2.1 Improving predictions beyond that of state-of-the-art physical models

First-principles models are used extensively in a wide range of engineering and environmental
applications. Even though these models are based on known physical laws, in most cases, they are

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 3

necessarily approximations of reality due to incomplete knowledge of certain processes, which

introduces bias. In addition, they often contain a large number of parameters whose values must be
estimated with the help of limited observed data, degrading their performance further, especially
due to heterogeneity in the underlying processes in both space and time. The limitations of physics-
based models cut across discipline boundaries and are well known in the scientific community (e.g.,
see Gupta et al. [103] in the context of hydrology).
ML models have been shown to outperform physics-based models in many disciplines (e.g.,
materials science [130, 220, 263], applied physics [15, 116], aquatic sciences [120], atmospheric
science [185], biomedical science [245], computational biology [3]). A major reason for this success
is that ML models (e.g., neural networks), given enough data, can find structure and patterns
in problems where complexity prohibits the explicit programming of a system’s exact physical
nature. Given this ability to automatically extract complex relationships from data, ML models
appear promising for scientific problems with physical processes that are not fully understood by
researchers, but for which data of adequate quality and quantity is available. However, the black-box
application of ML has met with limited success in scientific domains due to a number of reasons
[120, 126]: (i) while state-of-the-art ML models are capable of capturing complex spatiotemporal
relationships, they require far too much labeled data for training, which is rarely available in real
application settings, (ii) ML models often produce scientifically inconsistent results; and (iii) ML
models can only capture relationships in the available training data, and thus cannot generalize to
out-of-sample scenarios (i.e., those not represented in the training data).
The key objective here is to combine elements of physics-based modeling with state-of-the-art
ML models to leverage their complementary strengths. Such integrated physics-ML models are
expected to better capture the dynamics of scientific systems and advance the understanding of
underlying physical processes. Early attempts for combining ML with physics-based modeling in
several applications (e.g., modeling the lake phosphorus concentration [109] and lake temperature
dynamics [120, 128, 213]) have already demonstrated its potential for providing better prediction
accuracy with a much smaller number of samples as well as generalizability in out-of-sample
scenarios.

2.2 Downscaling
Complex first principle models are capable of capturing physical reality more precisely than
simpler models, as they often involve more diverse components that account for a greater number
of processes at finer spatial or temporal resolution. However, given the computation cost and
modeling complexity, many models are run at a resolution that is coarser than what is required
to precisely capture underlying physical processes. For example, cloud-resolving models (CRM)
need to run at sub-kilometer horizontal resolution to be able to effectively represent boundary-
layer eddies and low clouds, which are crucial for the modeling of Earth’s energy balance and the
cloud–radiation feedback [212]. However, it is not feasible to run global climate models at such fine
resolutions even with the most powerful commuters expected to be available in the near future.
Downscaling techniques have been widely used as a solution to capture physical variables that
need to be modeled at a finer resolution. In general, the downscaling techniques can be divided into
two categories: statistical downscaling and dynamical downscaling. Statistical downscaling refers
to the use of empirical models to predict finer-resolution variables from coarser-resolution variables.
Such a mapping across different resolutions can involve complex non-linear relationships that
cannot be precisely modeled by simple empirical models. Recently, artificial neural networks have
shown a lot of promise for this problem, given their ability to model non-linear relationships [232,
254]. In contrast, dynamical downscaling makes use of high-resolution regional simulations to
dynamically simulate relevant physical processes at regional or local scales of interest. Due to

, Vol. 1, No. 1, Article . Publication date: July 2020.

4 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

the substantial time cost of running such complex models, there is an increasing interest in using
ML models as surrogate models (a model approximating simulation-driven input–output data) to
predict target variables at a higher resolution [90, 237].
Although the state-of-the-art ML methods can be used in both statistical and dynamical downscal-
ing, it remains a challenge to ensure that the learned ML component is consistent with established
physical laws and can improve the overall simulation performance.

2.3 Parameterization
Complex physics-based models (e.g., for simulating phenomena in climate, weather, turbulence
modeling, astrophysics) often use an approach known as parameterization to account for missing
physics. In parameterization (note that this term has a different meaning when used in mathematics
and geometry), specific complex dynamical processes are replaced by simplified physical approxi-
mations whose associated parameter values are estimated from data, often through a procedure
referred to parameter calibration. The failure to correctly identify parameter values can make the
model less robust, and errors that result from imperfect parameterization can also feed into other
components of the entire physics-based model and deteriorate the modeling of important physical
processes. Hence, there is great interest in using ML models to learn new parameterizations directly
from observations and/or high-resolution model simulation. Already, ML-based parameterizations
have shown success in geology [43, 95] and atmospheric science [32, 33, 90, 186]. A major bene-
fit of ML-based parameterizations is the reduction of computation time compared to traditional
physics-based simulations. In chemistry, Hansen et al. [108] find that parameterizations of atomic
energy using ML take seconds compared to multiple days for more standard quantum-calculation
calculations, and Behler et al. [20] find that neural networks can vastly improve the efficiency of
finding potential energy surfaces of molecules.
Most of the existing work uses standard black box ML models for parameterization, but there is
an increasing interest in integrating physics in the ML models [23], as it has the potential to make
them more robust and generalizable to unseen scenarios as well as reduce the number of training
samples needed for training.
Note that both parameterization and downscaling described in Section 2.2 are used to replace
components of larger models. To avoid confusion, we distinctly use downscaling to relate to
resolutions (e.g., resolving certain physical processes to higher resolutions) and use parameterization
for creating simplified parameterized approximations of complex dynamical processes.

2.4 Reduced-Order Models

Reduced-Order Models (ROMs) are computationally inexpensive representations of more complex
models. Usually, constructing ROMs involves dimensionality reduction that attempts to capture
the most important dynamical characteristics of often large, high-fidelity simulations and models
of physical systems (e.g., in fluid dynamics [145]). This can also be viewed as a controlled loss
of accuracy. A common way to do this is to project the governing equations of a system onto a
linear subspace of the original state space using a method such as principal components analysis
or dynamic mode decomposition [201]. However, limiting the dynamics to a lower-dimensional
subspace inherently limits the accuracy of any ROM.
ML is beginning to assist in constructing ROMs for increased accuracy and reduced computational
cost in several ways. One approach is to build an ML-based surrogate model for full-order models
[46], where the ML model can be considered a ROM. Other ways include building an ML-based
surrogate model of an already built ROM by another dimensionality reduction method [273] or
building an ML model to mimic the dimensionality reduction mapping from a full-order model to a
reduced-order model [179]. ML and ROMs can also be combined by using the ML model to learn

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 5

the residual between a ROM and observational data [257]. ML models have the potential to greatly
augment the capabilities of ROMs because of their typically quick forward execution speed and
ability to leverage data to model high dimensional phenomena.
One area of recent focus of ML-based ROMs is in approximating the dominant modes of the
Koopman (or composition) operator, as a method of dimensionality reduction. The Koopman
operator is an infinite-dimension linear operator that encodes the temporal evolution of the system
state through nonlinear dynamics [36]. This allows linear analysis methods to be applied to nonlinear
systems and enables the inference of properties of dynamical systems that are too complex to
express using traditional analysis techniques. Applications span many disciplines, including fluid
dynamics [233], oceanography [93], molecular dynamics [267], and many others. Though dynamic
mode decomposition [178] is the most common technique for approximating the Koopman operator,
many recent approaches have been made to approximate Koopman operator embeddings with
deep learning models that outperform existing methods [150, 163, 171, 180, 188, 190, 244, 262, 286].
Adding physics-based knowledge to the learning of the Koopman operator has the potential to
augment generalizability and interpretability, which current ML methods in this area tend to lack
[190].
ROMs often lack robustness with respect to parameter changes within the systems they are
representing [5], or are not cost-effective enough when trying to predict complex dynamical systems
(e.g., multiscale in space and time). Thus, incorporating principles from physics-based models could
potentially reduce the search space to enable more robust training of ROMs, and also allow the
model to be trained with less data in many scenarios.

2.5 Inverse Modeling

The forward modeling of a physical system uses the physical parameters of the system (e.g., mass,
temperature, charge, physical dimensions or structure) to predict the next state of the system or its
effects (outputs). In contrast, inverse modeling uses the (possibly noisy) output of a system to infer
the intrinsic physical parameters. Inverse problems often stand out as important in physics-based
modeling communities because they can potentially shed light on valuable information that cannot
be observed directly. One example is the use of x-ray images from a CT scan to create a 3D image
reflecting the structure of part of a person’s body. This can be viewed as a computer vision problem,
where, given many training datasets of x-ray scans of the body at different capture angles, a model
can be trained to inversely reconstruct textures and 3D shapes of different organs or other areas of
interest. Allowing for better reconstruction while reducing scan time could potentially increase
patient satisfaction and reduce overall medical costs.
Often, the solution of an inverse problem can be computationally expensive due to the potentially
millions of forward model evaluations needed for estimator evaluation or characterization of
posterior distributions of physical parameters [84]. ML-based surrogate models (in addition to
other methods such as reduced-order models) are becoming a realistic choice since they can
model high-dimensional phenomena with lots of data and execute much faster than most physical
simulations.
Inverse problems are traditionally solved using regularized regression techniques. Data-driven
methods have seen success in inverse problems in remote sensing of surface properties [59],
photonics [197], and medical imaging [161], among many others. Recently, novel algorithms using
deep learning and neural networks have been applied to inverse problems. While still in their infancy,
these techniques exhibit strong performance for applications such as computerized tomography
[47, 174], seismic processing [253], or various sparse data problems.
There is also increasing interest in the inverse design of materials using ML, where desired target
properties of materials are used as input to the model to identify atomic structures that exhibit such

, Vol. 1, No. 1, Article . Publication date: July 2020.

6 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

properties [137, 151, 202, 226]. Physics-based constraints and stopping conditions based on material
properties can be used to guide the optimization process [151]. These constraints and similar
physics-guided techniques have the potential to alleviate noted challenges in inverse modeling,
particularly in scenarios with a small sample size and a paucity of ground-truth labels [127]. The
integration of prior physical knowledge is common in current approaches to the inverse problem,
and its integration into ML-based inverse models has the potential to improve data efficiency and
increase its ability to solve ill-posed inverse problems.

2.6 Forward Solving Partial Differential Equations

In many physical systems, governing equations are known, but direct numerical solutions of partial
differential equations (PDEs) using common methods, such as the Finite Elements Method or the
Finite Difference Method [80], are prohibitively expensive. In such cases, traditional methods are
not ideal or sometimes even possible. A common technique is to use an ML model as a surrogate
for the solution to reduce computation time [65, 140]. In particular, NN solvers can reduce the
high computational demands of traditional numerical methods into a single forward-pass of a NN.
Notably, solutions obtained via NNs are also naturally differentiable and have a closed analytic form
that can be transferred to any subsequent calculations, a feature not found in more traditional solving
methods [140]. Especially with the recent advancement of computational power, neural networks
models have shown success in approximating solutions across different kinds of physics-based
PDEs [9, 131, 215], including the difficult quantum many-body problem [42] and many-electron
Schrodinger equation [107]. As a step further, deep neural networks models have shown success
in approximating solutions across high dimensional physics-based PDEs previously considered
unsuitable for approximation by ML [106, 235]. However, slow convergence in training, limited
applicability to many complex systems, and reduced accuracy due to unawareness of physical laws
can prove problematic.

2.7 Discovering Governing Equations

When the governing equations of a dynamical system are known explicitly, they allow for more
robust forecasting, control, and the opportunity for analysis of system stability and bifurcations
through increased interpretability [217]. Furthermore, if the learned mathematical model accurately
describes the processes governing the observed data, it therefore can generalize to data outside
of the training domain. However, in many disciplines (e.g., neuroscience, cell biology, finance,
epidemiology) dynamical systems have no formal analytic descriptions. Often in these cases, data
is abundant, but the underlying governing equations remain elusive. In this section, we discuss
equation discovery systems that do not assume the structure of the desired equation (as in Section
2.6), but rather explore a space a large space of possibly nonlinear mathematical terms.
Advances in ML for the discovery of these governing equations has become an active research
area with rich potential to integrate principles from applied mathematics and physics with modern
ML methods. Early works on the data-driven discovery of physical laws relied on heuristics and
expert guidance and were focused on rediscovering known, non-differential, laws in different
scientific disciplines from artificial data [92, 143, 144, 149]. This was later expanded to include
real-world data and differential equations in ecological applications [72]. Recently, general and
robust data-driven discovery of potentially unknown governing equations has been pioneered by
[30, 227], where they apply symbolic regression to differences between computed derivatives and
analytic derivatives to determine underlying dynamical systems. More recently, works have used
sparse regression built on a dictionary of functions and partial derivatives to construct governing
equations [37, 200, 216]. Lagergren et al. [141] expand on this by using ANNs to construct the
dictionary of functions. These sparse identification techniques are based on the principle of Occam’s

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 7

Razor, where the goal is that only a few equation terms be used to describe any given nonlinear
system.

2.8 Data Generation

Data generation approaches are useful for creating virtual simulations of scientific data under
specific conditions. For example, these techniques can be used to generate potential chemical
compounds with desired characteristics (e.g., serving as catalysts or having a specific crystal
structure). Traditional physics-based approaches for generating data often rely on running physical
simulations or conducting physical experiments, which tend to be very time consuming. Also,
these approaches are restricted by what can be produced by physics-based models. Hence, these is
an increasing interest in generative ML approaches that learn data distributions in unsupervised
settings and thus have the potential to generate novel data beyond what could be produced by
traditional approaches.
Generative ML models have found tremendous success in areas such as speech recognition and
generation [187], image generation [63], and natural language processing [100]. These models have
been at the forefront of unsupervised learning in recent years, mostly due to their efficiency in
understanding unlabeled data. The idea behind generative models is to capture the inner probabilistic
distribution in order to generate similar data. With the recent advances in deep learning, new
generative models, such as the generative adversarial network (GAN) and variational autoencoder
(VAE), have been developed. These models have shown much better performance in learning non-
linear relationships to extract representative latent embeddings from observation data. Hence the
data generated from the latent embeddings are more similar to true data distribution. In particular,
the adversarial component of GAN consists of a two-part framework of a generative network and
discriminative network, where the generative network’s objective is to generate fake data to fool
the discriminative network, while the discriminative network attempts to determine true data from
fake data.
In the scientific domain, GANs can generate data like the data generated by the physics-based
models. Using GANs often incurs certain benefits, including reduced computation time and a
better reproduction of complex phenomenon, given the ability of GANs to represent nonlinear
relationships. For example, Farimani et al. [78] have shown that Conditional Generative Adversarial
Networks (cGAN) can be trained to simulate heat conduction and fluid flow purely based on
observations without using knowledge of the underlying governing equations. Such approaches
that use generative models have been shown to significantly accelerate the process of generating
new data samples.
However, a well-known issue of GANs is that they incur dramatically high sample complexity.
Therefore, a growing area of research is to engineer GANs that can leverage prior knowledge
of physics in terms of physical laws and invariance properties. For example, GAN-based models
for simulating turbulent flows can be further improved by incorporating physical constraints,
e.g., conservation laws [288] and energy spectrum [269], in the loss function. Cang et al. [40]
also imposed a physics-based morphology constraint on a VAE-based generative model used for
simulating artificial material samples. The physics-based constraint forces the generated artificial
samples to have the same morphology distribution as the authentic ones and thus greatly reduces
the large material design space.

2.9 Uncertainty Quantification

Uncertainty quantification (UQ) is of great importance in many areas of computational science
(e.g., climate modeling [64], fluid flow [55], systems engineering [195], among many others). At
its core, UQ requires an accurate characterization of the entire distribution p(y|x), where y is the

, Vol. 1, No. 1, Article . Publication date: July 2020.

8 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

response and x is the covariate of interest, rather than just making a point prediction y = f (x).
This makes it possible to characterize all quantiles and skews in the distribution, which allows for
analysis such as examining how close predictions are to being unacceptable, or sensitivity analysis
of input features.
Applying UQ tasks to physics-based models using traditional methods such as Monte Carlo
(MC) is usually infeasible due to the thousands or millions of forward model evaluations needed to
obtain convergent statistics. In the physics-based modeling community, a common technique is to
perform model reduction (described in Section 2.4) or create an ML surrogate model, in order to
increase model evaluation speed since ML models often execute much faster [87, 170, 250]. With a
similar goal, the ML community has often employed Gaussian Processes as the main technique for
quantifying uncertainty in simulating physical processes [26, 211], but neither Gaussian Processes
nor reduced models scale well to higher dimensions or larger datasets (Gaussian Processes scale as
O(N 3 ) with N data points).
Consequently, there is an effort to fit deep learning models, which have exhibited countless
successes across disciplines, as a surrogate for numerical simulations in order to achieve faster
model evaluations for UQ that have greater scalability than Gaussian Processes [250]. However,
since artificial neural networks do not have UQ naturally built into them, variations have been
developed. One such modification uses a probabilistic drop-out strategy in which neurons are
periodically "turned off" as a type of Bayesian approximation to estimate uncertainty [86]. There
are also Bayesian variants of neural networks that consist of distributions of weights and biases
[164, 296, 300], but these suffer from high computation times and high dependence on reliable
priors. Another method uses an ensemble of neural networks to create a distribution from which
uncertainty is quantified [142].
The integration of prior physics knowledge into ML for UQ has the potential to allow for a
better characterization of uncertainty. For example, ML surrogate models run the risk of producing
physically inconsistent predictions, and incorporating elements of physics could help with this
issue. Also, note that the reduced data needs of ML due to constraints for adherence to known
physical laws could alleviate some of the high computational cost of Bayesian neural networks for
UQ.

3 PHYSICS-ML METHODS
Given the diversity of forms in which scientific knowledge is represented in different disciplines
and applications, a diverse set of methodologies are needed to integrate physical principles into
ML models. This section creates a taxonomy of five classes of methodologies to merge principles of
physics-based modeling with ML; (i) physics-guided loss function, (ii) physics-guided initialization,
(iii) physics-guided design of architecture, (iv) residual modeling, and (v) hybrid physics-ML models.

3.1 Physics-Guided Loss Function

Scientific problems often exhibit a high degree of complexity due to relationships between many
physical variables varying across space and time at different scales. Standard ML models can fail to
capture such relationships directly from data, especially when provided with limited observation
data. This is one reason for their failure to generalize to scenarios not encountered in training
data. Researchers are beginning to incorporate physical knowledge into loss functions to help ML
models capture generalizable dynamic patterns consistent with established physical laws.
One of the most common techniques to make ML models consistent with physical laws is to
incorporate physical constraints into the loss function of ML models as follows [126],
Loss = LossTRN (Ytrue , Ypred ) + λR(W ) + γ LossPHY (Ypred ) (1)

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 9

where the training loss LossTRN measures a supervised error (e.g., RMSE or cross-entropy) between
true labels Ytrue and predicted labels Ypred , and λ is a hyper-parameter to control the weight of
model complexity loss R(W ). These first two terms are the standard loss of ML models. The addition
of physics-based loss LossPHY aims to ensure consistency with physical laws and it is weighted by
a hyper-parameter γ .
Steering ML predictions towards physically consistent outputs has numerous benefits. First,
the computation of physics-based loss LossPHY requires no observation data and thus optimiz-
ing LossPHY allows including unlabeled data in training. Second, the regularization by physical
constraints can reduce the possible search space of parameters. This provides possibility to learn
with less labeled data while also ensuring the consistency with physical laws during optimization.
Third, ML models which follow desired physical properties are more likely to be generalizable to
out-of-sample scenarios, and thus become acceptable for use by domain scientists and stakeholders
in physical science applications. Loss function terms corresponding to physical constraints are ap-
plicable across many different types of ML frameworks and objectives. In the following paragraphs,
we demonstrate the use of physics-based loss functions for different objectives described in Section
2.
Improving predictions. In the context of complex natural systems, incorporation of physics-based
loss has shown great success in improving prediction ability of ML models. In the context of
lake temperature modeling, Karpatne et al. [128] includes an additional physics-based penalty
that ensures that predictions of denser water predictions are at lower depths than predictions of
less dense water, a known monotonic relationship. Jia et al. [119] and Read et al. [213] further
extended this work to capture even more complex and general physical relationships that happen
on a temporal scale. Specifically, they use a physics-based penalty for energy conservation in
the loss function to ensure the lake thermal energy gain across time is consistent with the net
thermodynamic fluxes in and out of the lake. A diagram of this model is shown in Figure 1. Note
that the recurrent structure contains additional nodes (shown in blue) to represent physical variable
(lake energy, etc) that are computed using purely physics-based equations. These are needed to
incorporate energy conservation in the loss function. Similar structure can be used to model other
physical laws such as mass conservation.
Solving and Discovering PDEs and Governing Equations. Another strand of work that involves loss
function alterations is solving PDEs for dynamical systems modeling, in which adherence to the
governing equations is enforced in the loss function. In Raissi et al. [208], this concept is developed
and shown to create data-efficient spatiotemporal function approximators to both solve and find
parameters of basic PDEs like Burgers Equation or Schrodinger Equation. Going beyond a simple
feed-forward network, Zhu et al. [301] propose an encoder-decoder for predicting transient PDEs
with governing PDE constraints. Geneva et al. [89] extended this approach to deep auto-regressive
dense encoder-decoders with a Bayesian framework using stochastic weight averaging to quantify
uncertainty.
Qualitative mathematical properties of dynamical systems modeling have also shown promise in
informing loss functions for improved prediction beyond that of the physics model. Erichson et al.
[74] penalize autoencoders based on physically meaningful stability measures in dynamical systems
for improved prediction of fluid flow and sea surface temperature. They showed an improved
mapping of past states to future states in time in both modeling scenarios in addition to improved
generalizability to new data.
Physics-based loss function terms have also been used in the discovery of governing equations.
Loiseau et al. [155] uses constrained least squares [96] to incorporate energy-preserving nonlineari-
ties or to enforce symmetries in the identified equations to the equation learning process described

, Vol. 1, No. 1, Article . Publication date: July 2020.

10 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

Fig. 1. The flow of the Physics-Guided Recurrent Neural Network model proposed in Jia et al. [120]. They
include the standard RNN flow (black arrows) and the energy flow (blue arrows) in the recurrent process.
Here UT represents the thermal energy of the lake at time T , and both the energy output and temperature
output yT are used in calculating the loss function value. Figure adapted from Jia et al. [120].

in Section 2.7. Though these loss functions are mostly seen in common variants of NNs, they can
also be seen in architectures such as echo state networks. In Doan et al. [66], they find integrating
the physics-based loss from the governing equations in a Lorenz system, a commonly studied
system in dynamical systems, strongly improves the echo state network’s time-accurate prediction
of the system and also reduces convergence time.
Inverse modeling. For applications in vortex induced vibrations, Raissi et al. [210] pose the inverse
modeling problem of predicting the lift and drag forces of a system given sparse data about its
velocity field. Kahana et al. [122] uses a loss function term pertaining to the physical consistency
of the time evolution of waves for the inverse problem of identifying the location of an underwater
obstacle from acoustic measurements. In both cases, the addition of physics-based loss terms made
results more accurate and more robust to out-of-sample scenarios.
Parameterization. While ML has been used for parameterization, adding physics-based loss
terms can further benefit this process by ensuring physically consistent outputs. Zhang et al.
[292] parameterizes atomic energy for molecular dynamics using a NN with a loss function that,
in addition to loss due to accuracy, takes into account atomic force, atomic energy, and terms
relating kinetic and potential energy. Furthermore, in the context of climate modeling, Beucler
et al. show that enforcing energy conservation laws improves prediction when emulating cloud
processes [23, 24].
Uncertainty quantification. In Yang et al. [281] and Yang et al. [282], physics-based loss is imple-
mented in a deep probabilistic generative model for uncertainty quantification based on adherence
to the structure imposed by PDEs. To accomplish this, they construct probabilistic representations
for the system states, and use an adversarial inference procedure to train them on data using a
physics-based loss function that enforces adherence to the governing laws. This is expanded in Zhu
et al. [301], where a physics-informed encoder-decoder network is defined in conjunction with a

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 11

conditional flow-based generative model for similar purposes. A similar loss function modification is
performed in other works [89, 129, 278], but for the purpose of solving high dimensional stochastic
PDEs with uncertainty propagation. In these cases, physics-guided constraints provide effective
regularization for training deep generative models to serve as surrogates of physical systems where
the cost of acquiring data is high and the data sets are small [282].
Another direction for encoding physics knowledge into ML UQ applications is to create a
physics-guided Bayesian NN. This is explored in Yang et al. [277], where they use a Bayesian NN,
which naturally encodes uncertainty, as a surrogate for a PDE solution. Additionally, they add
a PDE constraint for the governing laws of the system to serve as a prior for the Bayesian net,
allowing for more accurate predictions in situations with significant noise due to the physics-based
regularization.

Generative models. In recent years, GANs have been used to efficiently generate solutions to
PDEs and there is interest in using physics knowledge to improve them. Work by Yang et al. [279]
shows GANs with loss functions based on PDEs can be used to solve stochastic elliptic PDEs in up to
30 dimensions. In a similar vein, Wu et al. [268] showed that physics-based loss functions in GANs
can lower the amount of data and training time needed to converge on solutions of turbulence
PDEs, while Shan et al. [231] saw similar results in the generation of microstructures satisfying
certain physical properties in computational materials science.

3.2 Physics-Guided Initialization

Since many ML models require an initial choice of model parameters before training, researchers
have explored different ways to physically inform a model starting state. For example, in NNs,
weights are often initialized according to a random distribution prior to training. Poor initialization
can cause models to anchor in local minima, which is especially true for deep neural networks.
However, if physical or other contextual knowledge can be used to help inform the initialization
of the weights, model training can be accelerated or improved [120]. One way to inform the
initialization to assist in model training and escaping local minima is to use an ML technique
known as transfer learning. In transfer learning, a model can be pre-trained on a related task prior
to being fine-tuned with limited training data to fit the desired task. The pre-trained model serves
as an informed initial state that ideally is closer to the desired parameters for the desired task
than random initialization. One way to harness physics-based modeling knowledge is to use the
physics-based model’s simulated data to pre-train the ML model, which also alleviates data paucity
issues. This is similar to the common application of pre-training in computer vision, where CNNs
are often pre-trained with very large image datasets before being fine-tuned on images from the
task at hand [243].
Jia et al. use this strategy in the context of modeling lake temperature dynamics [119, 120]. They
pre-train their Physics-Guided Recurrent Neural Network (PGRNN) models for lake temperature
modeling on simulated data generated from a physics-based model and fine tune the NN with
little observed data. They showed that pre-training, even using data from a physical model with
an incorrect set of parameters, can still significantly reduce the training data needed for a quality
model. In addition, Read et al. [213] demonstrated that such models are able to generalize better to
unseen scenarios than pure physics-based models.
Another application can be seen in computer vision in robotics, where images from robotics
simulations have been shown to be sufficient without any real-world data for the task of object
localization [248], while reducing data requirements by a factor of 50 for object grasping [31]. Then,
in autonomous vehicle training, Shah et al. [230] showed that pre-training the driving algorithm in
a simulator built on a video game physics engine can drastically lessen data needs. More generally,

, Vol. 1, No. 1, Article . Publication date: July 2020.

12 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

we see that simulation pre-training of applications allows for significantly less expensive data
collection than is possible with physical robots.
Another significant application of physics-guided initialization is in computational biophysics,
where high computational cost of molecular dynamics simulations limits the number of mutant
molecules that can be investigated from a given baseline molecule. Sultan et al. [239] showed that
the molecular dynamics of a mutant protein can be estimated by training a variational autoencoder
on the simulation of the protein from which the target protein mutated, allowing for surrogate
models to offset the high cost of simulations. This method promises to be able to quickly sample
related molecular systems to enable probing of the variation in large complex molecular structures,
an otherwise very computationally expensive task. In a similar vein, in the context of predicting
protein structures, Hurtado et al. [114] shows that a deep NN can be pre-trained on a large database
of already-known protein structures to enhance performance.
Physics-guided model initialization has also been employed in chemical process modeling [158,
159, 276]. Yan et al. [276] uses Gaussian process regression for process modeling that has been
transferred and adapted from a similar task. To adapt the transferred model, they use scale-bias
correcting functions optimized through maximum likelihood estimation of parameters. Furthermore,
Gaussian process models come equipped with uncertainty quantification which is also informed
through initialization. A similar transfer and adapt approach is seen in Lu et al. [158], but for an
ensemble of NNs transferred from related tasks. In both of the previous studies, the similarity metrics
used to find similar systems are defined by considering various common process characteristics
and behaviors.

3.3 Physics-Guided Design of Architecture

Although the physics-based loss and initialization in the previous sections helps constrain the search
space of ML models during training, the ML architecture is often still a black box. In particular, they
do not use any architectural properties to implicitly encode physical consistency or other desired
physical properties. A recent research direction has been to construct new ML architectures that
can make use of the specific characteristics of the problem being solved. . Though this section is
focused largely on neural network architecture, we also include subsections on multi-task learning
and structures of Gaussian processes. The modular and flexible nature of NNs in particular make
them prime candidates for architecture modification. For example, domain knowledge can used to
specify node connections that capture physics-based dependencies among variables. Furthermore,
incorporating physics-based guidance into architecture has the added bonus of also making the
previously black box algorithm more interpretable, a desirable but typically missing feature of ML
models used in physical modeling. In the following paragraphs, we discuss several contexts in
which physics-guided ML architectures have been used.

Intermediate Physical Variables. One way to embed known physical principles into NN design is
to ascribe physical meaning for certain neurons in the NN. It is also possible to declare physically
relevant variables explicitly. In lake temperature modeling, Daw et al. [58] incorporate a physical
intermediate variable as part of a monotonicity-preserving structure in the LSTM architecture as
shown in Figure 2. This model produces physically consistent predictions in addition to appending a
dropout layer to quantify uncertainty. In Muralidlar et al. [182], a similar approach is taken to insert
physics-constrained variables as the intermediate variables in the convolutional neural network
(CNN) architecture and achieve significant improvement over state-of-the-art physics-based models
on the problem of predicting drag force on particle suspensions in moving fluids.
An additional benefit of adding physically relevant intermediate variables in an ML architecture
is that they can help extract physically meaningful hidden representation that can be interpreted

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 13

Fig. 2. Monotonicity-preserving LSTM Architecture. Components in red represent novel physics-informed

innovations in LSTM. Zd represents the density values at a specific depth d, δd is the increase of density
from layer d − 1 to d. The value of δ is gauranteed to be non-negative as it is generated from an ReLU layer.
Such an architecture ensures the monotonic increase of density values as depth increases. The following link
contains the code for this paper: https://github.com/arkadaw9/PGN. Figure adapted from [58]
,

by domain scientists. This is particularly valuable, as standard deep learning models are limited in
their interpretability since they can only extract abstract hidden variables using highly complex
connected structure. This is further exacerbated given the randomness involved in the optimization
process.
Another related approach is to fix one or more weights within the NN to physically meaningful
values or parameters and make them non-modifiable during training. A recent approach is seen in
geophysics where researchers use NNs for the waveform inversion modeling to find subsurface
parameters from seismic wave data. In Sun et al. [240], they assign most of the parameters within a
network to mimic seismic wave propagation during forward propagation of the NN, with weights
corresponding to values in known governing equations. They show this leads to more robust
training, in addition to creating a more interpretable NN with meaningful intermediate variables.

Encoding invariances and symmetries. In physics, there is a deep connection between symmetries
and invariant quantities of a system and its dynamics. For example, Noether’s Law, a common
paradigm in physics, demonstrates a mapping between conserved quantities of a system and the
system’s symmetries (e.g. translational symmetry can be shown to correspond to the conservation
of momentum within a system). Therefore, if an ML model is created that is translation-invariant,
the conservation of momentum becomes more likely and the model’s prediction becomes more
robust and generalizable.
State-of-the-art deep learning architectures already encode certain types of invariance; for
example, RNNs encode time invariance and CNNs can implicitly encode spatial translation, rotation,
scale variance. In the same way, scientific modeling tasks may require other invariances based
on physical laws. In turbulence modeling and fluid dynamics, Ling et al [152] defines a tensor
basis neural network to embed the fundamental principle of rotational invariance into a NN for
improved prediction accuracy. This solves a key problem in ML models for turbulence modeling
because without rotational invariance, the model evaluated on identical flows with axes defined in
other directions could yield different predictions. This work alters the NN architecture by adding a
higher-order multiplicative layer that ensures the prediction lies on a rotationally invariant tensor
basis. In a molecular dynamics application, Anderson et al. [6] show that a rotationally covariant
NN architecture can learn the behavior and properties of complex many-body physical systems.

, Vol. 1, No. 1, Article . Publication date: July 2020.

14 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

In a general setting, Wang et al. [260] show how spatiotemporal models can be made more
generalizable by incorporating symmetries into deep NNs. More specifically, they demonstrated
the encoding of translational symmetries, rotational symmetries, scale invariances, and uniform
motion into NNs using customized convolutional layers in CNNs that enforce desired invariance
properties. They also provided theoretical guarantees of invariance properties across the different
designs and showed additional to significant increases in generalization performance.
Incorporating symmetries, by informing the structure of the solution space, also has the potential
to reduce the search space of an ML algorithm. This is important in the application of discovering
governing equations, where the space of mathematical terms and operators is exponentially large.
Though in its infancy, physics-informed architectures for discovering governing equations are
beginning to be investigated by researchers. In Section 2.7, symbolic regression is mentioned as
an approach that has shown success. However, given the massive search space of mathematical
operators, analytic functions, constants, and state variables, the problem can quickly become NP-
hard. Udrescu et al. [251] designs a recursive multidimensional version of symbolic regression that
uses a NN in conjunction with techniques from physics to narrow the search space. Their idea is to
use NNs to discover hidden signs of "simplicity", such as symmetry or separability in the training
data, which enables breaking the massive search space into smaller ones with fewer variables to be
determined.
In the context of molecular dynamics applications, a number of researchers [21, 292] have used
a NN per individual atom to calculate each atom’s contribution to the total energy. Then, to ensure
the energy invariance with respect to the possibility of interchanging two atoms, the structure of
each NN and the values of each network’s weight parameters are constrained to be identical for
atoms of the same element. More recently, novel deep learning architectures have been proposed for
fundamental invariances in chemistry. Schutt et al. [228] proposes continuous-filter convolutional
(cfconv) layers for CNNs to allow for modeling objects with arbitrary positions such as atoms in
molecules, in contrast to objects described by Cartesian-gridded data such as images. Furthermore,
their architecture uses atom-wise layers that incorporate inter-atomic distances that enabled the
model to respect quantum-chemical constraints such as rotationally invariant energy predictions
as well as energy-conserving force predictions. Furthermore, Zhang et al. [293] propose an end-to-
end modeling framework that preserves all natural symmetries of a molecular system using an
embedding procedure that maps the input to symmetry-preserving components. This is further
expanded for 3D point clouds in the "Tensor field networks" proposed by Thomas et al. [246], where
they use continuous convolution layers on point cloud data, and also constrain the network filters
to be the product of a learnable radial function and a spherical harmonic that make them rotation-
and translation-equivariant. As we can see, because molecular dynamics often ascribes importance
to different important geometric properties of molecules (e.g. rotation), network architectures
dealing with invariances can be effective for improving performance and robustness of ML models.
Architecture modifications incorporating symmetry are also seen extensively in dynamical
systems research involving differential equations. In a pioneering work by Ruthotto et al [218],
three variations of CNNs are proposed for solving PDEs. Each variation uses mathematical theories
to guide the design of the CNN based on fundamental properties of the PDEs. Multiple types of
modifications are made, including adding symmetry layers to guarantee the stability expressed
by the PDEs and layers that convert inputs to kinematic eigenvalues that satisfy certain physical
properties. They define a parabolic CNN inspired by anisotropic filtering, a hyperbolic CNN based
on Hamiltonian systems, and a second order hyperbolic CNN. Hyperbolic CNNs were found to
preserve the energy in the system as intended, which set them apart from parabolic CNNs that
smooth the output data, reducing the energy.

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 15

A recent direction also relating to conserved or invariant quantities is the incorporation of

the Hamiltonian operator into NNs [54, 98, 249, 299]. The Hamiltonian operator in physics is the
primary tool for modeling the time evolution of systems with conserved quantities, but until recently
the formalism had not been integrated with NNs. Greydanus et al. [98] design a NN architecture that
naturally learns and respects energy conservation and other invariance laws in simple mass-spring
or pendulum systems. They accomplish this through predicting the Hamiltonian of the system and
re-integrating instead of predicting the state of physical systems themselves. This is taken a step
further in Toth et al. [249], where they show that not only can NNs learn the Hamiltonian, but also
the abstract phase space (assumed to be known in Greydanus et al. [98]) to more effectively model
expressive densities in similar physical systems and also extend more generally to other problems
in physics. Recently, the Hamiltonian-parameterized NNs above have also been expanded into NN
architectures that perform additional differential equation-based integration steps based on the
derivatives approximated by the Hamiltonian network [51].

Encoding other domain-specific physical knowledge. Various other domain-specific physical infor-
mation can be encoded into architecture that doesn’t exactly correspond to known invariances
but provides meaningful structure to the optimization process depending on the task at hand. This
can take place in many ways, including using domain-informed convolutions for CNNs, additional
domain-informed discriminators in GANs, or structures informed by the physical characteristics of
the problem. For example, Sadoughi et al. [221] prepend a CNN with a Fast Fourier Transform layer
and a physics-guided convolution layer based on known physical information pertaining to fault
detection of rolling element bearings. A similar approach is used in Sturmfels et al. [238], but the
added beginning layer instead serves to segment different areas of the brain for domain guidance
in neuroimaging tasks. In the context of generative models, Xie et al. [274] introduce tempoGAN,
which augments a general adversarial network with an additional discriminator network along with
additional loss function terms that preserve temporal coherence in the generation of physics-based
simulations of fluid flow. This type of approach, though found mostly in NN models, has been
extended to non-NN models in Baseman et al. [17], where they introduce a physics-guided Markov
Random Field that encodes spatial and physical properties of computer memory devices into the
corresponding probabilistic dependencies.
A number of advances in encoding physical knowledge have been made in the context of quantum
chemistry and molecular dynamics. In computational quantum chemistry, Pfau et al. [196] develop
the paradigm of spin-based neural-network quantum states by defining a custom NN architecture
to map electron states to the atom’s wave function, which characterizes its atomic state. Key
architecture components are intermediate layers in the NN to take the mean of the activation
functions from each electron stream, and a final layer which applies a quantum spin-dependent
linear transformation, weights the desired outputs as the solution wave function, and enforces
any boundary conditions. In molecular kinetics, reduced order modeling frameworks are used for
approximating the Koopman operator [171, 262]. Mardt et al. [171] describe a dual NN architecture
that transforms the molecular configurations found at given time delays (e.g. t and t + τ ) to Markov
states along simulation trajectories to perform nonlinear dimensionality reduction. They show this
outperforms state-of-the-art existing methods and also remains interpretable due to the mapping
to realistic physical states. A similar approach is seen in [262], where they take a similar time-delay
approach to improve over state-of-the-art but instead use autoencoders, which allows for the
addition of a decoding step that can predict configurations in the original coordinate space from
latent space points. This encoder-decoder architecture is further expanded on by Otto et al. [188],
where they go beyond linear reconstruction using Koopman modes in the decoder step to an

, Vol. 1, No. 1, Article . Publication date: July 2020.

16 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

architecture that does nonlinear reconstruction in order to learn ever lower dimensional Koopman
subspaces.
Fan et al. [77] define new architectures to solve the inverse problem of electrical impedance
tomography, where the goal is to determine the electrical conductivity distribution of an unknown
medium from electrical measurements along its boundary. They define new NN layers based on a
linear approximation of both the forward and inverse maps relying on the nonstandard form of the
wavelet decomposition [25].
Architecture modifications are also seen in dynamical systems research encoding principles
from differential equations. Chen et al. [48] develops a continuous depth NN based on the Residual
Network [110] for solving ordinary differential equations. They change the traditionally discretized
neuron layer depths into continuous equivalents such that hidden states can be parameterized
by differential equations in continuous time. This allows for increased computational efficiency
due to the simplification of the backpropagation step of training, and also creates a more scalable
normalizing flow, an architectural component for solving PDEs. This is done by by parameterizing
the derivative of the hidden states of the NN as opposed to the states themselves. Then, in a similar
application, Chang et al. [44] uses principles from the stability properties of differential equations in
dynamical systems modeling to guide the design of the gating mechanism and activation functions
in an RNN.
Currently, human experts have manually developed the majority of domain knowledge-encoded
employed architectures, which can be a time-intensive and error-prone process. Because of this,
there is increasing interest in automated neural architecture search methods [13, 73, 115]. A young
but promising direction in ML architecture design is to embed prior physical knowledge into neural
architecture searches. Ba et al. [12] adds physically meaningful input nodes and physical operations
between nodes to the neural architecture search space to enable the search algorithm to discover
more ideal physics-guided ML architectures.
Auxiliary Task in Multi-Task Learning. Domain knowledge can be incorporated into ML architec-
ture as auxiliary tasks in a multi-task learning framework. Multi-task learning allows for multiple
learning tasks to be solved at the same time, ideally while exploiting commonalities and differences
across tasks. This can result in improved learning efficiency and predictions for one or more of the
tasks. Therefore, another way to implement physics-based learning constraints is to use an auxiliary
task in a multi-task learning framework. Here, an example of an auxiliary task in a multi-task
framework might be related to ensuring physically consistent solutions in addition to accurate
predictions. The promise of such an approach was demonstrated for a computer vision task by
integrating auxiliary information (e.g. pose estimation) for facial landmark detection [297]. In this
paradigm, a task-constrained loss function can be formulated to allow errors of related tasks to be
back-propagated jointly to improve model generalization. Early work in a computational chemistry
application showed that a NN could be trained to predict energy by constructing a loss function
that had penalties for both inaccuracy and inaccurate energy derivatives with respect to time as
determined by the surrounding energy force field [199]. In particle physics, De Oliveira et al. [62]
uses an additional task for the discriminator network in a generative adversarial network (GAN) to
satisfy certain properties of particle interaction for the production of jet images of particle energy.
Physics-guided Gaussian process regression. Gaussian process regression (GPR) [265] is a nonpara-
metric, Bayesian approach to regression that is increasingly being used in ML applications. GPR
has several benefits, including working well on small amounts of data and enabling uncertainty
measurements on predictions. In GPR, first a Gaussian process prior must be assumed in the form
of a mean function and a matrix-valued kernel or covariance function. One way to incorporate
physical knowledge in GPR is to encode differential equations into the kernel [242]. This is a key

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 17

feature in Latent Force Models, which attempt to use equations in the physical model of the system
to inform the learning from data [4, 160]. Alvarez et al. [4] draws inspiration from similar applica-
tions in bioinformatics [88, 146], which showed an increase in predictive ability in computational
biology, motion capture, and geostatistics datasets. More recently, Glielmo et al. [94] propose a
vectorial GPR that encodes physical knowledge in the matrix-valued kernel function. They show
rotation and reflection symmetry of the interatomic force between atoms can be encoded in the
Gaussian process with specific invariance-preserving covariant kernels. Furthermore, Raissi et
al. [206] show that the covariance function can explicitly encode the underlying physical laws
expressed by differential equations in order to solve PDEs and learn with smaller datasets.

3.4 Residual modeling

The oldest and most common approach for directly addressing the imperfection of physics-based
models in the scientific community is residual modeling, where an ML model (usually linear
regression) learns to predict the errors, or residuals, made by a physics-based model [82, 247]. A
visualization is shown in Figure 3. The key concept is to learn biases of the physical model (relative
to observations) and use it to make corrections to the physical model’s predictions. However, one
key limitation in residual modeling is their inability to enforce physics-based constraints (like
in Section 3.1) because such approaches model the errors instead of the physical quantities in
physics-based problems.

Fig. 3. Machine Learning model f M L trained to model the error made by Physics model f P H Y . Figure adapted
from [82]

Recently, a key area in which residual modeling has been applied is in reduced order models
(ROMs) of dynamical systems (described in Section 2.4). After reducing model complexity to create
a ROM, an ML model can be used to model the residual due to the truncation. ROM methods were
created in response to the problem of many detailed simulations being too expensive to be used in
various engineering tasks including design optimization and real-time decision support. In San et
al. [223, 224], a simple NN used to model the error due to the model reduction is shown to sharply
reduce high error regions when applied to known differential equations. Also, in Wan et al. [257],
an RNN is used to model the residual between a ROM for prediction of extreme weather events
and the available data projected to a reduced-order space.
As another example, in Kani et al. [125] a physics-driven "deep residual recurrent neural network
(DR-RNN)" is proposed to find the residual minimiser of numerically discretized PDEs. Their
architecture involves a stacked RNN embedded with the dynamical structure of the PDEs such
that each layer of the NN solves a layer of the residual equations. They showed that DR-RNN
sharply reduces both computational cost and time discretization error of the reduced order modeling
framework.

, Vol. 1, No. 1, Article . Publication date: July 2020.

18 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

3.5 Hybrid Physics-ML Models

Residual modeling is just one way to combine physics-based models with ML where both are
operating simultaneously. We will call a generalized version of this Hybrid Physics-ML Models.
One straightforward method to combine physics-based and ML models is to feed the output
of a physics-based model as input to an ML model. Karpatne et al [128] showed that using the
output of a physics-based model as one feature in an ML model along with inputs used to drive the
physics-based model for lake temperature modeling can improve predictions. Visualization of this
method is shown in Figure 4. As we discuss below, there are multiple other ways of constructing a
hybrid model, including replacing part of a larger physical model, or weighting predictions from
different modeling paradigms depending on context.

Fig. 4. Diagram of a hybrid physics-ML model. (Figure adapted from Karpatne et al. [128]). fPHY represents
the physics-based model which converts the input drivers D to simulated outputs YPHY , fPHY represents the
hybrid physics-ML model that combines D and YPHY as input to make final predictions.

In one variant of hybrid physics-data models, ML models are used to replace one or more
components of a physics-based model or to predict an intermediate quantity that is poorly modeled
using physics. For example, to improve predictions of the discrepancy of Reynolds-Averaged
Navier–Stokes (RANS) solvers in fluid dynamics, Parish et al. [192] propose a NN to estimate
variables in the turbulence closure model to account for missing physics. They show this correction
to traditional turbulence models results in convincing improvements of predictions. In Hamilton et
al. [105], a subset of the mechanistic model’s equations are replaced with data-driven nonparametric
methods to improve prediction beyond the baseline process model. In Zhang et al. [294], a physics-
based architecture for power system state estimation embeds a deep learning model in place of
traditional predicting and optimization techniques. To do this, they substitute NN layers into
an unrolled version of an existing solution framework which drastically reduced the overall
computational cost due to the fast forward evaluation property of NNs, but kept information
of the underlying physical models of power grids and of physical constraints.
ML models for parameterization (see Section 2.3) can also be viewed as a type of hybrid modeling.
Most of these efforts use black box ML models (citations covered in first column of Table 1), but some
of these parameterization models can use more sophisticated physics-guided versions of ML. For
example, in geodynamics modeling, Goldstein et al. [95] showed that an ML model for oscillatory
flow ripples can be substituted into an established model for geologic material being moved by
fluid flow. They showed that the hybrid model was able to capture dynamics previously absent
from the model, which led to the discovery of two new observed pattern modes in the geology. In

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 19

chemical physics, Manzhos et al. [169] first showed that NNs could be used to represent component
functions within high dimensional model representations (HDMR) of energy potentials. They show
this improves the quality of the final representation and reduces the cost of the fits of component
functions for HDMR traditionally done with explicit equations. Furthermore, the generality of the
hybrid framework allows the HDMR to be constructed in a black box, molecule-independent way.
In another class of hybrid frameworks, the overall prediction is a combination of predictions
from a physical model and an ML model, where the weights depend on prediction circumstance.
For example, long-range interactions (e.g. gravity) can often be more easily modeled by classical
physics equations than more stochastic short-range interactions (quantum mechanics) that are
better modeled using data-driven alternatives. Hybrid frameworks like this have been used to
adaptively combine ML predictions for short-range processes and physics model predictions for long-
range processes for applications in chemical reactivity [284] and seismic activity prediction [191].
Estimator quality at a given time and location can also be used to determine whether a prediction
comes from the physical model or the ML model, which was shown in Chen et al. [50] for air
pollution estimation and in Vlachas et al. [255] for dynamical system forecasting more generally. In
the context of solving PDEs, Malek et al [166] showcases a hybrid NN and traditional optimization
technique to find the closed analytical form of the solution of a PDE. In this hybrid solver, there
exist two terms, a term described by the NN and a term described by traditional optimization
techniques.
Moreover, in inverse modeling, there is a growing use of hybrid models that first use physics-
based models to perform the direct inversion, then use deep learning models to further refine
the inverse model’s predictions. Multiple works have shown an effective application for this in
computed tomography (CT) reconstructions [38, 121]. Another common technique in inverse
modeling of images (e.g. medical imaging, particle physics imaging), is the use of CNNs as deep
image priors [252]. To simultaneously exploit data and prior knowledge, Senouf et al. [229] embed
a CNN that serves as the image prior for a physics-based forward model for MRIs.

, Vol. 1, No. 1, Article . Publication date: July 2020.

20 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

Table 1. Table of literature classified by objective and method

Physics-Guided Physics-Guided Physics-Guided

Basic ML Loss Function Initialization Architecture Residual Model Hybrid Model

Improve prediction [287] [35] [102] [247] [4] [199] [158] [276] [162] [152] [56] [17] [247] [275] [258] [95] [99] [222]
beyond physical model [104] [271] [139][128] [160] [248] [31] [230] [238] [6] [11] [223] [257] [271] [105] [128] [236]
[183] [66] [74] [239] [114] [119] [183] [182] [193] [153] [50] [156] [191]
[119] [153] [182] [213] [276] [196] [221] [113] [284] [294] [255]
[213] [295] [109] [260] [154] [293] [280] [97] [165]
[113] [165] [79] [232] [292] [289]
[285] [291] [246] [228]

Parameterization [198] [157] [203] [292] [23] [24] [21] [23] [294]
[135] [169] [95] [285]
[43] [90] [186]
[212] [29] [33]
[189]

Reduced Order Models [244] [179] [101] [188] [10] [148] [125] [262] [171] [125] [223] [224] [57]
[273] [225] [286] [74] [188] [190] [257] [101] [190]
[173]

Downscaling [237] [254] [81]

[232]

Uncertainty [275] [142] [250] [270] [281] [266] [276] [162] [58] [277] [266] [67]
quantification [283] [282] [301] [89]

Inverse [59] [47] [161] [210] [122] [27] [77] [240] [111] [192] [105] [121]
modeling [197] [132] [241] [39] [38]
[68] [229]

Discover [30] [227] [37] [207] [155] [251]

Governing [167] [168] [216]
Equations [123] [200][205]
[52] [141] [290]

Solve [147] [65] [140] [209] [234] [279] [48] [218] [44] [166]
PDEs [9] [215] [42] [281] [60] [177] [60] [172] [235]
[264] [106] [235] [208] [231] [301] [19] [51] [131]
[107] [131] [71] [89] [129] [76] [278] [194]
[268] [194] [176] [176] [206]
[16]

Data [187] [78] [62] [40] [28] [49] [274] [231]

Generation [133] [268] [288]
[298]

4 DISCUSSION
Table 1 provides a systematic organization and taxonomy of the objectives and methods of existing
physics-based ML applications. This table provides a convenient organization for papers discussed in
this survey and other papers that could not be discussed because of space limitations. Importantly,
analysis of works within our taxonomy uncovers knowledge gaps and potential crossovers of
methods between disciplines that can serve as ideas for future research.
Indeed, there are a myriad of opportunities for taking ideas across applications, objectives, and
methods, as well as bringing them back to the traditional ML discipline. For example, the physics-
guided NN approach developed for aquatic sciences [109, 120] can be used in any application
where an imperfect mechanistic model is available. Raissi et al. [210] take physics-guided loss
function methods developed for solving PDEs and extend them to inverse modeling problems in
fluid dynamics. Furthermore, Daw et al. [58] develop a monotonicity-preserving architecture for

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 21

lake temperature based on previous work done using loss function terms and a hybrid physics-ML
architecture [128].
Note that some techniques for incorporating physical principles into ML models can be applied
to a wide range of objectives. For example, the physics-guided loss function spans nearly all the
rows of Table 1. Methods also have the potential for overlap, where, for example, physics-guided
loss functions, physics-guided architecture, and physics-guided initialization could all be applied to
an ML model used in a hybrid framework.
From Table 1, it is easy to see that several boxes in the Table 1 are rather sparse or entirely
empty, many of which represent opportunities for future work. For example, ML models for
parameterization are increasingly being used in domains such as climate science and weather
forecasting [135], all of which can benefit from the integration of physical principles. Furthermore,
principles from super-resolution frameworks, originally developed in the context of computer
vision applications, are beginning to be applied to downscaling to create higher resolution climate
predictions [254]. However, most of these do not incorporate physics (e.g., through additional
loss function terms or an informed design of architecture). The fact that nearly all of the other
methods (columns) except for hybrid modeling are applicable to this task shows that there is
tremendous scope for new exploration, where research pursued in these columns in the context
of other objectives can be applied for this objective. We also see a lot of opportunities for new
research on physics-guided initialization, where, for example, pre-training an ML algorithm for
inverse modeling or parameterization on simulation data related to the desired task could improve
performance or lessen data needs.
We note that many boxes in Table 1 may not make sense for the pairing of an objective with
a method. In particular, residual modeling (which is one of the oldest approaches for integrating
physical models with statistical/machine learning models) cannot be naturally used for many
of the objectives. For example, ML methods for solving PDEs often serve the purpose to reduce
computation compared to an existing solving method rather than improve over a physical model,
as in the case of residual modeling. Similarly, for discovering governing equations, there often
isn’t a known physical model to compare to for creating a residual model or for use in a hybrid
approach. Data generation also does not make sense for residual modeling since the purpose is to
simulate a data distribution rather than improve on a physical model.
Not all promising research directions are covered by Table 1 and the previous discussion. For
instance, one promising direction is to forecast future events using ML and continuously updating
model states by incorporating ideas from data assimilation [124]. An instance of this is pursued
by Dua et al. [69] to build an ML model that predicts the parameters of a physical model using
past time series data as an input. Another instance of this approach is seen in epidemiology, where
Magri et al. [165] used a NN for data assimilation to predict a parameter vector that captures the
time evolution of a COVID-19 epidemiological model. Such approaches can benefit from the ideas
in Section 3 (e.g., physics-based loss, intermediate physics variables, etc.).

5 CONCLUDING REMARKS
Given the current deluge of sensor data and advances in ML methods, we envision that the merging
of principles from ML and physics will play an invaluable role in the future of scientific modeling
to address the pressing environmental and physical modeling problems facing society. Although
both physics-based modeling and ML are rapidly advancing fields, knowledge integration for
mutual gain requires clear lines of communication between disciplines. Currently, most of the
diverse set of techniques for bringing ML and physical knowledge together have been developed in
isolated applications by researchers working in distinct disciplines. Our hope is that this survey
will accelerate the cross-pollination of ideas among these diverse research communities.

, Vol. 1, No. 1, Article . Publication date: July 2020.

22 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

The contributions and structure provided in this survey also serve to benefit the ML community,
where, for example, techniques for adding physical constraints to loss functions can be used
to enforce fairness in predictive models, or realism for data generated by GANs. Further, novel
architecture designs can enable new ways to incorporate prior domain information (beyond what
is usually done using Bayesian frameworks) and can lead to better interpretability.
This survey focuses primarily on improving the modeling of engineering and environmental
systems that are traditionally solved used using physics-based modeling. However, the general ideas
of physics-informed ML have wider applicability, and such research is already being pursued in
many other contexts. For example, there are several interesting works in system control which often
involves reinforcement learning techniques (e.g., combining model predictive control with Gaussian
processes in robotics [7], informed priors for neuroscience modeling [91], physics-based reward
functions in computational chemistry [53], and fluidic feedback control from a cylinder [134]).
Other examples include identifying features of interest in the output of computational simulations
of physics-based models (e.g., high-impact weather predictions [175], segmentation of climate
models [181], and tracking phenomena from climate model data [85, 219]). There is also recent
work on encoding domain knowledge for geometric deep learning [34, 41] that is finding increasing
use in computational chemistry [61, 70, 83], physics [18, 45], geostatistics [8, 272], and neuroscience
[136]. We expect that there will be a lot of potential for cross-over of ideas amongst these different
efforts that will greatly expedite research in this nascent field.

REFERENCES
[1] Somak Aditya, Yezhou Yang, and Chitta Baral. 2019. Integrating knowledge and reasoning in image understanding.
arXiv preprint arXiv:1906.09954 (2019).
[2] Mark Alber, Adrian Buganza Tepole, William R Cannon, Suvranu De, Salvador Dura-Bernal, Krishna Garikipati,
George Karniadakis, William W Lytton, Paris Perdikaris, Linda Petzold, et al. 2019. Integrating machine learning
and multiscale modeling—perspectives, challenges, and opportunities in the biological, biomedical, and behavioral
sciences. npj Digital Medicine 2, 1 (2019), 1–11.
[3] Babak Alipanahi, Andrew Delong, Matthew T Weirauch, and Brendan J Frey. 2015. Predicting the sequence specificities
of DNA-and RNA-binding proteins by deep learning. Nature biotechnology 33, 8 (2015), 831–838.
[4] Mauricio Alvarez, David Luengo, and Neil D Lawrence. 2009. Latent force models. In Artificial Intelligence and
Statistics. 9–16.
[5] David Amsallem and Charbel Farhat. 2008. Interpolation method for adapting reduced-order models and application
to aeroelasticity. AIAA journal 46, 7 (2008), 1803–1813.
[6] Brandon Anderson, Truong Son Hy, and Risi Kondor. 2019. Cormorant: Covariant molecular neural networks. In
Advances in Neural Information Processing Systems. 14510–14519.
[7] Olov Andersson, Fredrik Heintz, and Patrick Doherty. 2015. Model-based reinforcement learning in continuous
environments using real-time constrained optimization. In Twenty-Ninth AAAI Conference on Artificial Intelligence.
[8] Gabriel Appleby, Linfeng Liu, and Liping Liu. 2020. Kriging Convolutional Networks.. In AAAI. 3187–3194.
[9] L Arsenault, A Lopez-Bezanilla, O von Lilienfeld, and A Millis. 2014. Machine learning for many-body physics: The
case of the Anderson impurity model. Phys. Rev. B (2014).
[10] Omri Azencot, N Benjamin Erichson, Vanessa Lin, and Michael W Mahoney. 2020. Forecasting Sequential Data using
Consistent Koopman Autoencoders. arXiv preprint arXiv:2003.02236 (2020).
[11] Yunhao Ba, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. 2019. Physics-based neural networks for
shape from polarization. arXiv preprint arXiv:1903.10210 (2019).
[12] Y Ba, G Zhao, and A Kadambi. 2019. Blending diverse physical priors with neural networks. arXiv:1910.00201 (2019).
[13] Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar. 2016. Designing neural network architectures using
reinforcement learning. arXiv preprint arXiv:1611.02167 (2016).
[14] N Baker, F Alexander, T Bremer, A Hagberg, Y Kevrekidis, H Najm, M Parashar, A Patra, J Sethian, S Wild, et al.
2019. Workshop report on basic research needs for scientific machine learning: Core technologies for artificial intelligence.
Technical Report. USDOE Office of Science.
[15] Pierre Baldi, Kevin Bauer, Clara Eng, Peter Sadowski, and Daniel Whiteson. 2016. Jet substructure classification in
high-energy physics with deep neural networks. Physical Review D 93, 9 (2016), 094034.

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 23

[16] Yohai Bar-Sinai, Stephan Hoyer, Jason Hickey, and Michael P Brenner. 2019. Learning data-driven discretizations for
partial differential equations. Proceedings of the National Academy of Sciences 116, 31 (2019), 15344–15349.
[17] E Baseman, N DeBardeleben, S Blanchard, J Moore, O Tkachenko, K Ferreira, T Siddiqua, and V Sridharan. 2018.
Physics-Informed Machine Learning for DRAM Error Modeling. In IEEE DFT.
[18] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. 2016. Interaction networks for learning
about objects, relations and physics. In Advances in neural information processing systems. 4502–4510.
[19] Christian Beck, E Weinan, and Arnulf Jentzen. 2019. Machine learning approximation algorithms for high-dimensional
fully nonlinear partial differential equations and second-order backward stochastic differential equations. Journal of
Nonlinear Science 29, 4 (2019), 1563–1619.
[20] Jörg Behler. 2011. Neural network potential-energy surfaces in chemistry: a tool for large-scale simulations. Physical
Chemistry Chemical Physics 13, 40 (2011), 17930–17955.
[21] Jörg Behler and Michele Parrinello. 2007. Generalized neural-network representation of high-dimensional potential-
energy surfaces. Physical review letters 98, 14 (2007), 146401.
[22] Karianne J Bergen, Paul A Johnson, V Maarten, and Gregory C Beroza. 2019. Machine learning for data-driven
discovery in solid Earth geoscience. Science 363, 6433 (2019), eaau0323.
[23] Tom Beucler, Michael Pritchard, Stephan Rasp, Pierre Gentine, Jordan Ott, and Pierre Baldi. 2019. Enforcing Analytic
Constraints in Neural-Networks Emulating Physical Systems. arXiv preprint arXiv:1909.00912 (2019).
[24] Tom Beucler, Stephan Rasp, Michael Pritchard, and Pierre Gentine. 2019. Achieving Conservation of Energy in Neural
Network Emulators for Climate Modeling. arXiv preprint arXiv:1906.06622 (2019).
[25] Gregory Beylkin, Ronald Coifman, and Vladimir Rokhlin. 1991. Fast wavelet transforms and numerical algorithms I.
Communications on pure and applied mathematics 44, 2 (1991), 141–183.
[26] I Bilionis and N Zabaras. 2012. Multi-output local Gaussian process regression: Applications to uncertainty quantifi-
cation. J. Comput. Phys (2012).
[27] Reetam Biswas, Mrinal K Sen, Vishal Das, and Tapan Mukerji. 2019. Prestack and poststack inversion using a
physics-guided convolutional neural network. Interpretation 7, 3 (2019), SE161–SE174.
[28] Mathis Bode, Michael Gauding, Zeyu Lian, Dominik Denker, Marco Davidovic, Konstantin Kleinheinz, Jenia Jitsev,
and Heinz Pitsch. 2019. Using Physics-Informed Super-Resolution Generative Adversarial Networks for Subgrid
Modeling in Turbulent Reactive Flows. arXiv preprint arXiv:1911.11380 (2019).
[29] Thomas Bolton and Laure Zanna. 2019. Applications of deep learning to ocean data inference and subgrid parameter-
ization. Journal of Advances in Modeling Earth Systems 11, 1 (2019), 376–399.
[30] J Bongard and H Lipson. 2007. Automated reverse engineering of nonlinear dynamical systems. PNAS (2007).
https://doi.org/10.1073/pnas.0609476104
[31] Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs,
Julian Ibarz, Peter Pastor, Kurt Konolige, et al. 2018. Using simulation and domain adaptation to improve efficiency of
deep robotic grasping. In 2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 4243–4250.
[32] Noah D Brenowitz and Christopher S Bretherton. 2018. Prognostic validation of a neural network unified physics
parameterization. Geophysical Research Letters 45, 12 (2018), 6289–6298.
[33] Noah D Brenowitz and Christopher S Bretherton. 2019. Spatially Extended Tests of a Neural Network Parametrization
Trained by Coarse-Graining. Journal of Advances in Modeling Earth Systems 11, 8 (2019), 2728–2744.
[34] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. 2017. Geometric deep
learning: going beyond euclidean data. IEEE Signal Processing Magazine 34, 4 (2017), 18–42.
[35] M Brown, D Lary, A Vrieling, D Stathakis, and H Mussa. 2008. Neural networks as a tool for constructing continuous
NDVI time series from AVHRR and MODIS. Int. J. Remote Sens (2008).
[36] Steven Brunton, Eurika Kaiser, and Nathan Kutz. 2017. Koopman operator theory: Past, present, and future. APS
(2017), L27–004.
[37] S Brunton, J Proctor, and J Kutz. 2016. Discovering governing equations from data by sparse identification of nonlinear
dynamical systems. PNAS (2016).
[38] T A Bubba, G Kutyniok, M Lassas, M Maerz, W Samek, S Siltanen, and V Srinivasan. 2019. Learning the invisible: A
hybrid deep learning-shearlet framework for limited angle computed tomography. Inverse Problems (2019).
[39] Gustau Camps-Valls, Luca Martino, Daniel H Svendsen, Manuel Campos-Taberner, Jordi Muñoz-Marí, Valero Laparra,
David Luengo, and Francisco Javier García-Haro. 2018. Physics-aware Gaussian processes in remote sensing. Applied
Soft Computing 68 (2018), 69–82.
[40] Ruijin Cang, Hechao Li, Hope Yao, Yang Jiao, and Yi Ren. 2018. Improving direct physical properties prediction of
heterogeneous materials from imaging data via convolutional neural network and a morphology-aware generative
model. Computational Materials Science 150 (2018), 212–221.
[41] Wenming Cao, Zhiyue Yan, Zhiquan He, and Zhihai He. 2020. A comprehensive survey on geometric deep learning.
IEEE Access 8 (2020), 35929–35949.

, Vol. 1, No. 1, Article . Publication date: July 2020.

24 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

[42] Giuseppe Carleo and Matthias Troyer. 2017. Solving the quantum many-body problem with artificial neural networks.
Science 355, 6325 (2017), 602–606.
[43] S Chan and A Elsheikh. 2017. Parametrization and generation of geological models with generative adversarial
networks. arXiv:1708.01810 (2017).
[44] B Chang, M Chen, E Haber, and E Chi. 2019. AntisymmetricRNN: A dynamical system view on recurrent neural
networks. arXiv:1902.09689 (2019).
[45] Michael B Chang, Tomer Ullman, Antonio Torralba, and Joshua B Tenenbaum. 2016. A compositional object-based
approach to learning physical dynamics. arXiv preprint arXiv:1612.00341 (2016).
[46] Gang Chen, Yingtao Zuo, Jian Sun, and Yueming Li. 2012. Support-vector-machine-based reduced-order model for
limit cycle oscillation prediction of nonlinear aeroelastic system. Mathematical problems in engineering 2012 (2012).
[47] H Chen, Y Zhang, M Kalra, F Lin, Y Chen, P Liao, J Zhou, and G Wang. 2017. Low-dose CT with a residual encoder-
decoder convolutional neural network. IEEE T-MI (2017).
[48] TQ Chen, Y Rubanova, J Bettencourt, and D K Duvenaud. 2018. Neural ordinary differential equations. In NIPS.
[49] W Chen and M Fuge. 2018. BezierGAN: Automatic Generation of Smooth Curves from Interpretable Low-Dimensional
Parameters. arXiv:1808.08871 (2018).
[50] X Chen, X Xu, X Liu, S Pan, J He, HY Noh, L Zhang, and P Zhang. 2018. Pga: Physics guided and adaptive approach
for mobile fine-grained air pollution estimation. In Ubicomp.
[51] Zhengdao Chen, Jianyu Zhang, Martin Arjovsky, and Léon Bottou. 2019. Symplectic Recurrent Neural Networks.
arXiv preprint arXiv:1909.13334 (2019).
[52] In Ho Cho, Qiang Li, Rana Biswas, and Jaeyoun Kim. 2020. A framework for glass-box physics rule learner and its
application to nano-scale phenomena. Communications Physics 3, 1 (2020), 1–9.
[53] Youngwoo Cho, Sookyung Kim, Peggy Pk Li, Mike P Surh, T Yong-Jin Han, and Jaegul Choo. 2019. Physics-guided
Reinforcement Learning for 3D Molecular Structures. In Workshop at the 33rd Conference on Neural Information
Processing Systems (NeurIPS).
[54] Anshul Choudhary, John F Lindner, Elliott G Holliday, Scott T Miller, Sudeshna Sinha, and William L Ditto. 2019.
Physics enhanced neural networks predict order and chaos. arXiv preprint arXiv:1912.01958 (2019).
[55] M Christie, V Demyanov, and D Erbas. 2006. Uncertainty quantification for porous media flows. J. Comput. Phys
(2006).
[56] Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. 2018. Spherical cnns. arXiv preprint arXiv:1801.10130
(2018).
[57] Thomas Daniel, Fabien Casenave, Nissrine Akkari, and David Ryckelynck. 2020. Model order reduction assisted by
deep neural networks (ROM-net). Advanced Modeling and Simulation in Engineering Sciences 7 (2020), 1–27.
[58] A Daw, RQ Thomas, CC Carey, JS Read, AP Appling, and A Karpatne. 2019. Physics-Guided Architecture (PGA) of
Neural Networks for Quantifying Uncertainty in Lake Temperature Modeling. arXiv:1911.02682 (2019).
[59] MS Dawson, J Olvera, AK Fung, and MT Manry. 1992. Inversion of surface parameters using fast learning neural
networks. IGARSS International Geoscience and Remote Sensing Symposium (1992).
[60] Emmanuel de Bezenac, Arthur Pajot, and Patrick Gallinari. 2019. Deep learning for physical processes: Incorporating
prior scientific knowledge. Journal of Statistical Mechanics: Theory and Experiment 2019, 12 (2019), 124009.
[61] Nicola De Cao and Thomas Kipf. 2018. MolGAN: An implicit generative model for small molecular graphs. arXiv
preprint arXiv:1805.11973 (2018).
[62] L de Oliveira, M Paganini, and B Nachman. 2017. Learning particle physics by example: location-aware generative
adversarial networks for physics synthesis. Comput. Softw. Big Sci (2017).
[63] EL Denton, S Chintala, R Fergus, et al. 2015. Deep generative image models using a laplacian pyramid of adversarial
networks. In NIPS.
[64] C Deser, A Phillips, V Bourdette, and H Teng. 2012. Uncertainty in climate change projections: the role of internal
variability. Climate dynamics (2012).
[65] MWMG Dissanayake and N Phan-Thien. 1994. Neural-network-based approximations for solving partial differential
equations. communications in Numerical Methods in Engineering 10, 3 (1994), 195–201.
[66] NAK Doan, W Polifke, and L Magri. 2019. Physics-informed echo state networks for chaotic systems forecasting. In
ICCS. Springer.
[67] B Dong, Z Li, SMM Rahman, and R Vega. 2016. A hybrid model approach for forecasting future residential electricity
consumption. Energy and Buildings (2016).
[68] Jonathan E Downton and Daniel P Hampson. 2019. Use of theory-guided neural networks to perform seismic
inversion. In GeoSoftware, CGG.
[69] Vivek Dua. 2011. An artificial neural network approximation based decomposition approach for parameter estimation
of system of ordinary differential equations. Computers & chemical engineering 35, 3 (2011), 545–553.

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 25

[70] David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alán Aspuru-Guzik,
and Ryan P Adams. 2015. Convolutional networks on graphs for learning molecular fingerprints. In Advances in
neural information processing systems. 2224–2232.
[71] Vikas Dwivedi and Balaji Srinivasan. 2020. Solution of Biharmonic Equation in Complicated Geometries with Physics
Informed Extreme Learning Machine. Journal of Computing and Information Science in Engineering (2020), 1–10.
[72] Sašo Džeroski, Ljupčo Todorovski, Ivan Bratko, Boris Kompare, and Viljem Križman. 1999. Equation discovery with
ecological applications. In Machine learning methods for ecological applications. Springer, 185–207.
[73] T Elsken, JH Metzen, and F Hutter. 2018. Neural architecture search: A survey. arXiv:1808.05377 (2018).
[74] N Benjamin Erichson, Michael Muehlebach, and Michael W Mahoney. 2019. Physics-informed autoencoders for
lyapunov-stable fluid flow prediction. arXiv preprint arXiv:1905.10866 (2019).
[75] James H Faghmous and Vipin Kumar. 2014. A big data guide to understanding climate change: The case for theory-
guided data science. Big data 2, 3 (2014), 155–163.
[76] Yuwei Fan, Lin Lin, Lexing Ying, and Leonardo Zepeda-Núnez. 2019. A multiscale neural network based on hierarchical
matrices. Multiscale Modeling & Simulation 17, 4 (2019), 1189–1213.
[77] Yuwei Fan and Lexing Ying. 2020. Solving electrical impedance tomography with deep learning. J. Comput. Phys. 404
(2020), 109119.
[78] Amir Barati Farimani, Joseph Gomes, and Vijay S Pande. 2017. Deep learning the physics of transport phenomena.
arXiv preprint arXiv:1709.02432 (2017).
[79] Ferdinando Fioretto, Terrence WK Mak, and Pascal Van Hentenryck. 2020. Predicting AC optimal power flows:
Combining deep learning and lagrangian dual methods. In Proceedings of the AAAI Conference on Artificial Intelligence,
Vol. 34. 630–637.
[80] J Fish and T Belytschko. 2007. A first course in finite elements. Wiley.
[81] Christian Folberth, Artem Baklanov, Juraj Balkovič, Rastislav Skalskỳ, Nikolay Khabarov, and Michael Obersteiner.
2019. Spatio-temporal downscaling of gridded crop model yield estimates based on machine learning. Agricultural
and forest meteorology 264 (2019), 1–15.
[82] Urban Forssell and Peter Lindskog. 1997. Combining Semi-Physical and Neural Network Modeling: An Example ofIts
Usefulness. IFAC Proceedings Volumes 30, 11 (1997), 767–770.
[83] Alex Fout, Jonathon Byrd, Basir Shariat, and Asa Ben-Hur. 2017. Protein interface prediction using graph convolutional
networks. In Advances in neural information processing systems. 6530–6539.
[84] Michalis Frangos, Youssef Marzouk, Karen Willcox, and B van Bloemen Waanders. 2010. Surrogate and reduced-order
modeling: A comparison of approaches for large-scale statistical inverse problems. Large-scale inverse problems and
quantification of uncertainty 123149 (2010).
[85] David John Gagne, Amy McGovern, Sue Ellen Haupt, Ryan A Sobash, John K Williams, and Ming Xue. 2017. Storm-
based probabilistic hail forecasting with machine learning applied to convection-allowing ensembles. Weather and
forecasting 32, 5 (2017), 1819–1840.
[86] Y Gal and Z Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep
learning. In ICML.
[87] D Galbally, K Fidkowski, K Willcox, and O Ghattas. 2010. Non-linear model reduction for uncertainty quantification
in large-scale inverse problems. Int. J. Numer. Methods Fluids (2010).
[88] Pei Gao, Antti Honkela, Magnus Rattray, and Neil D Lawrence. 2008. Gaussian process modelling of latent chemical
species: applications to inferring transcription factor activities. Bioinformatics 24, 16 (2008), i70–i75.
[89] N Geneva and N Zabaras. 2020. Modeling the dynamics of PDE systems with physics-constrained deep auto-regressive
networks. J. Comput. Phys (2020).
[90] P Gentine, M Pritchard, S Rasp, G Reinaudi, and G Yacalis. 2018. Could machine learning break the convection
parameterization deadlock? Geophys. Res. Lett (2018).
[91] Samuel J Gershman. 2016. Empirical priors for reinforcement learning models. Journal of Mathematical Psychology
71 (2016), 1–6.
[92] Donald Gerwin. 1974. Information processing, data inferences, and scientific generalization. Behavioral Science 19, 5
(1974), 314–325.
[93] Dimitrios Giannakis, Joanna Slawinska, and Zhizhen Zhao. 2015. Spatiotemporal feature extraction with data-driven
Koopman operators. In Feature Extraction: Modern Questions and Challenges. 103–115.
[94] Aldo Glielmo, Peter Sollich, and Alessandro De Vita. 2017. Accurate interatomic force fields via machine learning
with covariant kernels. Physical Review B 95, 21 (2017), 214302.
[95] EB Goldstein, G Coco, AB Murray, and MO Green. 2014. Data-driven components in a model of inner-shelf sorted
bedforms: a new hybrid model. Earth Surface Dynamics (2014).
[96] Gene H Golub and Charles F Van Loan. 2012. Matrix computations. Vol. 3. JHU press.

, Vol. 1, No. 1, Article . Publication date: July 2020.

26 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

[97] Noel P Greis, Monica L Nogueira, Sambit Bhattacharya, and Tony Schmitz. 2020. Physics-Guided Machine Learning
for Self-Aware Machining. In AAAI Spring Symposium – AI and Manufacturing.
[98] S Greydanus, M Dzamba, and J Yosinski. 2019. Hamiltonian neural networks. In NIPS.
[99] Aditya Grover, Ashish Kapoor, and Eric Horvitz. 2015. A deep hybrid model for weather forecasting. In Proceedings of
the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 379–386.
[100] Jiaxian Guo, Sidi Lu, Han Cai, Weinan Zhang, Yong Yu, and Jun Wang. 2018. Long text generation via adversarial
training with leaked information. In Thirty-Second AAAI Conference on Artificial Intelligence.
[101] Mengwu Guo and Jan S Hesthaven. 2019. Data-driven reduced order modeling for time-dependent problems. Computer
Methods in Applied Mechanics and Engineering 345 (2019), 75–99.
[102] Xiaoxiao Guo, Wei Li, and Francesco Iorio. 2016. Convolutional neural networks for steady flow approximation. In
Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 481–490.
[103] HV Gupta et al. 2014. Debates—The future of hydrological sciences: A (common) path forward? Using models and
data to learn: A systems theoretic perspective on the future of hydrological science. Water Resources Research (2014).
[104] Yoo-Geun Ham, Jeong-Hwan Kim, and Jing-Jia Luo. 2019. Deep learning for multi-year ENSO forecasts. Nature 573,
7775 (2019), 568–572.
[105] F Hamilton, AL Lloyd, and KB Flores. 2017. Hybrid modeling and prediction of dynamical systems. PLOS Comput.
Biol (2017).
[106] J Han, A Jentzen, and E Weinan. 2018. Solving high-dimensional partial differential equations using deep learning.
PNAS (2018).
[107] Jiequn Han, Linfeng Zhang, and E Weinan. 2019. Solving many-electron Schrödinger equation using deep neural
networks. J. Comput. Phys. 399 (2019), 108929.
[108] Katja Hansen, Grégoire Montavon, Franziska Biegler, Siamac Fazli, Matthias Rupp, Matthias Scheffler, O Anatole
Von Lilienfeld, Alexandre Tkatchenko, and Klaus-Robert Muller. 2013. Assessment and validation of machine learning
methods for predicting molecular atomization energies. Journal of Chemical Theory and Computation 9, 8 (2013),
3404–3419.
[109] Paul C Hanson, Aviah B Stillman, Xiaowei Jia, Anuj Karpatne, Hilary A Dugan, Cayelan C Carey, Joseph Stachelek,
Nicole K Ward, Yu Zhang, Jordan S Read, et al. 2020. Predicting lake surface water phosphorus dynamics using
process-guided machine learning. Ecological Modelling 430 (2020), 109136.
[110] K He, X Zhang, S Ren, and J Sun. 2016. Deep residual learning for image recognition. In CVPR.
[111] Jonathan R Holland, James D Baeder, and Karthikeyan Duraisamy. 2019. Field inversion and machine learning with
embedded neural networks: Physics-consistent neural network training. In AIAA Aviation 2019 Forum. 3200.
[112] William W Hsieh. 2009. Machine learning methods in the environmental sciences: Neural networks and kernels.
Cambridge university press, Cambridge, United Kingdom.
[113] X Hu, H Hu, S Verma, and ZL Zhang. 2020. Physics-Guided Deep Neural Networks for PowerFlow Analysis.
arXiv:2002.00097 (2020).
[114] David Menéndez Hurtado, Karolis Uziela, and Arne Elofsson. 2018. Deep transfer learning in the assessment of the
quality of protein models. arXiv preprint arXiv:1804.06281 (2018).
[115] Frank Hutter, Lars Kotthoff, and Joaquin Vanschoren. 2019. Automated machine learning: methods, systems, challenges.
Springer Nature.
[116] Gabriel Ibarra-Berastegi, Jon Saénz, Ganix Esnaola, Agustin Ezcurra, and Alain Ulazia. 2015. Short-term forecasting of
the wave energy flux: Analogues, random forests, and physics-based models. Ocean Engineering 104 (2015), 530–539.
[117] Željko Ivezić, Andrew J Connolly, Jacob T VanderPlas, and Alexander Gray. 2019. Statistics, data mining, and machine
learning in astronomy: a practical Python guide for the analysis of survey data. Princeton University Press, Princeton,
NJ.
[118] Xiaowei Jia, Ankush Khandelwal, David J Mulla, Philip G Pardey, and Vipin Kumar. 2019. Bringing automated,
remote-sensed, machine learning methods to monitoring crop landscapes at scale. Agricultural Economics 50 (2019),
41–50.
[119] X Jia, J Willard, A Karpatne, J Read, J Zwart, M Steinbach, and V Kumar. 2019. Physics Guided RNNs for Modeling
Dynamical Systems: A Case Study in Simulating Lake Temperature Profiles. In Proceedings of SIAM International
Conference on Data Mining.
[120] Xiaowei Jia, Jared Willard, Anuj Karpatne, Jordan S Read, Jacob A Zwart, Michael Steinbach, and Vipin Kumar. 2020.
Physics-guided machine learning for scientific discovery: An application in simulating lake temperature profiles.
arXiv preprint arXiv:2001.11086 (2020).
[121] KH Jin, MT McCann, E Froustey, and M Unser. 2017. Deep convolutional neural network for inverse problems in
imaging. IEEE Trans. Image Process (2017).
[122] Adar Kahana, Eli Turkel, Shai Dekel, and Dan Givoli. 2020. Obstacle segmentation based on the wave equation and
deep learning. J. Comput. Phys. (2020), 109458.

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 27

[123] Eurika Kaiser, J Nathan Kutz, and Steven L Brunton. 2018. Sparse identification of nonlinear dynamics for model
predictive control in the low-data limit. Proceedings of the Royal Society A 474, 2219 (2018), 20180335.
[124] Eugenia Kalnay. 2003. Atmospheric modeling, data assimilation and predictability. Cambridge university press.
[125] J Kani and A Elsheikh. 2017. DR-RNN: A deep residual recurrent neural network for model reduction. arXiv:1709.00939
(2017).
[126] Anuj Karpatne, Gowtham Atluri, James H Faghmous, Michael Steinbach, Arindam Banerjee, Auroop Ganguly, Shashi
Shekhar, Nagiza Samatova, and Vipin Kumar. 2017. Theory-guided data science: A new paradigm for scientific
discovery from data. IEEE Transactions on knowledge and data engineering 29, 10 (2017), 2318–2331.
[127] Anuj Karpatne, Imme Ebert-Uphoff, Sai Ravela, Hassan Ali Babaie, and Vipin Kumar. 2018. Machine learning for
the geosciences: Challenges and opportunities. IEEE Transactions on Knowledge and Data Engineering 31, 8 (2018),
1544–1554.
[128] A Karpatne, W Watkins, J Read, and V Kumar. 2017. Physics-guided Neural Networks (PGNN): An Application in
Lake Temperature Modeling. arXiv:1710.11431 (2017).
[129] S Karumuri, R Tripathy, I Bilionis, and J Panchal. 2020. Simulator-free solution of high-dimensional stochastic elliptic
partial differential equations using deep neural networks. J. Comput. Phys (2020).
[130] Steven K Kauwe, Jake Graser, Antonio Vazquez, and Taylor D Sparks. 2018. Machine learning prediction of heat
capacity for solid inorganics. Integrating Materials and Manufacturing Innovation 7, 2 (2018), 43–51.
[131] Yuehaw Khoo, Jianfeng Lu, and Lexing Ying. 2019. Solving for high-dimensional committor functions using artificial
neural networks. Research in the Mathematical Sciences 6, 1 (2019), 1.
[132] Yuehaw Khoo and Lexing Ying. 2019. SwitchNet: a neural network model for forward and inverse scattering problems.
SIAM Journal on Scientific Computing 41, 5 (2019), A3182–A3201.
[133] Byungsoo Kim, Vinicius C Azevedo, Nils Thuerey, Theodore Kim, Markus Gross, and Barbara Solenthaler. 2019. Deep
fluids: A generative network for parameterized fluid simulations. In Computer Graphics Forum, Vol. 38. Wiley Online
Library, 59–70.
[134] Hiroshi Koizumi, Seiji Tsutsumi, and Eiji Shima. 2018. Feedback control of Karman vortex shedding from a cylinder
using deep reinforcement learning. In 2018 Flow Control Conference. 3691.
[135] Vladimir M Krasnopolsky and Michael S Fox-Rabinovitz. 2006. Complex hybrid models combining deterministic and
machine learning components for numerical climate modeling and weather prediction. Neural Networks 19, 2 (2006),
122–134.
[136] Sofia Ira Ktena, Sarah Parisot, Enzo Ferrante, Martin Rajchl, Matthew Lee, Ben Glocker, and Daniel Rueckert.
2017. Distance metric learning using graph convolutional networks: Application to functional brain networks. In
International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 469–477.
[137] Siddhant Kumar, Stephanie Tan, Li Zheng, and Dennis M Kochmann. 2020. Inverse-designed spinodoid metamaterials.
npj Computational Materials 6, 1 (2020), 1–10.
[138] J Nathan Kutz. 2017. Deep learning in fluid dynamics. Journal of Fluid Mechanics 814 (2017), 1–4.
[139] L’ubor Ladickỳ, SoHyeon Jeong, Barbara Solenthaler, Marc Pollefeys, and Markus Gross. 2015. Data-driven fluid
simulations using regression forests. ACM Transactions on Graphics (TOG) 34, 6 (2015), 1–9.
[140] IE Lagaris, A Likas, and DI Fotiadis. 1998. Artificial neural networks for solving ordinary and partial differential
equations. IEEE Trans. Neural Netw. Learn (1998).
[141] John H Lagergren, John T Nardini, G Michael Lavigne, Erica M Rutter, and Kevin B Flores. 2020. Learning partial
differential equations for biological transport models from noisy spatio-temporal data. Proceedings of the Royal Society
A 476, 2234 (2020), 20190800.
[142] B Lakshminarayanan, A Pritzel, and C Blundell. 2017. Simple and scalable predictive uncertainty estimation using
deep ensembles. In NIPS.
[143] Pat Langley. 1981. Data-driven discovery of physical laws. Cognitive Science 5, 1 (1981), 31–54.
[144] Pat Langley, Gary L Bradshaw, and Herbert A Simon. 1983. Rediscovering chemistry with the BACON system. In
Machine learning. Springer, 307–329.
[145] Toni Lassila, Andrea Manzoni, Alfio Quarteroni, and Gianluigi Rozza. 2014. Model order reduction in fluid dynamics:
challenges and perspectives. In Reduced Order Methods for modeling and computational reduction. Springer, 235–273.
[146] Neil D Lawrence, Guido Sanguinetti, and Magnus Rattray. 2007. Modelling transcriptional regulation using Gaussian
processes. In Advances in Neural Information Processing Systems. 785–792.
[147] Hyuk Lee and In Seok Kang. 1990. Neural algorithm for solving differential equations. J. Comput. Phys. 91, 1 (1990),
110–131.
[148] Kookjin Lee and Kevin T Carlberg. 2020. Model reduction of dynamical systems on nonlinear manifolds using deep
convolutional autoencoders. J. Comput. Phys. 404 (2020), 108973.
[149] Douglas B Lenat. 1983. The role of heuristics in learning by discovery: Three case studies. In Machine learning.
Springer, 243–306.

, Vol. 1, No. 1, Article . Publication date: July 2020.

28 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

[150] Qianxiao Li, Felix Dietrich, Erik M Bollt, and Ioannis G Kevrekidis. 2017. Extended dynamic mode decomposition
with dictionary learning: A data-driven adaptive spectral decomposition of the Koopman operator. Chaos: An
Interdisciplinary Journal of Nonlinear Science 27, 10 (2017), 103111.
[151] T Warren Liao and Guoqiang Li. 2020. Metaheuristic-based inverse design of materials–A survey. Journal of
Materiomics (2020).
[152] J Ling, A Kurzawski, and J Templeton. 2016. Reynolds averaged turbulence modelling using deep neural networks
with embedded invariance. J. Fluid Mech (2016).
[153] Dehao Liu and Yan Wang. 2019. Multi-Fidelity Physics-Constrained Neural Network and Its Application in Materials
Modeling. Journal of Mechanical Design 141, 12 (2019).
[154] Yuying Liu, Colin Ponce, Steven L Brunton, and J Nathan Kutz. 2020. Multiresolution Convolutional Autoencoders.
arXiv preprint arXiv:2004.04946 (2020).
[155] Jean-Christophe Loiseau and Steven L Brunton. 2018. Constrained sparse Galerkin regression. Journal of Fluid
Mechanics 838 (2018), 42–67.
[156] Yun Long, Xueyuan She, and Saibal Mukhopadhyay. 2018. HybridNet: integrating model-based and data-driven
learning to predict evolution of dynamical systems. arXiv preprint arXiv:1806.07439 (2018).
[157] Sönke Lorenz, Axel Groß, and Matthias Scheffler. 2004. Representing high-dimensional potential-energy surfaces for
reactions at surfaces by neural networks. Chemical Physics Letters 395, 4-6 (2004), 210–215.
[158] Junde Lu and Furong Gao. 2008. Model migration with inclusive similarity for development of a new process model.
Industrial & engineering chemistry research 47, 23 (2008), 9508–9516.
[159] Junde Lu, Ke Yao, and Furong Gao. 2009. Process similarity and developing new process models through migration.
AIChE journal 55, 9 (2009), 2318–2328.
[160] David Luengo, Manuel Campos-Taberner, and Gustau Camps-Valls. 2016. Latent force models for earth observation
time series prediction. In 2016 IEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP).
IEEE, 1–6.
[161] S Lunz, O Öktem, and CB Schönlieb. 2018. Adversarial regularizers in inverse problems. In NIPS.
[162] Linkai Luo, Yuan Yao, and Furong Gao. 2015. Bayesian improved model migration methodology for fast process
modeling by incorporating prior information. Chemical Engineering Science 134 (2015), 23–35.
[163] Bethany Lusch, J Nathan Kutz, and Steven L Brunton. 2018. Deep learning for universal linear embeddings of
nonlinear dynamics. Nature communications 9, 1 (2018), 4950.
[164] D JC MacKay. 1992. A practical Bayesian framework for backpropagation networks. Neural computation (1992).
[165] Luca Magri and Nguyen Anh Khoa Doan. 2020. First-principles machine learning modelling of COVID-19. arXiv
preprint arXiv:2004.09478 (2020).
[166] A Malek and R Shekari Beidokhti. 2006. Numerical solution for high order differential equations using a hybrid
neural network—optimization method. Appl. Math. Comput. 183, 1 (2006), 260–271.
[167] Niall M Mangan, Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. 2016. Inferring biological networks by sparse
identification of nonlinear dynamics. IEEE Transactions on Molecular, Biological and Multi-Scale Communications 2, 1
(2016), 52–63.
[168] Niall M Mangan, J Nathan Kutz, Steven L Brunton, and Joshua L Proctor. 2017. Model selection for dynamical
systems via sparse regression and information criteria. Proceedings of the Royal Society A: Mathematical, Physical and
Engineering Sciences 473, 2204 (2017), 20170009.
[169] Sergei Manzhos and Tucker Carrington Jr. 2006. A random-sampling high dimensional model representation neural
network for building potential energy surfaces. The Journal of chemical physics 125, 8 (2006), 084109.
[170] A Manzoni, S Pagani, and T Lassila. 2016. Accurate solution of Bayesian inverse uncertainty quantification problems
combining reduced basis methods and reduction error models. SIAM/ASA JUQ (2016).
[171] Andreas Mardt, Luca Pasquali, Hao Wu, and Frank Noé. 2018. VAMPnets for deep learning of molecular kinetics.
Nature communications 9, 1 (2018), 1–11.
[172] Marios Mattheakis, P Protopapas, D Sondak, M Di Giovanni, and E Kaxiras. 2019. Physical symmetries embedded in
neural networks. arXiv preprint arXiv:1904.08991 (2019).
[173] Romit Maulik, Arvind Mohan, Bethany Lusch, Sandeep Madireddy, Prasanna Balaprakash, and Daniel Livescu. 2020.
Time-series learning of latent-space dynamics for reduced-order model closure. Physica D: Nonlinear Phenomena 405
(2020), 132368.
[174] Michael T McCann, Kyong Hwan Jin, and Michael Unser. 2017. A review of convolutional neural networks for inverse
problems in imaging. arXiv preprint arXiv:1710.04011 (2017).
[175] Amy McGovern, Kimberly L Elmore, David John Gagne, Sue Ellen Haupt, Christopher D Karstens, Ryan Lagerquist,
Travis Smith, and John K Williams. 2017. Using artificial intelligence to improve real-time decision-making for
high-impact weather. Bulletin of the American Meteorological Society 98, 10 (2017), 2073–2090.

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 29

[176] Xuhui Meng and George Em Karniadakis. 2020. A composite neural network that learns from multi-fidelity data:
Application to function approximation and inverse PDE problems. J. Comput. Phys. 401 (2020), 109020.
[177] X Meng, Z Li, D Zhang, and GE Karniadakis. 2019. Ppinn: Parareal physics-informed neural network for time-
dependent pdes. arXiv:1909.10145 (2019).
[178] Igor Mezić. 2005. Spectral properties of dynamical systems, model reduction and decompositions. Nonlinear Dynamics
41, 1-3 (2005), 309–325.
[179] Arvind T Mohan and Datta V Gaitonde. 2018. A deep learning based approach to reduced order modeling for turbulent
flow control using LSTM neural networks. arXiv preprint arXiv:1804.09269 (2018).
[180] Jeremy Morton, Freddie D Witherden, and Mykel J Kochenderfer. 2019. Deep variational koopman models: Inferring
koopman observations for uncertainty-aware dynamics modeling and control. arXiv preprint arXiv:1902.09742 (2019).
[181] Mayur Mudigonda, Sookyung Kim, Ankur Mahesh, Samira Kahou, Karthik Kashinath, Dean Williams, Vincen
Michalski, Travis O’Brien, and Mr Prabhat. 2017. Segmenting and tracking extreme climate events using neural
networks. In Deep Learning for Physical Sciences (DLPS) Workshop, held with NIPS Conference.
[182] Nikhil Muralidhar, Jie Bu, Ze Cao, Long He, Naren Ramakrishnan, Danesh Tafti, and Anuj Karpatne. 2020. PhyNet:
Physics Guided Neural Networks for Particle Drag Force Prediction in Assembly. In Proceedings of the 2020 SIAM
International Conference on Data Mining. SIAM, 559–567.
[183] N Muralidhar, MR Islam, M Marwah, A Karpatne, and N Ramakrishnan. 2018. Incorporating prior domain knowledge
into deep neural networks. In IEEE Big Data. IEEE.
[184] Frank Noé, Alexandre Tkatchenko, Klaus-Robert Müller, and Cecilia Clementi. 2020. Machine learning for molecular
simulation. Annual review of physical chemistry 71 (2020), 361–390.
[185] Peer Nowack, Peter Braesicke, Joanna Haigh, Nathan Luke Abraham, John Pyle, and Apostolos Voulgarakis. 2018.
Using machine learning to build temperature-based ozone parameterizations for climate sensitivity simulations.
Environmental Research Letters 13, 10 (2018), 104016.
[186] Paul A O’Gorman and John G Dwyer. 2018. Using machine learning to parameterize moist convection: Potential for
modeling of climate, climate change, and extreme events. Journal of Advances in Modeling Earth Systems 10, 10 (2018),
2548–2563.
[187] A Oord, S Dieleman, H Zen, K Simonyan, O Vinyals, A Graves, N Kalchbrenner, A Senior, and K Kavukcuoglu. 2016.
Wavenet: A generative model for raw audio. arXiv:1609.03499 (2016).
[188] Samuel E Otto and Clarence W Rowley. 2019. Linearly recurrent autoencoder networks for learning dynamics. SIAM
Journal on Applied Dynamical Systems 18, 1 (2019), 558–593.
[189] Anikesh Pal. [n.d.]. Deep learning emulation of subgrid scale processes in turbulent shear flows. Geophysical Research
Letters ([n. d.]), e2020GL087005.
[190] Shaowu Pan and Karthik Duraisamy. 2020. Physics-informed probabilistic learning of linear embeddings of nonlinear
dynamics with guaranteed stability. SIAM Journal on Applied Dynamical Systems 19, 1 (2020), 480–509.
[191] R Paolucci, F Gatti, M Infantino, C Smerzini, AG Özcebe, and M Stupazzini. 2018. Broadband Ground Motions from
3D Physics-Based Numerical Simulations Using Artificial Neural NetworksBroadband Ground Motions from 3D PBSs
Using ANNs. BSSA (2018).
[192] EJ Parish and K Duraisamy. 2016. A paradigm for data-driven predictive modeling using field inversion and machine
learning. J. Comput. Phys (2016).
[193] Junyoung Park and Jinkyoo Park. 2019. Physics-induced graph neural network: An application to wind-farm power
estimation. Energy 187 (2019), 115883.
[194] Wei Peng, Weien Zhou, Jun Zhang, and Wen Yao. 2020. Accelerating Physics-Informed Neural Network Training
with Prior Dictionaries. arXiv preprint arXiv:2004.08151 (2020).
[195] CL Pettit. 2004. Uncertainty quantification in aeroelasticity: recent results and research challenges. Journal of Aircraft
(2004).
[196] David Pfau, James S Spencer, Alexander G de G Matthews, and WMC Foulkes. 2019. Ab-Initio Solution of the
Many-Electron Schr\" odinger Equation with Deep Neural Networks. arXiv preprint arXiv:1909.02487 (2019).
[197] L Pilozzi, FA Farrelly, G Marcucci, and C Conti. 2018. Machine learning inverse problem for topological photonics.
Communications Physics (2018).
[198] Frederico V Prudente and JJ Soares Neto. 1998. The fitting of potential energy surfaces using neural networks.
Application to the study of the photodissociation processes. Chemical physics letters 287, 5-6 (1998), 585–589.
[199] A Pukrittayakamee, M Malshe, M Hagan, LM Raff, R Narulkar, S Bukkapatnum, and R Komanduri. 2009. Simultaneous
fitting of a potential-energy surface and its corresponding force fields using feedforward neural networks. J. Chem.
Phys (2009).
[200] Markus Quade, Markus Abel, J Nathan Kutz, and Steven L Brunton. 2018. Sparse identification of nonlinear dynamics
for rapid model recovery. Chaos: An Interdisciplinary Journal of Nonlinear Science 28, 6 (2018), 063116.

, Vol. 1, No. 1, Article . Publication date: July 2020.

30 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

[201] Alfio Quarteroni, Gianluigi Rozza, et al. 2014. Reduced order methods for modeling and computational reduction. Vol. 9.
Springer.
[202] Paul Raccuglia, Katherine C Elbert, Philip DF Adler, Casey Falk, Malia B Wenny, Aurelio Mollo, Matthias Zeller,
Sorelle A Friedler, Joshua Schrier, and Alexander J Norquist. 2016. Machine-learning-assisted materials discovery
using failed experiments. Nature 533, 7601 (2016), 73–76.
[203] LM Raff, M Malshe, M Hagan, DI Doughan, MG Rockley, and R Komanduri. 2005. Ab initio potential-energy surfaces
for complex, multichannel systems using modified novelty sampling and feedforward neural networks. The Journal
of chemical physics 122, 8 (2005), 084104.
[204] Rahul Rai and Chandan K Sahu. 2020. Driven by Data or Derived Through Physics? A Review of Hybrid Physics
Guided Machine Learning Techniques With Cyber-Physical System (CPS) Focus. IEEE Access 8 (2020), 71050–71073.
[205] Maziar Raissi. 2018. Deep hidden physics models: Deep learning of nonlinear partial differential equations. The
Journal of Machine Learning Research 19, 1 (2018), 932–955.
[206] Maziar Raissi and George Em Karniadakis. 2018. Hidden physics models: Machine learning of nonlinear partial
differential equations. J. Comput. Phys. 357 (2018), 125–141.
[207] M Raissi, P Perdikaris, and GE Karniadakis. 2017. Physics informed deep learning (part II): Data-driven discovery of
nonlinear partial differential equations. arXiv:1711.10561 (2017).
[208] M Raissi, P Perdikaris, and GE Karniadakis. 2019. Physics-informed neural networks: A deep learning framework for
solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys (2019).
[209] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. 2017. Physics informed deep learning (part i): Data-driven
solutions of nonlinear partial differential equations. arXiv preprint arXiv:1711.10561 (2017).
[210] M Raissi, Z Wang, M Triantafyllou, and G Karniadakis. 2019. Deep learning of vortex-induced vibrations. J. Fluid
Mech (2019).
[211] MM Rajabi and H Ketabchi. 2017. Uncertainty-based simulation-optimization using Gaussian process emulation:
application to coastal groundwater management. J. Hydrol (2017).
[212] Stephan Rasp, Michael S Pritchard, and Pierre Gentine. 2018. Deep learning to represent subgrid processes in climate
models. Proceedings of the National Academy of Sciences 115, 39 (2018), 9684–9689.
[213] JS Read, X Jia, J Willard, AP Appling, JA Zwart, SK Oliver, A Karpatne, GJA Hansen, PC Hanson, W Watkins, et al.
2019. Process-guided deep learning predictions of lake water temperature. Water Resources Research (2019).
[214] Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, et al. 2019.
Deep learning and process understanding for data-driven Earth system science. Nature 566, 7743 (2019), 195–204.
[215] K Rudd and S Ferrari. 2015. A constrained integration (CINT) approach to solving partial differential equations using
artificial neural networks. Neurocomputing (2015).
[216] SH Rudy, SL Brunton, JL Proctor, and JN Kutz. 2017. Data-driven discovery of partial differential equations. Science
Advances (2017).
[217] Samuel H Rudy, J Nathan Kutz, and Steven L Brunton. 2019. Deep learning of dynamics and signal-noise decomposition
with time-stepping constraints. J. Comput. Phys. 396 (2019), 483–506.
[218] L Ruthotto and E Haber. 2018. Deep neural networks motivated by partial differential equations. Journal of
Mathematical Imaging and Vision (2018).
[219] Jonathan J Rutz, Christine A Shields, Juan M Lora, Ashley E Payne, Bin Guan, Paul Ullrich, Travis O’Brien, L Ruby
Leung, F Martin Ralph, Michael Wehner, et al. 2019. The atmospheric river tracking method intercomparison project
(ARTMIP): quantifying uncertainties in atmospheric river climatology. Journal of Geophysical Research: Atmospheres
(2019).
[220] Kevin Ryan, Jeff Lengyel, and Michael Shatruk. 2018. Crystal structure prediction via deep learning. Journal of the
American Chemical Society 140, 32 (2018), 10158–10168.
[221] M Sadoughi and C Hu. 2019. Physics-based convolutional neural network for fault diagnosis of rolling element
bearings. IEEE Sensors Journal (2019).
[222] Peter Sadowski, David Fooshee, Niranjan Subrahmanya, and Pierre Baldi. 2016. Synergies Between Quantum
Mechanics and Machine Learning in Reaction Prediction. Journal of Chemical Information and Modeling 56, 11 (2016),
2125–2128. https://doi.org/10.1021/acs.jcim.6b00351 arXiv:https://doi.org/10.1021/acs.jcim.6b00351 PMID: 27749058.
[223] O San and R Maulik. 2018. Machine learning closures for model order reduction of thermal fluids. Applied Mathematical
Modelling (2018).
[224] Omer San and Romit Maulik. 2018. Neural network closures for nonlinear model order reduction. Advances in
Computational Mathematics 44, 6 (2018), 1717–1750.
[225] Omer San, Romit Maulik, and Mansoor Ahmed. 2019. An artificial neural network framework for reduced order
modeling of transient flows. Communications in Nonlinear Science and Numerical Simulation 77 (2019), 271–287.
[226] Gabriel R Schleder, Antonio CM Padilha, Carlos Mera Acosta, Marcio Costa, and Adalberto Fazzio. 2019. From DFT to
machine learning: recent approaches to materials science–a review. Journal of Physics: Materials 2, 3 (2019), 032001.

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 31

[227] M Schmidt and H Lipson. 2009. Distilling free-form natural laws from experimental data. Science (2009).
[228] Kristof Schütt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-
Robert Müller. 2017. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. In
Advances in neural information processing systems. 991–1001.
[229] O Senouf, S Vedula, T Weiss, A Bronstein, O Michailovich, and M Zibulevsky. 2019. Self-supervised learning of
inverse problem solvers in medical imaging. In Domain Adaptation and Representation Transfer and Medical Image
Learning with Less Labels and Imperfect Data. Springer.
[230] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. 2018. Airsim: High-fidelity visual and physical
simulation for autonomous vehicles. In Field and service robotics. Springer, 621–635.
[231] V Shah, A Joshi, S Ghosal, B Pokuri, S Sarkar, B Ganapathysubramanian, and C Hegde. 2019. Encoding invariances in
deep generative models. arXiv:1906.01626 (2019).
[232] Ehsan Sharifi, B Saghafian, and R Steinacker. 2019. Downscaling satellite precipitation estimates with multiple
linear regression, artificial neural networks, and spline interpolation techniques. Journal of Geophysical Research:
Atmospheres 124, 2 (2019), 789–805.
[233] Ati S Sharma, Igor Mezić, and Beverley J McKeon. 2016. Correspondence between Koopman mode decomposition,
resolvent mode decomposition, and invariant solutions of the Navier-Stokes equations. Physical Review Fluids 1, 3
(2016), 032402.
[234] R Sharma, AB Farimani, J Gomes, P Eastman, and V Pande. 2018. Weakly-supervised deep learning of heat transport
via physics informed loss. arXiv:1807.11374 (2018).
[235] Justin Sirignano and Konstantinos Spiliopoulos. 2018. DGM: A deep learning algorithm for solving partial differential
equations. J. Comput. Phys. 375 (2018), 1339–1364.
[236] D Solle, B Hitzmann, C Herwig, M Pereira R, S Ulonska, L Wuerth, A Prata, and T Steckenreiter. 2017. Between the
poles of data-driven and mechanistic modeling for process operation. Chemie Ingenieur Technik (2017).
[237] Prashant K Srivastava, Dawei Han, Miguel Rico Ramirez, and Tanvir Islam. 2013. Machine learning techniques for
downscaling SMOS satellite soil moisture using MODIS land surface temperature for hydrological application. Water
resources management 27, 8 (2013), 3127–3144.
[238] P Sturmfels, S Rutherford, M Angstadt, M Peterson, C Sripada, and J Wiens. 2018. A domain guided CNN architecture
for predicting age from structural brain images. arXiv:1808.04362 (2018).
[239] Mohammad M Sultan, Hannah K Wayment-Steele, and Vijay S Pande. 2018. Transferable neural networks for
enhanced sampling of protein dynamics. Journal of chemical theory and computation 14, 4 (2018), 1887–1894.
[240] Jian Sun, Zhan Niu, Kristopher A Innanen, Junxiao Li, and Daniel O Trad. 2020. A theory-guided deep-learning
formulation and optimization of seismic waveform inversion. Geophysics 85, 2 (2020), R87–R99.
[241] Daniel Heestermans Svendsen, Luca Martino, Manuel Campos-Taberner, Francisco Javier García-Haro, and Gustau
Camps-Valls. 2017. Joint Gaussian processes for biophysical parameter retrieval. IEEE Transactions on Geoscience and
Remote Sensing 56, 3 (2017), 1718–1727.
[242] Laura Swiler, Mamikon Gulian, Ari Frankel, Cosmin Safta, and John Jakeman. 2020. A Survey of Constrained Gaussian
Process Regression: Approaches and Implementation Challenges. arXiv preprint arXiv:2006.09319 (2020).
[243] Nima Tajbakhsh, Jae Y Shin, Suryakanth R Gurudu, R Todd Hurst, Christopher B Kendall, Michael B Gotway, and
Jianming Liang. 2016. Convolutional neural networks for medical image analysis: Full training or fine tuning? IEEE
transactions on medical imaging 35, 5 (2016), 1299–1312.
[244] Naoya Takeishi, Yoshinobu Kawahara, and Takehisa Yairi. 2017. Learning Koopman invariant subspaces for dynamic
mode decomposition. In Advances in Neural Information Processing Systems. 1130–1140.
[245] Christian Tesche, Carlo N De Cecco, Stefan Baumann, Matthias Renker, Tindal W McLaurin, Taylor M Duguay,
Richard R Bayer 2nd, Daniel H Steinberg, Katharine L Grant, Christian Canstein, et al. 2018. Coronary CT angiography–
derived fractional flow reserve: machine learning algorithm versus computational fluid dynamics modeling. Radiology
288, 1 (2018), 64–72.
[246] Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. 2018. Tensor field
networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219
(2018).
[247] Michael L Thompson and Mark A Kramer. 1994. Modeling chemical processes using prior knowledge and neural
networks. AIChE Journal 40, 8 (1994), 1328–1340.
[248] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. 2017. Domain randomization
for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on
intelligent robots and systems (IROS). IEEE, 23–30.
[249] Peter Toth, Danilo Jimenez Rezende, Andrew Jaegle, Sébastien Racanière, Aleksandar Botev, and Irina Higgins. 2019.
Hamiltonian generative networks. arXiv preprint arXiv:1909.13789 (2019).

, Vol. 1, No. 1, Article . Publication date: July 2020.

32 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

[250] RK Tripathy and I Bilionis. 2018. Deep UQ: Learning deep neural network surrogate models for high dimensional
uncertainty quantification. J. Comput. Phys (2018).
[251] Silviu-Marian Udrescu and Max Tegmark. 2020. AI Feynman: A physics-inspired method for symbolic regression.
Science Advances 6, 16 (2020), eaay2631.
[252] D Ulyanov, A Vedaldi, and V Lempitsky. 2018. Deep image prior. In CVPR.
[253] J Vamaraju and MK Sen. 2019. Unsupervised physics-based neural networks for seismic migration. Interpretation
(2019).
[254] Thomas Vandal, Evan Kodra, Sangram Ganguly, Andrew Michaelis, Ramakrishna Nemani, and Auroop R Ganguly.
2017. Deepsd: Generating high resolution climate change projections through single image super-resolution. In
Proceedings of the 23rd acm sigkdd international conference on knowledge discovery and data mining. 1663–1672.
[255] PR Vlachas, W Byeon, ZY Wan, TP Sapsis, and P Koumoutsakos. 2018. Data-driven forecasting of high-dimensional
chaotic systems with long short-term memory networks. Proc. R. Soc. A (2018).
[256] Laura von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven Giesselbach, Raoul Heese, Birgit Kirsch,
Julius Pfrommer, Annika Pick, Rajkumar Ramamurthy, et al. 2020. Informed machine learning–a taxonomy and
survey of integrating knowledge into learning systems. arXiv preprint arXiv:1903.12394 (2020).
[257] ZY Wan, P Vlachas, P Koumoutsakos, and T Sapsis. 2018. Data-assisted reduced-order modeling of extreme events in
complex dynamical systems. PloS one (2018).
[258] Jian-Xun Wang, Jin-Long Wu, and Heng Xiao. 2017. Physics-informed machine learning approach for reconstructing
Reynolds stress modeling discrepancies based on DNS data. Physical Review Fluids 2, 3 (2017), 034603.
[259] Lijing Wang, Jiangzhuo Chen, and Madhav Marathe. 2020. TDEFSI: Theory-guided Deep Learning-based Epidemic
Forecasting with Synthetic Information. ACM Transactions on Spatial Algorithms and Systems (TSAS) 6, 3 (2020),
1–39.
[260] Rui Wang, Robin Walters, and Rose Yu. 2020. Incorporating Symmetry into Deep Dynamics Models for Improved
Generalization. arXiv preprint arXiv:2002.03061 (2020).
[261] Zhen Wang, Haibin Di, Muhammad Amir Shafiq, Yazeed Alaudah, and Ghassan AlRegib. 2018. Successful leveraging
of image processing and machine learning in seismic structural interpretation: A review. The Leading Edge 37, 6
(2018), 451–461.
[262] Christoph Wehmeyer and Frank Noé. 2018. Time-lagged autoencoders: Deep learning of slow collective variables for
molecular kinetics. The Journal of chemical physics 148, 24 (2018), 241703.
[263] Han Wei, Shuaishuai Zhao, Qingyuan Rong, and Hua Bao. 2018. Predicting the effective thermal conductivities of
composite materials and porous media by machine learning methods. International Journal of Heat and Mass Transfer
127 (2018), 908–916.
[264] E Weinan, Jiequn Han, and Arnulf Jentzen. 2017. Deep learning-based numerical methods for high-dimensional
parabolic partial differential equations and backward stochastic differential equations. Communications in Mathematics
and Statistics 5, 4 (2017), 349–380.
[265] Christopher KI Williams and Carl Edward Rasmussen. 2006. Gaussian processes for machine learning. Vol. 2. MIT
press Cambridge, MA.
[266] Nick Winovich, Karthik Ramani, and Guang Lin. 2019. ConvPDE-UQ: Convolutional neural networks with quantified
uncertainty for heterogeneous elliptic partial differential equations on varied domains. J. Comput. Phys. 394 (2019),
263–279.
[267] Hao Wu, Feliks Nüske, Fabian Paul, Stefan Klus, Péter Koltai, and Frank Noé. 2017. Variational Koopman models:
slow collective variables and molecular kinetics from short off-equilibrium simulations. The Journal of chemical
physics 146, 15 (2017), 154104.
[268] JL Wu, K Kashinath, A Albert, D Chirila, H Xiao, et al. 2020. Enforcing statistical constraints in generative adversarial
networks for modeling chaotic dynamical systems. J. Comput. Phys (2020).
[269] Jin-Long Wu, Karthik Kashinath, Adrian Albert, Dragos Chirila, Heng Xiao, et al. 2019. Enforcing statistical constraints
in generative adversarial networks for modeling chaotic dynamical systems. J. Comput. Phys. (2019), 109209.
[270] Jin-Long Wu, Jian-Xun Wang, Heng Xiao, and Julia Ling. 2016. Physics-informed machine learning for predictive
turbulence modeling: A priori assessment of prediction confidence. arXiv preprint arXiv:1607.04563 (2016).
[271] Jin-Long Wu, Heng Xiao, and Eric Paterson. 2018. Physics-informed machine learning approach for augmenting
turbulence models: A comprehensive framework. Physical Review Fluids 3, 7 (2018), 074602.
[272] Yuankai Wu, Dingyi Zhuang, Aurelie Labbe, and Lijun Sun. 2020. Inductive Graph Neural Networks for Spatiotemporal
Kriging. arXiv preprint arXiv:2006.07527 (2020).
[273] D Xiao, CE Heaney, L Mottet, F Fang, W Lin, IM Navon, Y Guo, OK Matar, AG Robins, and CC Pain. 2019. A reduced
order model for turbulent flows in the urban environment using machine learning. Building and Environment 148
(2019), 323–337.

, Vol. 1, No. 1, Article . Publication date: July 2020.

Integrating Physics-Based Modeling With Machine Learning: A Survey 33

[274] You Xie, Erik Franz, Mengyu Chu, and Nils Thuerey. 2018. tempogan: A temporally coherent, volumetric gan for
super-resolution fluid flow. ACM Transactions on Graphics (TOG) 37, 4 (2018), 1–15.
[275] TF Xu and AJ Valocchi. 2015. Data-driven methods to improve baseflow prediction of a regional groundwater model.
Computers & Geosciences (2015).
[276] Wenjin Yan, Shuangquan Hu, Yanhui Yang, Furong Gao, and Tao Chen. 2011. Bayesian migration of Gaussian process
regression for rapid process modeling and optimization. Chemical engineering journal 166, 3 (2011), 1095–1103.
[277] Liu Yang, Xuhui Meng, and George Em Karniadakis. 2020. B-PINNs: Bayesian physics-informed neural networks for
forward and inverse PDE problems with noisy data. arXiv preprint arXiv:2003.06097 (2020).
[278] L Yang, S Treichler, T Kurth, K Fischer, D Barajas-Solano, J Romero, V Churavy, A Tartakovsky, M Houston, M
Prabhat, et al. 2019. Highly-Ccalable, Physics-Informed GANs for Learning Solutions of Stochastic PDEs. In 2019
IEEE/ACM Deep Learning on Supercomputers. IEEE.
[279] L Yang, D Zhang, and GE Karniadakis. 2018. Physics-informed generative adversarial networks for stochastic
differential equations. arXiv:1811.02033 (2018).
[280] Tao Yang, Fubao Sun, Pierre Gentine, Wenbin Liu, Hong Wang, Jiabo Yin, Muye Du, and Changming Liu. 2019.
Evaluation and machine learning improvement of global hydrological model-based flood simulations. Environmental
Research Letters 14, 11 (2019), 114027.
[281] Y Yang and P Perdikaris. 2018. Physics-informed deep generative models. arXiv:1812.03511 (2018).
[282] Y Yang and P Perdikaris. 2019. Adversarial uncertainty quantification in physics-informed neural networks. J. Comput.
Phys (2019).
[283] Yibo Yang and Paris Perdikaris. 2019. Conditional deep surrogate models for stochastic, high-dimensional, and
multi-fidelity systems. Computational Mechanics 64, 2 (2019), 417–434.
[284] K Yao, JE Herr, DW Toth, R Mckintyre, and J Parkhill. 2018. The TensorMol-0.1 model chemistry: a neural network
augmented with long-range physics. Chemical Science (Jan 2018).
[285] Alireza Yazdani, Maziar Raissi, and George Em Karniadakis. 2019. Systems biology informed deep learning for
inferring parameters and hidden dynamics. bioRxiv (2019), 865063.
[286] Enoch Yeung, Soumya Kundu, and Nathan Hodas. 2019. Learning deep neural network representations for Koopman
operators of nonlinear dynamical systems. In 2019 American Control Conference (ACC). IEEE, 4832–4839.
[287] J Yi and VR Prybutok. 1996. A neural network model forecasting for prediction of daily maximum ozone concentration
in an industrialized urban area. Environmental pollution (1996).
[288] Yang Zeng, Jin-Long Wu, and Heng Xiao. 2019. Enforcing Deterministic Constraints on Generative Adversarial
Networks for Emulating Physical Systems. arXiv preprint arXiv:1911.06671 (2019).
[289] Leonardo Zepeda-Núñez, Yixiao Chen, Jiefu Zhang, Weile Jia, Linfeng Zhang, and Lin Lin. 2019. Deep Density:
circumventing the Kohn-Sham equations via symmetry preserving neural networks. arXiv preprint arXiv:1912.00775
(2019).
[290] Jun Zhang and Wenjun Ma. 2020. Data-driven discovery of governing equations for fluid dynamics based on molecular
simulation. Journal of Fluid Mechanics 892 (2020).
[291] Linfeng Zhang, Jiequn Han, Han Wang, Roberto Car, and Weinan E. 2018. DeePCG: Constructing coarse-grained
models via deep neural networks. The Journal of chemical physics 149, 3 (2018), 034101.
[292] Linfeng Zhang, Jiequn Han, Han Wang, Roberto Car, and E Weinan. 2018. Deep potential molecular dynamics: a
scalable model with the accuracy of quantum mechanics. Physical review letters 120, 14 (2018), 143001.
[293] Linfeng Zhang, Jiequn Han, Han Wang, Wissam Saidi, Roberto Car, and E Weinan. 2018. End-to-end symmetry
preserving inter-atomic potential energy model for finite and extended systems. In Advances in Neural Information
Processing Systems. 4436–4446.
[294] L Zhang, G Wang, and GB Giannakis. 2018. Real-time power system state estimation via deep unrolled neural
networks. In GlobalSIP. IEEE.
[295] R Zhang, Y Liu, and H Sun. 2019. Physics-guided Convolutional Neural Network (PhyCNN) for Data-driven Seismic
Response Modeling. arXiv:1909.08118 (2019).
[296] X Zhang, F Liang, R Srinivasan, and M Van Liew. 2009. Estimating uncertainty of streamflow simulation using
Bayesian neural networks. Water Resources Research (2009).
[297] Z Zhang, P Luo, CC Loy, and X Tang. 2014. Facial landmark detection by deep multi-task learning. In ECCV. Springer.
[298] Qiang Zheng, Lingzao Zeng, and George Em Karniadakis. 2019. Physics-informed semantic inpainting: Application
to geostatistical modeling. arXiv preprint arXiv:1909.09459 (2019).
[299] Yaofeng Desmond Zhong, Biswadip Dey, and Amit Chakraborty. 2019. Symplectic ODE-Net: Learning Hamiltonian
Dynamics with Control. arXiv preprint arXiv:1909.12077 (2019).
[300] Y Zhu and N Zabaras. 2018. Bayesian deep convolutional encoder–decoder networks for surrogate modeling and
uncertainty quantification. J. Comput. Phys (2018).

, Vol. 1, No. 1, Article . Publication date: July 2020.

34 Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar

[301] Y Zhu, N Zabaras, PS Koutsourelakis, and P Perdikaris. 2019. Physics-constrained deep learning for high-dimensional
surrogate modeling and uncertainty quantification without labeled data. J. Comput. Phys (2019).

, Vol. 1, No. 1, Article . Publication date: July 2020.