See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/245437600

Robust Multiscale Simulation-based Design of Multifunctional Materials

Conference Paper · January 2005

CITATIONS READS
2 297

5 authors, including:

Jitesh Panchal Hae-Jin Choi

Purdue University Chung-Ang University
299 PUBLICATIONS 3,352 CITATIONS 88 PUBLICATIONS 1,219 CITATIONS

SEE PROFILE SEE PROFILE

Janet Allen Farrokh Mistree

University of Oklahoma University of Oklahoma
471 PUBLICATIONS 10,655 CITATIONS 640 PUBLICATIONS 13,146 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Farrokh Mistree on 01 July 2014.

The user has requested enhancement of the downloaded file.

 International Conference on Advanced Materials Design & Development
14-16 December 2005, Goa, India

ICAMDD-2005

ROBUST MULTISCALE SIMULATION-BASED DESIGN OF MULTIFUNCTIONAL MATERIALS

Jitesh H. Panchal*, Hae-Jin Choi*,

# 1
Janet K. Allen*, David L. McDowell , Farrokh Mistree *
*Systems Realization Laboratory
#
Mechanical Properties Research Laboratory
The G. W. Woodruff School of Mechanical Engineering,
Georgia Institute of Technology, Atlanta, GA 30332-0405

ABSTRACT Keywords – Multiscale Modeling, Multiscale Simulation-

With the advances in understanding material behavior at Based Design, Materials Design
atomic and higher length scales, multiscale modeling is gaining
momentum in support of the field of materials design. Complex 1 MULTISCALE SYSTEMS – AN EMERGING
multiscale material models are shown to be useful in predicting CHALLENGE FOR ENGINEERING SYSTEMS
the overall behavior of materials by accounting for phenomena DESIGN
at much smaller scales. However, we believe that multiscale
The US Department of Energy sponsored three workshops
modeling is just a means to an end. For purposes of designing a
on multiscale mathematics [1] to identify research and funding
material - tailoring the microstructure to suit multiple
opportunities in multiscale modeling. During the two
performance requirements - multiscale models should be
workshops, a number of issues related to multiscale systems
developed in the context of their ultimate use – design decision-
were identified. The participants identified various application
making. In this paper, we focus the attention of the multiscale
domains that would benefit from multiscale modeling. These
modeling community on multiscale design, which introduces
domains include environmental sciences and geosciences,
different types challenges from the decision -making standpoint
climate, material science, combustion, biosciences, power grids
that are currently not addressed in multiscale materials
and information networks, development of biomimetic sensors
modeling efforts. The key difference is that multiscale
and devices, homeland security, etc. Materials are perhaps one
modeling deals with efficient integration of information from
of the most representative multiscale systems. As expressed by
various models to gain a holistic understanding of the system,
Olsen [2, 3], the objective in materials design is to first
whereas multiscale design deals with efficient utilization of
understand the links between processing-structure-property-
information to satisfy design objectives.
performance in a deductive manner and then use that
In this paper, we highlight the challenges in both
knowledge in an inductive manner to achieve desired
multiscale modeling and design and provide a brief overview of
performance by determining the appropriate properties,
the existing efforts in multiscale modeling. In order to fill the
structure, and processing (see Figure 1). The deductive
research gaps in multiscale design, we present the foundations
approach of understanding material properties and performance
for Multiscale Simulation-Based Design (MSBD), which is a
from the structure is the focus of materials modeling.
robust decision-based methodology for the integrated design of
Simulation-based design of materials refers to the inductive,
multiscale, multifunctional materials and systems. MSBD
goal-oriented activity. Olson’s conception of inductive design
refers to design decision-making using multidisciplinary and
can be extended to the design of materials in association with
multi-physics simulation models across multiple time and
products in which the materials are used, which is illustrated in
length scales that are utilized to achieve system level
performance requirements. The concepts and challenges in the Error! Reference source not found.. In the figure, the higher
paper are illustrated throughout the paper using a multiscale level system is hierarchically partitioned into assemblies, parts,
materials example from the design of multifunctional energetic materials, microstructure, etc. All components of the hierarchy
structural materials. can be inductively designed to achieve the performance at the
system level. The infusion of materials design in conventional

1
Corresponding Author, Professor, Associate Chair, The GW Woodruff School of Mechanical Engineering, Savannah Campus

1
product design provides designers with flexibility to achieve sustaining reactions. A certain level of porosity in the mixture
design objectives that were not previously conceivable. promotes shock-induced reaction initiation, as dynamic
) plasticity and void collapse engender substantial local
e
c tiv temperature rise (hot spots). In addition to reactivity, strength
du
s (in Performance is desirable for storage of these materials for various energy
an
s /me applications. Accordingly, safe handling of these materials
a l
Go Properties during transportation or storage demands certain
multifunctional requirements of combined strength and
e )
Structure tiv resistance to reaction initiation, as shown in Figure 3. We may
uc
ed therefore regard the resulting materials as so-called
c t (d
Processing ffe Multifunctional Energetic Structural Materials (MESMs).
de
e an In the design of MESMs, multiscale analytical,
us
Ca experimental and computational tools are employed to engineer
(Olsen, 1997)
the MESMs from micro- to nano-scales and evaluate their
Figure 1 - Olson's concept of 'materials by design' [2]. performance. Ab initio and molecular dynamics calculations
are conducted to estimate the equation of state for the
individual constituents and to explore both temperature- and
stress-induced initiation of reactions in MESMs. These
calculations provide information to mesoscale continuum
dynamic simulations of shock wave propagation through
System
energetic material mixtures composed of discrete particles that
s
od Assembly are characteristic of actual material microstructures. In
M eth
sis od
s addition to providing statistics regarding probability of reaction
aly Part eth
An initiation, results of these continuum hydrocode calculations are
f fec
t
ig nM
e/E D es then homogenized into equivalent nonequilibrium (extended
us t ed
Ca irreversible thermodynamics) models for pressure- and
Continuum
ri en
O
al- temperature-dependent mechanical behavior, as well as
Material Go
Selection
temperature-induced reaction initiation and propagation.
Microscale Effectively, bridges between models at multiple length and time
Molecular scales are created to analyze both energetic materials and
Quantum energetic structural materials, in close connection with
companion experiments.

Figure 2 - Multiscale simulation-based systems design.

From a multiscale materials modeling standpoint, the

blast
primary challenge is to integrate information generated by
different simulation models in a consistent manner so that the
overall material behavior can be predicted from the individual
constituent models. Recent developments in material modeling
have enabled prediction of behavior ranging from lower level reinforced temperature
quantum- and nano-scales to higher macroscopic and
continuum scales. Since it is possible to predict the behavior of MESM
systems (such as materials) at multiple scales, the natural next
step is to use these models for designing systems at multiple
scales, which is called Multiscale Simulation-Based Design
(MSBD). In contrast to multiscale modeling, the primary
challenge faced by MSBD is to effectively and efficiently impact
utilize information generated by wide range of models that
predict system behavior at different scales. For example, as Figure 3 – Schematic of MESM designed to resist
shown in Fig. 2, system level design is performed concurrently reaction initiation at elevated temperature and/or under
in conjunction with parts, sub-assemblies and the associated impact or blast loading.
material.
The objective is to achieve desired performance at the
An example of a material-product design problem is the
system level (e.g., a reinforced MESM). Overall system
design of multifunctional materials for energetic applications.
performance is a function of the reinforced energetic material
Thermite mixtures are multiphase mixtures of metal and metal-
behavior, which depends on the material properties. Material
oxide or intermetallic powders, often with a binder phase.
properties in turn depend on reinforcement strategy, micro-
They represent an effective means to store available energy.
scale interfaces between the mixture and reinforcement, as well
When elevated in temperature or subjected to a shock loading
as interfaces between constituent particles in the mixture.
environment, this energy can be released with exothermic, self-
Properties at micro-scale interfaces depend on nanoscale

2
interactions of lattices of constituents and at reactant interfaces. performance of complete system provides a more accurate
This hierarchy must be considered when designing energetic representation of the system. However, running these models at
material systems at multiple scales. This results in a greater smaller scales at a large enough domain to capture larger
coupling in the design, thereby increasing problem complexity. effects is computational prohibitive. For example, in a
Although design complexity is a challenge in multiscale design, theoretical sense, just by using the behavior at atomistic level,
the advantage of designing products and materials at multiple emergent properties of the whole system can be determined.
scales is increased design freedom (i.e., greater flexibility in The overall properties of a material can be predicted using the
configuring the system to achieve desired behavior), which interaction of individual atoms. However, using the current
enables designers to achieve better performance. For example, computational capabilities, it is not physically possible to
by designing energetic materials concurrently with the predict the behavior of the system using just the lowest scale
containment systems, designers have more options for models. Miller [4] argues that the current massively parallel
customizing the mass and size of containers by employing computers can handle only 0.01% of a typical aluminum grain.
custom-designed MESMs instead of relying only on container He also estimated that the total number of atoms simulated
design. This directly places requirements on both strength and worldwide in the past 30 years is on the order of 1012 – which is
reaction initiation threshold for these materials. This added only about 10% of the atoms in a single grain of aluminum.
flexibility, available through integrated design of materials and “Multiscale modeling is a new paradigm, where a variety of
systems, is a primary motivation for designing multiscale mathematical models at different levels of details can be
systems. considered and the right combination of models is selected
Having discussed the motivation for multiscale modeling during the process of computation according to the accuracy
and design, we proceed to discuss the challenges associated needs” [5]. This is the first challenge in modeling multiscale
with modeling and design of multiscale systems in Sections 2 systems - balancing the behavior prediction accuracy with
and 3. computational cost.
Multiscale systems often consist of multiple components.
2 MULTISCALE MODELING – CHALLENGES AND For example, in the material simulation, the components
APPROACHES include atoms, voids, interfaces, cracks, grain boundaries, etc.
All these components of the system interact with each other.
A multiscale model is defined as a model that takes
Different material models consider different number of
advantage of information from various scales present in the
components, which in turn determines their accuracy. By
system in order to gain a better understanding of the system
reducing the number of components and interactions modeled
while reducing the computational cost. In the previous section,
in the system, the number of required calculations reduces but
we discussed an example of multiscale system from the
the accuracy increases. Hence, modelers at each scale must
materials domain. Successful modeling of multiscale systems
consider the appropriate components in the system in order to
must overcome the challenges highlighted in Table 1 and
faithfully simulate component interactions.
discussed in Section 2.1. The approaches currently used for
Modeling required physical phenomena relevant to the
addressing these challenges are discussed in Section 2.2. A
system: Most current multiscale systems are also multi-physics
comparison of the multiscale systems with conventional
in nature. For example, the material model for MESMs consists
systems is carried out in Section 2.3 to understand the
of phenomena such as reaction initiation, reaction propagation,
challenges specific to multiscale systems and those shared with
impact, and shock propagation. These physical phenomena are
other complex systems.
governed by different physical laws and mathematical
Table 1 – Challenges in modeling of multiscale systems equations. These phenomena may either be dependent on or
coupled with each other. The impact of considering different
Multiscale Modeling Challenges
phenomena on the accuracy of the overall system behavior
Balancing the behavior prediction accuracy with computational cost prediction is different. Hence, in order to gain a reasonable
Appropriate components in the system in order to faithfully simulate
understanding of the system, it is important to model required
component interactions
Modeling required physical phenomena relevant to the system phenomena that are related to each other.
Modeling interactions between scales and interfacing them such that Modeling interactions between scales and interfacing them
they are physically compatible such that they are physically compatible: “The ability to
Appropriate selection of models and model parameters at each scale simulate complete systems requires faithfully modeling how the
Resolution of model mismatch to ensure compatibility between system is connected and controlled at all the levels”[1]. An
models important challenge in multiscale modeling is that phenomena
Managing large quantities of information (statistical datasets) at at different length and time scales are generally modeled with
different levels of abstraction
Quantifying and handling uncertainty, and its propagation
different sets of physical laws, mathematical equations, and
Targeted refinement of models parameters. The assumptions at different levels are also
Adaptive selection of details and resolution different. Each of these models provides different kinds of
insight into the system behavior and hence, must be integrated
in a manner such that the overall model provides consistent
2.1 Challenges in Modeling Multiscale Systems behavior of the system. Integration of models at different scales
requires consistent mathematical and physical description of
One of the primary challenges in multiscale systems is to coupling between scales. The challenge is that various scales
balance the need for accuracy and computational cost. depend on each other, which makes it important to determine
Generally, using a smaller scale model for predicting the how the different scales are linked with each other. Hence, as

3
Rudd and Broughton [6] point out, “any successful multiscale Quantifying and handling uncertainty: Any simulation
model must faithfully reproduce the intertwined nature of model has some amount of uncertainty. Uncertainty in
length scales”. simulation models is categorized as aleatory and epistemic
In summary, multiscale systems represent a special type of uncertainty. Aleatory uncertainty refers to the uncertainty due
complex systems, characterized by multiple components, to the inherent randomness in the physical processes, whereas
multiple physics, and multiple scales. Appropriate modeling of epistemic uncertainty refers to the uncertainty due to lack of
all these there aspects and their interactions (between knowledge about the system, which can be due to lack of
components, physics, and scales) is the key to multiscale information about model parameters and approximations in the
modeling. The properties of such complex systems can not be model. In order to make appropriate use of information
predicted merely through determination of individual sub- generated by simulation models, uncertainty quantification
system properties. It is important to model the interactions plays an important role. Capturing the information about range
between physical phenomena at various scales. of validity of models is also important. Uncertainty is
Appropriate selection of models and model parameters at especially important in multiscale models due to the
each scale (which models are appropriate at which scale): interactions between phenomena at different scales and
Simulation models can be developed at different fidelity levels quantification of this uncertainty in the models is difficult.
by changing the scope and the assumptions underlying the Managing uncertainty propagation: In addition to
model. For example, a material system may be modeled in one- quantifying the uncertainty in multiscale simulation models,
dimension, two-dimensions, or three dimensions. Depending on uncertainty propagates from one model to another along with
the system under consideration, different models may be the information flow between models. The uncertainty may
appropriate for predicting its behavior. It is important to select either get amplified or may remain under bounds. If the
the right set of models and assumptions. It can be argued that uncertainty gets amplified while passing information from one
this is a rather general requirement for any kind of simulation model to another, the overall system level simulation model
model development. However, the requirement is important for may not be acceptable although the uncertainty bounds of
multiscale systems because it has to be considered at multiple individual models are acceptable.
scales with information from one model feeding into another. Targeted refinement of models: The accuracy of the overall
The appropriateness of models also depends on the multiscale simulation model is dependent on the accuracy of
compatibility between assumptions made in different models. constituting models at individual scales and the manner in
Hence, a related requirement for multiscale modeling is which uncertainty is amplified due to information flow from
resolution of model mismatch to ensure compatibility between one model to another. Hence, in order to improve the accuracy
models. Appropriate selection of models has great impact on of the overall model, it is important to identify the model that
the accuracy and the time required for executing the models. has the maximum impact on the overall uncertainty and then
Bridging the gap between different types of information refine that model in a targeted fashion (i.e, the most critical link
(such as stochastic to deterministic, discrete to continuous, etc.) in the model chain and improve that link).
Enormous amount of information is generated in multiscale Adaptive selection of details and resolution: Although
models. This information from different sources is generally uncertainty is an important aspect of multiscale modeling and it
available in different forms such as graphs, images, text, should be controlled, many multiscale models can be simplified
numerical and experimental data, etc. Mathematical (and significantly reducing the model execution time without
software) bridges across levels of lengths and type such as reducing the accuracy. The objective of multiscale modeling is
stochastic to deterministic, discrete to continuous [1, 5] are to exploit such scenarios and to select appropriate levels of
required to integrate information from different scales. detail in the models.
Managing large quantities of information (statistical As a summary, the objective of multiscale modeling is to
datasets) at different levels of abstraction: At each finer scale, a take advantage of multiple scales in order to gain a holistic
more detailed theory has to be used, giving rise to more understanding of the system. The key in multiscale modeling is
detailed information about the system [5]. In addition to interactions between models at different scales. As suggested
mathematical challenges in coupling information at different by Rudd and co-authors, “no one of those models alone would
scales, the integration also needs to be carried out at a software suffice to describe the entire multiscale system, but it may be
infrastructure level. Issues such as synchronization of possible to combine the models of different scales, effectively
information generated by models at different scales, long run concentrating the computational power where it is needed the
times, load balancing, capturing information at various levels of most”[6]. Hence, during multiscale modeling, there is a tradeoff
abstraction in a consistent database, integration of distributed between the computational resources and overall model fidelity
computational models and hardware resources are pervasive in due to integration of knowledge from multiple scales. Although
multiscale modeling. efforts have been made to address some of these challenges for
Managing complex, multidisciplinary models: In addition individual application domains, a domain independent
to managing the data, management of simulation codes is also framework for addressing these challenges associated with
important. A repository of simulation models from which the multiscale problems is not currently available in the literature.
designers or analysts can extract the models appropriate for Some of the approaches currently used for multiscale modeling
their needs. The simulation model repository should be are discussed in the following Section 2.2.
developed considering the issues such as capturing of
assumptions, range of validity, and model’s context are
important while developing simulation based design framework
for multiscale systems.

4
 Weinan E and co-authors [7] at Princeton University have
2.2 Multiscale Modeling Approaches categorized multiscale modeling methods into classical and
Currently, there are two commonly adopted techniques for modern techniques. Classical techniques refine the macroscale
linking these multiscale models – a) parameterization and b) model using microscale models. In other words, classical
concurrent coupling. techniques are essentially microscale solvers applied to the
macroscale domain of interest. Examples of classical multiscale
a) Parameterization is a technique through with
information from lower level models is captured into a set of techniques include multi-grid method, domain decomposition,
parameters and their values. The parameters can be empirical or wavelet-based methods, adaptive mesh refinement, fast multi-
semi-empirical and can be used to approximate average pole method, and conjugate gradient method. However, the
behavior of physics at a lower scale. For example, the modern multiscale methods utilize the microscale models only
properties of materials are generally captured in terms of a few in the domain where it is required. For example, during the
parameters such as Elastic Modulus, Poisson’s Ratio, yield analysis of fracture, molecular level microscale model is
strength, etc. All models contain parameterization at certain important only at regions close to crack-tip. The regions far
level. No model is capable of simulating a phenomenon away from the crack tip can be modeled using macroscale
completely using first principles. The advantage of (continuum) models. This results in a more efficient multiscale
parameterization is its simplicity in accounting for the model, which is a combination of the microscale and
phenomena at lower scales; the disadvantage being its low macroscale model. Examples of modern multiscale models
accuracy. include Car-Parrinello method, Quasi-continuum method,
Heterogeneous Multiscale Method (HMM), Gap-Tooth scheme,
b) Coupling refers to the technique of using the model at
Coarse-Grained Monte Carlo Models, and Adaptive Model
one scale “on-the-fly” while performing calculations using
Refinement.
model at another scale. Many climate modeling efforts are
Although these techniques for multiscale modeling have
focused on increasing the resolution by reducing the minimum
proven successful in providing a greater understanding of
feature size modeled in the multiscale model. This requires
multiscale problems by increasing their accuracy, these
dynamic utilization of many levels of lower scale models in the
techniques are developed for and employed in very specific
overall simulation. This requirement directly translates to the
applications. This is mainly because these multiscale models
need for high performance computing tools. Hence, the ability
are based on specific insights into the coupling between
to couple models at multiple scales is mainly dependent on the
different scales. This coupling is not only problem dependent
power of computational tools available. This is evident from
but also dependent on the models used to describe the physics
the advances in climate modeling, which are greatly dependent
at these scales. Specific methods for coupling atomistic and
on the development in computational power, such as, the
continuum models are characterized by Miller [4]. Miller has
development of the world’s No. 1 super computer (ranked in
argues that the main challenge in coupling the atomistic and
2004) – the Earth Simulator. Using the Earth Simulator, the
continuum models is to model the transition region between the
resolution of climate level simulations has improved from the
two domains. The multiscale methods are different based on the
common 500 kilometers down to 25 kilometers, thereby
way in which they model this transition region. These models
generating more accurate descriptions of the underlying
include the FEAt (Finite Element and Atomistic) method, the
physics. Coupled links between multiscale models render the
QC (quasicontinuum) method, CLS (coupling of length scales)
overall model to be more accurate, but at the cost of increased
method, CGMD (Coarse-Grained Molecular Dynamics)
complexity and computational cost.
method, and CADD (Coupled Atomistic and Discrete
It is important to realize that even with the most
Dislocation) method. Rudd and Broughton [6] provide an
sophisticated super computers available today; there is an upper
overview of coupled multiscale methods for generating
bound on the complexity of problems that can be solved.
accurate description of materials spanning electronic to
Hence, there is a need for appropriate combination of
macroscopic length scales. The authors review methods
parameterization and coupling while linking multiscale models,
developed for seamless coupling between finite element
such that there is a balance between accuracy and
models, molecular dynamics models, and semi-empirical tight
computational cost. The basic question that material modelers
binding, where models at different scales are run concurrently.
need to answer is – “How much detail is required in modeling
The second approach – CGMD, which is a generalization of
the materials?” Currently, this question is answered by
finite elements that pass smoothly from higher scales to
modelers based on their experience. Hence, multiscale
molecular dynamics as the mesh size is reduced to atomic
modeling is as much an art as a science. This challenge is
spacing. All these methods developed in the multiscale
common across all multiscale systems. Hence, the main
materials research community are focused on using different
challenge in multiscale modeling is to reduce the complexity
models for different material regions and developing a hybrid
and recognize the simplicity of multiscale problems in order to
coupled model. Once the issues of interfacing between the
generate an accurate system description. In the past, there have
different regions have been solved, the overall hybrid
been a number of efforts in modeling the scales individually,
multiscale model is computationally efficient because it uses
but multiscale modelers have shown that by appropriate
the right tool for the right part of the system [6].
combination of different scales of models, it is possible to gain
Since all the methods developed above are developed for
a holistic understanding of the system. Multiscale modeling
specific applications, it is clear that a common (domain
employs models at different physical scales in combination to
independent) mathematical framework for multiscale modeling
build a comprehensive description of systems that could not be
is required to bridge the gap between heterogeneous models
modeled otherwise [6].
and information generated by them [7]. Heterogeneous

5
Multiscale Modeling (HMM) framework is an example of Coupling between components
effort in that direction. The HMM framework consists of two at the same scale
main components – the first being overall macroscopic scheme (Horizontal Coupling)
and the second is to estimate the missing macroscopic data

ale
needed for the implementation of macroscopic scheme by

es

Sc
tween scal
solving the microscopic model locally.

oupling)

er
gh
Before going into the design of multiscale systems we

Hi
would like to ask ourselves the following question - “How are

(Vertical C
multiscale systems different from conventional complex

Coupling be

Scale
systems?” This is a valid question because any conventional
complex system such as an automobile, an airplane, a satellite,
etc. spans multiple length scales – complete systems are at the

le
ca
order to few meters and the smallest components such as

rS
electronic sensors at the order of few millimeters (or even

we
Lo
microns). The key differences between emerging multiscale
systems and the conventional systems are the focus of the
following Section 2.3. Figure 4 - Horizontal and vertical couplings in
multiscale systems.
2.3 Conventional Complex Systems Vs. Multiscale Complex multiscale systems of the future should be
Systems designed by considering both horizontal and vertical coupling.
It is not an exaggeration to say that most of the engineering Although independent analysis can be carried out individually
design problems are multiscale in nature [5]. For example, the at each scale, new physical insight is developed by coupling
design of a car involves design of overall systems such as these scales. Specific methods are developed for solving
engine, transmission, cooling, body, etc. and their integration. domain specific multiscale design problems but a domain
Each of these systems consists of sub-systems and components independent structured methodology for multiscale systems
that interact with each other. These complex hierarchical design is not available. Current methods in simulation based
systems are defined by Koch through a hierarchical structure of design do not encompass the full set of performance criteria to
system, subsystem, and component level information, for which produce better designs considering variables from all scales –
compatible solutions are sought concurrently [8]. The material microstructure through overall system [9]. In order to
difference however, is that in multiscale systems, the coupling address this challenge, there is a need for domain independent
is between physical phenomena at different scales for the same methodology for designing multiscale systems.
component, whereas, in hierarchical systems considered so far All complex systems are characterized by three types of
(such as the one considered by Koch in [8]), the coupling is couplings – between components of the system, between
primarily between subsystems (physical components). The physical phenomena, and between different scales (see Figure
complexity in such conventional multiscale systems is due to 5). The strengths of each of these couplings are different in
coupling between components at the same level (scale). This is different systems. Some of the couplings are weak and may be
referred to as horizontal coupling (see Figure 4). However, the ignored during modeling and design, while others are strong
scales are not tightly coupled with each other, thereby allowing and should be considered. Science abounds with examples of
for independent design of components that can be used in the multiscale systems in which the scales are only weakly coupled.
system level design. The complexity in multiscale design arises Were this not so, we would have made little progress in the
when a) these scales are tightly linked with each other, i.e., theoretical sciences [6]. In the conventional systems, the
vertical coupling (see Figure 4); b) each scale is described by a strength of coupling between system components is high and
different set of physical principles. Hence, in conventional the strength of couplings between physical phenomena and
hierarchical systems, the coupling is simpler to consider as across different scales is weak. In the multiscale systems such
compared to multiscale systems. In hierarchical systems, the as the environment, all the three types of couplings are strong
challenges include integration of multiple disciplines, and and must be explicitly modeled and accounted for decision
require resolution of multiple conflicting objectives. The main making.
reason for considering coupling between physics is to gain an Interactions
understanding of the system efficiently and making right Between Physics
Physical
decisions in an efficient manner. The main reason for Physical
Phenomena
Physical
Phenomena
considering coupling between subsystems in a concurrent Phenomena
fashion is to make right decisions. The coupling is handled in
hierarchical systems by introducing intermediate responses,
linking variables, and compatibility constraints. Interactions Between
Components Interactions
Between Scales

System Multiple
System Multiple
Components
System Scales
Multiple
Components Scales
Components Scales

Figure 5 – Complex systems with multiple components,

scales and physical phenomena interacting with each other.

6
 Management of uncertainty: Uncertainty in design of
multiscale systems arises from three sources – a) inherent
3 MULTISCALE DESIGN – A NEED FOR EFFICIENT randomness in the system (natural uncertainty, variability), b)
DECISION MAKING lack of information about the parameters in a model (model
parameter uncertainty), and c) the error introduced in the
Multiscale design refers to the engineering field which
models due to simplification of simulation models (model
involves accounting for all aspects of systems from lower scale
structure uncertainty) and d) propagated uncertainty in analysis
materials to larger scale systems throughout the product
and design chains. Effective management of uncertainty
lifecycle in order to make risk informed design decisions at all
involves making decisions robust to uncertainty in the
scales. The primary design challenge due to multiscale nature
simulation models and mitigating uncertainty through model
of the problem is interactions between scales, which
refinement. Proper accounting of uncertainty is especially
necessitates designers to appropriately account for coupling
important in multiscale design because of the propagation of
between scales that affect the ultimate behavior of the complete
uncertainty across different models and scales. Design methods
system. As mentioned in Section 1, the objective in designing is
for robust decision making under model uncertainty and
to utilize information generated by multiscale models in a goal-
propagated uncertainty are required.
oriented manner to satisfy the requirements. “Current methods
Increased number of design variables and coupling:
in simulation based design do not encompass the full set of
Couplings between multiscale models induce complexity in the
performance criteria to produce better designs considering
associated design processes, which is further increased if the
variables from all scales – material microstructure through
design is multi-functional. Multiscale, multi-functional design
overall system” [9]. Design of multiscale systems is
processes involve different domain experts with distributed
characterized by the challenges of multiscale modeling and
simulation models. All these factors further complicate the
additional challenges associated with design exploration at
design processes. In order to reduce the complexity of design
different scales. The manner in which design activities are
processes, it is important that only the couplings that are most
carried out reflects a problem solving emphasis rather than
important for decision making be considered in the design
information gathering emphasis. The overall objective in
process. Hence, designers are faced with decisions related to
designing is to enable the design and manufacturing of
the product but also the decisions related to design processes.
increasingly complex products at lower cost and in less time. In
These decisions constitute the meta-design and are called meta-
order to perform multiscale design, the ability to develop
level decisions. The meta-level decisions are concerned with
multiscale models is a pre-requisite. Multiscale design poses
the tradeoff between the complexity of design processes, and
additional challenges due to the need for design synthesis and
the satisfaction of design objectives.
associated decision making. In contrast to multiscale modeling,
Design process configuration: It is clear that the processes
where the tradeoff is between simulation time and accuracy, the
used to design the multiscale systems have a significant impact
tradeoff in multiscale design is between the satisfaction of
on the computational costs and design efficiency. Appropriately
design objectives and the time for design. The challenges
designed design processes can lead to better design solutions
associated with multiscale design are highlighted in Table 2 and
faster. Hence in multiscale design, there is a need for careful
discussed in this section. The strategies for addressing these
consideration of the design process. The design of design
challenges are discussed in 4.
processes involves decisions at three levels – a) architecture
Table 2 - Challenges and proposed strategies for level, where the precedence relationships between decisions
simulation-based multiscale design and tasks are determined, b) individual decision level, where
the simulation models used to generate information for making
Multiscale Design
design decisions are selected, and c) at the individual model
Challenges Strategies level, where the parameters, approximations, etc. associated
1 Management and Robust design Type I-IV with the simulation model are decided upon.
handling of uncertainty (see Section 4.1)
Evolving simulation models and requirements: Another set
Integrated design of
2. Design process of challenges inherent in multiscale design of systems are – a)
products and design processes (see
configuration the fidelity of simulation models increases with time because of
Section 4.2, Figure 7)
Methods and metrics for the evolving system knowledge, and b) the design requirements
consciously and systematically evolve with time. The objective, hence, is to utilize the models
3. Increased design deciding which couplings are available at various fidelity levels and develop a preliminary
coupling important and should be considered design, which can be refined when additional knowledge about
for decision making (see Section 4.2, the models or the requirements. “The hierarchical nature of
Figure 8) multiscale models offers the promise of obtaining
Using value of information- computational improvement, especially in early stages of the
based metrics for determining
4. Evolving optimization process, by considering only as much model
appropriate levels of details of
simulation models resolution as necessary to obtain sufficient progress at a given
simulation models (see Section 4.3,
Figure 9) iteration” [9]. It is important to focus the refinement effort on
Development of new meta- aspects that have the most impact on the final satisfaction of
5. Efficient design modeling techniques for non- design requirements. Hence, design methods for multiscale
exploration deterministic systems (see Section systems should be open to refinement of simulation models.
4.4) Efficient design exploration: Multiscale problems are
generally characterized by an increase in number of parameters

7
that can be selected appropriately to achieve design goals. This data. Qian and the co-authors [11] propose a modified
increases the efforts for design exploration using conventional calibrated model by modeling the scale term as an unknown
techniques and calls for the development of faster and more linear approximate regression function. These two methods
efficient design exploration techniques. These include design of are particularly useful for decreasing the computational expense
computer experiments, approximation techniques, etc. Further, by reducing the number of samples in fine mesh analyses.
multiscale design problems are characterized by long Tobias [12] propose detailed guidelines for choosing the best
simulation runtime and large degrees of freedom. Using such model among available mathematical or computer models by
models in the design exploration loops is computationally measuring levels of detail, complexity, and corresponding
prohibitive. Hence, efficient design of experiments and meta- model performance. Sargent [13] develops a guideline for
modeling techniques are required to create simplified model validation, which includes data validity, conceptual
mathematical relationships between the design variables and model validity, computerized model verification, and
responses that can be used for design space exploration. operational validity. Jin and co-authors [14] test various
Distributed decision makers: Geographical distribution of metamodeling techniques for different optimization
designers adds to the complexity of design processes by formulations under uncertainty and compare the accuracy of the
increasing the bandwidth of information transfer and associated approximation results. Simpson and co-authors [15] also survey
design time. The distribution of design expertise and functional sampling and metamodeling techniques and recommend a
knowledge dictates the way in which design problem must be guideline for the appropriate use of statistical approximation
partitioned. In such cases, the design processes are defined not techniques in a given situation.
only based on the physics based coupling between parameters
but also on designers’ expertise and how effectively they can Robust Design
exchange information. This necessitates development of design The second approach for managing uncertainties is
methods that account for such organizational and geographical designing a system to be insensitive to uncertainty without
considerations. eliminating or reducing its sources in the system; this is called
All these design related challenges can be summed up into robust design. Robust design is useful approach when reducing
the following requirement – “efficient and effective use of the uncertainty is virtually impossible or very expensive. Two
information and computational resources in design of types of robust design have been investigated.
multiscale systems”. In order to address this requirement, there
Type I Robust Design: Identify control factor (design
is a need for a systematic, domain independent framework for variable) values that satisfy a set of performance
designing multiscale systems. We believe that one of the main requirement targets despite variation in noise factors
aspects of such a framework is the consideration of decisions (uncontrollable variable).
related to both products and design processes. Hence, in this
Type II Robust Design: Identify control factor (design
paper, we focus our attention to the design of design processes variable) values that satisfy a set of performance
(meta-design) for increasing the efficiency and effectiveness requirement targets despite variation in control and noise
of design. We present a set of constructs for the design of factors.
multiscale systems in the remaining part of the paper. Type I robust design is most commonly used in practice.
Taguchi method for Type I robust design is widely used. Type
4 SYSTEMS-BASED DESIGN CONSTRUCTS FOR II robust design has been documented in the literature (e.g.,
ADDRESSING MULTISCALE CHALLENGES [16]) and its use is not as prevalent in practice as Type I.
Clearly, Types I and II robust design are needed in the design
In this section, we present our strategy for addressing the of multidisciplinary engineering systems. We have shown
challenges presented in Section 3. The strategy is discussed in that the two are not enough in engineering design problems
three subsections – management and handling of uncertainty in involving multiple disciplines and length scales [17]. To
the design of multiscale systems in Section 4.1; integrated address this shortcoming we hypothesize that the capability to
design of systems and design processes in Section 4.2; address the twin issues of variability embodied in the model
systematic consideration of design process simplification and and uncertainty propagation in an analysis and design chain is
model refinement in Section 4.3; and efficient design needed. Hence, we proposed two additional types of robust
exploration in Section 4.4. The components of the strategy and design [18-20].
their relationship with the challenges are highlighted in Table 2.
Type III Robust Design: Identify adjustable ranges for
control factors, that satisfy a set of performance
4.1 Management of Uncertainty requirement targets and/or performance requirement
For managing the sources of uncertainty discussed in ranges and are insensitive to the variability within the
Section 3, two primary approaches are available. One approach model.
is reducing the uncertainty itself, and the other is designing a For example, in materials design, it is common practice to
system to be insensitive to uncertainty without reducing or analyze a representative statistical volume element which
eliminating it. specifies a microstructure distribution. Repeated analyses of
various statistical volume elements with identical inputs yield
Reducing Uncertainty variability in performance. As shown in Figure 6, due to the
Reducing uncertainty is feasible when a designer has larger non-deterministic behavior of a system, the deviation or
amounts of data or complete (or better) knowledge of a system. objective function cannot be represented as single deterministic
Kennedy and O’Hagan [10] employ a Gaussian Process model function (the solid graph). Instead, it may be represented as
(known as kriging in spatial statistics) for fitting simple model upper and lower limit functions (such as, some percent

8
confidence/prediction interval of a response surface model). In environment. Patterns represent regularities that recur in a
this case, the optimum solution (the peak point of the solid particular design domain, and have been successfully used in
graph) and Type I and II robust solution (the flat region of the architecture and design of software programs. According to
solid graph) are not the best solution since the deviations at Alexander and co-authors [22], a pattern describes a problem
those solutions are larger than that at Type I, II, and III robust that occurs over and over again in our environment, and then
solution with consideration of those limit functions. Robust describes the core of solution to a problem, in such a way that
Concept Exploration Method with Error Margin Indices the solution can be used a million times. Patterns can be
(RCEM-EMI)[19] are proposed as a method for finding Type I, defined in many different ways – in software design, Gamma
II and III robust solution and validated based on the example of and co-authors [23] describe patterns in terms of the behavior
the MESMs Design introduced in Section 1. of objects, structure of interactions between objects, and the
Y manner in which they are created. Alexander describes patterns
Deviation
at Optimal
for architectural design in terms of neighborhood boundary,
Solution Upper Limit main gateways, arcades, etc.
It is important to note that there is no unique way of
Deviation
at Type I, II describing the patterns. Patterns are generally identified by
Robust Solution Deviation or
Objective
recognizing certain characteristics of the system that are
Function important in a given context. Since the context in this paper is
simulation-based multiscale design, we have identified patterns
Lower Limit based on interactions between simulation models at multiple
scales and between decisions. Interaction between decisions
Deviation
and models is taken as a basis for defining patterns in design
at Type I, II, III processes because a) interactions primarily define the flow of
Robust Solution
information between tasks, and b) the types of interaction
dictate the kind of design process to be used for design. The
Optimal Type I, II Type I, II, III X
Design patterns based on interactions between decisions and models
Variable
Solution Robust Robust are also useful in simplifying the design. The simplification of
Solution Solution
design processes using the interaction patterns is discussed in
Figure 6 Types I, II, and III Robust Design. Section 4.3.
The interaction patterns used as patterns in the design

Type IV Robust Design: Identify adjustable ranges of method are shown in Figure 7 and organized in a matrix form.
control factor (design variable) values under the The three columns of the matrix represent three different types
propagation and potential amplification of uncertainty in of interactions – i) independent, ii) dependent, and iii) coupled.
a process chain. In the independent scenario, both the microscale and
Type IV robust design is focused on uncertainty associated macroscale simulation models can be executed in a parallel
with multi-time and length scaled simulation and analysis fashion. The dependent scenario represents one way
model chain. Inductive Design Exploration Method (IDEM) (sequential) flow of information, where the information
[18] is proposed and validated based on a multiscale, generated by microscale model is fed into the macroscale
multifunctional materials and products design example. model. In the coupled scenario, both these models need to be
executed together with a two-way flow of information between
them. Such a classification is also common the design literature
4.2 Integrated Design of Multiscale Materials and [24], where the flow of information between tasks is defined as
Design Processes independent tasks, dependent tasks, and interdependent tasks
In order to perform integrated design of products and respectively.
design processes, our approach consists of viewing processes as The classification in three rows of the matrix is based on
systems that can be partitioned into sub-systems with well- design variables and responses associated with different models
defined interfaces. From a hierarchical systems standpoint, and multi-functionality. In the first row, the macroscale model
design processes can be progressively broken down into sub- has design variables and response variables associated to it,
processes that can be further represented in terms of basic whereas in the second row, both the microscale model and
design process building blocks, namely the information macroscale models are associated with design and response
transformations. Specifically, we identify standardized design variables. The third row represents a multifunctional design
process patterns with clearly defined inputs and outputs that scenario where at each level, there are different models that
facilitate hierarchical modeling of design processes. These predict the system behavior for different functional
process patterns are captured as reusable templates used to characteristics (such as thermal, impact, vibration, etc.).
model any multiscale design processes. The design processes,
modeled in such a manner, provide the ability to easily archive
and reuse design process knowledge. Any design process can
be described in terms of these building blocks. The building
blocks used for modeling design processes are identified by
observing regularities in design processes. According to Nikos
A. Salingaros [21], the ability to observe patterns gives us the
human advantage of both adapting to, and changing our

9
Information Passing Information Passing
Validation
from Micro-scale to Low Design Freedom
from Micro-scale to Validation
Macro-scale
Macro-scale
P1 P2 P3
Increasing
P1 P2 P3
Y1
Complexity of
X1 Y1
X1
Design Freedom at X1 Y1
Design
Y1 both Micro and Higher Design Freedom
Design Freedom at X1 Y1 X1 Y1 X1
Macro-scale X2 Y2
X 2 Y 2
Y2
Processes
X2
both Micro and
Macro-scale X2 Y2 X2 Y2 Y2 P4 P5 P6
X2

X1 Y1 Y1
P4 P5 P6 X1 Y1 X1

Multi-functional X2 Y2
Y2 Y2
X2
Y1 X2
X1 Y1 X1 Y1 X1
P7 P8 P9
Multi-functional
X2 Y2 Y2 Y2 Dependent Coupled
X2
X2
Independent
(one-way flow of (Two-way flow of
Macro-scale model information) information)
P7 P8 P9
Micro-scale model
Parameter
Dependent Coupled
Design Variable/Response
Independent (one-way flow of information) (Two-way flow of information)
Increasing Complexity of Design Processes
Figure 7 – Interaction patterns in multiscale design.
Figure 8 – Complexity in the simulation model
interaction patterns.
4.3 Design Process Simplification and Model
Related work on measuring the complexity of processes
Refinement
[25, 26] is focused mainly on the quantifying the interactions
In order to increase the efficiency of design processes, it is between different systems. Some of the metrics capture only
important that analysis model development and design the number of interacting tasks, whereas other metrics capture
exploration be carried out in a parallel fashion, instead of the strength of these interactions. In other words, existing
developing all simulation models to completion and then measures of quantify the coupling between system’s inherent
executing starting the design process in a sequential fashion. behavior only. However, the simplification of design processes
The objective here is to maximize design process efficiency by is dependent on three factors – a) the coupling inherent in the
narrowing down the design space in the preliminary design system behavior, b) the designers’ preferences, and c) the stage
phase by using approximate models and then performing finer in the design process. Coupling in system’s behavior is
design exploration using exact models. This is particularly captured in the simulation models and the relationship between
important when the simulation models evolve with time. design variables and responses. Designer’s preferences may
Multiscale, multifunctional design processes are generally either amplify or diminish the effect of coupling in the system.
complex because of the inherent coupling between various Similarly, the stage in the design process may dictate whether
scales. An independent type of design process takes less time to some coupling between design variables and responses is
execute when compared to a decoupled (dependent) design important. For example, a simple design process may be good
process, which in turn takes less time than its coupled enough in the preliminary design phases where the objective is
counterpart. By de-coupling a coupled system, designers can to reduce the number of options to a few promising options.
reduce the complexity of design processes but increase the However, in the detailed design phase, it may not be
uncertainty in the design. Hence, the designers are faced with appropriate to simplify the design process. This also indicates
the meta-level decision involving tradeoff between the appropriateness of simplification is dependent on the design
simplification of design process and effectiveness of the final timeline. Designers’ preferences and stage in the design process
design. Appropriate simplification of the design process by are not considered in the design literature to determine whether
reducing coupling is very important for design exploration. In a coupling between design variables and response is important
the context of interaction patters discussed in Section 4.2, the or not.
complexity increases from left-to-right and from top-to-bottom In this paper, we overcome this limitation by developing a
in the matrix of interaction patterns, as shown in Figure 8. metric that considers both system coupling and designer’s
Interaction Pattern P9 results in the most complex design preferences for determining whether a design process
process whereas the processes associated with Pattern P1 is the simplification is appropriate. A decision making perspective is
simplest. While going from the left-to-right column, the adopted for making the meta-level decision under
complexity increases because of increased coupling. The consideration. The guiding principle used for determining
complexity in second row is higher than the first row because whether process simplification is appropriate is the answer to
design exploration needs to be carried out at both scales. The the following question – “What is the impact of process
increase in complexity from second to third row is because of simplification on the design decisions?” If the impact on the
the additional coupling between functions at a given scale in a decision is small and process simplification reduces the design
multifunctional scenario. Hence, the objective during process exploration cost drastically, then the designer should go ahead
simplification is to systematically go from Pattern P9 to Pattern and simplify the design process, otherwise not. In order to
P1 as shown by the circular arrows. This objective gives rise to quantify the impact on decision, metrics based on value of
the need for defining a metric that guides designers to information are currently being developed. Value of
determine the right level of simplification of interaction information refers to the benefit of additional information due
patterns. to preserved couplings per unit cost of computation and
extended design time. Expected value of information, as
defined by Howard [27], and later applied to catalog selection

10
problems in engineering design by Bradley and Agogino [28], these factors have an effect on the accuracy of simulation
is given by the difference between the expected value of the models. It is obvious that the accuracy of simulation models is
option selected with the benefit of information less than important for better design decisions. However, it is not the
without. Comparing two patterns (see Figure 7) such as P2 only factor to consider in decision making. The impact of
(sequential interaction) and P3 (coupled interaction), be inaccuracy in simulation models is not directly proportional to
including the coupling in P3, we are adding information about the impact on decisions. This is because of the impact of
the system that is not accounted for in P2. If the expected value designers’ preferences that act on the outputs of these
of this added information is greater than certain threshold simulation models. In other words, although there is inaccuracy
value, pattern P3 should be used instead of P2. in the prediction of a response, the designer’s overall
preference may not be sensitive to that response. The
inaccuracy in simulation models may either get amplified or
Ideal Maximum
Payoff
Value = (Payoff achieved using actual behavior)
– (Payoff achieved using predicted diminished by the designers’ preferences. If the inaccuracy is
behavior)
amplified, then a small amount of inaccuracy has major impact
Value
on the overall result of the decision, and hence, the simulation
Achieved Payoff
Using Predicted model must be refined. In the case where inaccuracy is
Behavior
diminished, the inaccuracy in simulation models does not have
strong impact on the overall result. Hence, the simulation
model need not be refined.
Overall Payoff (Utility)

Predicted System Behavior

Actual System Behavior

4.4 Efficient Design Exploration
If a system, mostly computer simulation, is deterministic,
then material designers need to perform single simulation at
X1
‘Ideal’ Decision if real
X2
Decision Using
Design Variable each evaluation point during a design exploration process (such
behavior is known Model Predictions
as, optimization process). However, if a system is represented
Figure 9 - Conceptual description of value of as a non-deterministic model, which is the case of most
information in simplified models. material simulation models with explicit microstructure
representation, a large number of samples are required at each
Consider a scenario shown in Figure 9, where the evaluation point in the design exploration process. The most
horizontal axis is the value of design variable and the vertical accurate way of designing the non-deterministic system is to
axis is the corresponding payoff that is achieved by selecting perform a statistical uncertainty analysis (e.g., Monte Carlo
the design variable. The design variable can be some physical simulation, Latin HyperCube sampling, etc.) and to obtain the
dimensions that the designer has control over, whereas the statistical variability of the non-deterministic responses due to
payoff represents profit, which depends on system behavior input variation. However, this approach requires an extremely
such as performance, strength, and cost. The designer’s large number of experiments (more than 1,000~10,000 in
objective is to maximize the payoff by appropriate selection of Monte Carlo simulation) for the variability analysis at a single
the design variable value. The solid line represents the actual evaluation point during the design exploration process.
system behavior and the dashed line represents the system Therefore, based on the direct simulation approach, it is
behavior predicted by the simulation model. The difference in virtually impossible to consider the non-deterministic
actual and predicted behavior is due to the inaccuracy in the variability in designing the system. In this need, we propose a
model. In Figure 9, it is assumed that the decision is metamodeling approach called integrated mean model and
characterized by no uncertainty, but only imprecision due to the prediction interval estimation [19]. Based on this approach, we
inaccuracy in the simulation model. build a metamodel that represents the non-deterministic system
If a designer makes a decision using only the simulation behavior. A designer incorporates the metamodel, instead of the
model, the decision point is X2, because it maximizes the expensive computer simulation or experiment, in a design
payoff based on the predicted behavior. However a designer exploration process that requires a large number of function
would have selected decision point - X1 if the actual (real) evaluations
behavior of the system were known (by using a perfect model). In a distributed environment, it is a tedious work to
Hence, the value of using the perfect model over simpler model integrate multiple (multiscale) models running in
is the difference in payoff actually achieved by using the exact heterogeneous platforms distributed geographically. An
model. It is important to note that the value of information is approach may be taken for solving this problem is to build
evaluated using the difference in payoff using the actual system metamodels of the distributed models and a moderator collects
behavior. and synthesizes them for the decision-making. However, this
The same value of information based metrics can also be approach is also limited to handle discrete and continuous
used for determining the appropriate levels of simplification of design spaces, which is the cases of molecular and quantum
simulation models. Although the simulation models can be level simulations. The Inductive Discrete Constraints
refined in a variety of ways such as consideration of additional Evaluation in the IDEM [18] is passing discrete data sets
physical phenomena, modeling some phenomena with greater among the distributed designers instead of computationally
accuracy, inclusion of interactions between different interfacing multiscale models (or metamodels) side-by-side.
phenomena, refinement of mesh, making convergence criteria This approach needs only a distributed database but does not
more stringent, use of better microscale model, and so on. All required integration frameworks or a moderator. For example,

11
continuum-level designers find the feasible spaces based on the 7. Weinan E, B. Engquist, and Z. Huang, Heterogeneous
requirements passed from the part designer. The feasible spaces Multiscale Method: A General Methodology for
are represented as discrete feasible points within design spaces. Multiscale Modeling. Physical Review B, 2003.
The feasible points are stored in shared database or a data file. 67(092101): p. 1-4.
The micro-level designers retrieve the feasible points stored by 8. Koch, P.N., Hierarchical Modeling and Robust
continuum-level designers and use them as performance or Synthesis for the Preliminary Design of Large Scale
design space feasible points in their design task. This process Complex Systems, in Mechanical Engineering. 1997,
may be extended until the lowest level of design tasks (e.g., Georgia Institute of Technology: Atlanta, GA.
quantum-level design tasks). This approach saves large amount 9. Workshop Report, Simulation Based Engineering
of time and effort that is necessary to integrate and synthesize Science. 2004, National Science Foundation:
the distributed multiscale multidisciplinary models. This Arlington, VA. p. 1-18.
approach is also applicable to the models that have discrete and 10. Kennedy, M.C. and A. O'Hagan, Predicting the output
continuous input spaces. from a complex computer code when fast
approximations are available. Biometrika, 2000.
5 CLOSURE 87(1): p. 1-13.
In this paper, we provide an overview of the challenges in 11. Qian, Z., C.C. Seepersad, V.R. Joseph, C.F.J. Wu, and
modeling and design of multiscale systems. Existing efforts for J.K. Allen. Building Surrogate Models based on
modeling multiscale systems are discussed. An overview of our Detailed and Approximate Simulations. in ASME
strategy for addressing the design related challenges is DETC04. 2004. Salt Lake City, UT: ASME. Paper No.
provided. We are approaching the multiscale materials design DETC2004/DAC57486.
problem from a robust simulation-based approach. The 12. Brooks, R.J. and A.M. Tobias, Choosing the Best
elements of this strategy include a) making decisions that are Model: Level of Detail, Complexity, and Model
robust to noise factors, uncertainty in design variables, Performance. Mathl. Comput. Modeling, 1996. 24(4):
uncertainty inherent to simulation models, and uncertainty p. 1-14.
propagated throughout the design process, b) use of value of 13. Sargent, R.G. Verification and Validation of
information based metrics for making meta-level decisions such Simulation Models. in 2003 Winter Simulation
as determining the right level of simulation model refinement, Conferences. 2003. New Orleans, LA.
appropriateness of consideration of couplings in decisions and 14. Jin, R., X. Du, and W. Chen, The Use of
simulation models, c) use of interaction based patterns for Metamodeling Techniques for Optimization under
modeling design processes, and d) use of meta-modeling Uncertainty. Journal of Structural & Multidisciplinary
techniques that are based on non-deterministic simulation Optimization, 2001.
models. 15. Simpson, T.W., J.D. Peplinski, P.N. Koch, and J.K.
Allen, Metamodels for Computer-based Engineering
Design: Survey and Recommendations. Engineering
6 ACKNOWLEDGEMENTS
with Computers, 2001. 17: p. 129-150.
We gratefully acknowledge the support from Air Force 16. Myers, R.H. and D.C. Montgomery, Response Surface
Office of Scientific Research (AFOSR) grants F49620-03-1- Methdology: Product and Process Optimization Using
0348 and Multi-University Research Initiative Grant Designed Experiments. 1995, New York: Wiley.
(1606U81), without which, the research presented in this paper 17. Choi, H.-J., R. Austin, J. Shepherd, J.K. Allen, D.L.
would not be possible. McDowell, F. Mistree, and D.J. Benson. An Approach
for Robust Micro-Scale Materials Design under
7 REFERENCES Unparameterizable Variability. in 10th AIAA/ISSMO
1. Dolbow, J., M.A. Khaleel, and J. Mitchell, Multiscale Multidisciplinary Analysis and Optimization
Mathematics Initiative: A Roadmap. 2004, U.S. Conference. 2004. Albany, NY: AIAA. Paper No.
Department of Energy. AIAA-2004-4331.
2. Olsen, G.B., Computational Design of Hierarchically 18. Choi, H.-J., J.K. Allen, D. Rosen, D.L. McDowell, and
Structured Materials. Science, 1997. 277(5330): p. F. Mistree. An Inductive Design Exploration Method
1237-1242. for the Integrated Design of Multi-scale Materials and
3. Olsen, G.B., Designing a New Material World. Products. in ASME DETC05. 2005. Long Beach, CA:
Science, 2000. 288(5468): p. 993-998. ASME. Paper No. DETC2005-85335.
4. Miller, R.E., Direct Coupling of Atomistic and 19. Choi, H.-J., R. Austin, J.K. Allen, D.L. McDowell, F.
Continuum Mechanics in Computational Materials Mistree, and D.J. Benson, An Approach for Robust
Science. International Journal of Computational Design of Reactive Powder Metal Mixtures Based on
Materials Science, 2003. 1(1): p. 57-72. Non-deterministic Micro-Scale Shock Simulation.
5. Weinan E and B. Engquist, Multiscale Modeling and Journal of Computer-Aided Materials Design, 2005.
Computation. Notices of the AMS, 2003. 50(9): p. 12(1): p. 57-85.
1062-1070. 20. Seepersad, C.C., M.G. Fernandez, J.H. Panchal, H.-J.
6. Rudd, R.E. and J.Q. Broughton, Concurrent Coupling Choi, J.K. Allen, D.L. McDowell, and F. Mistree.
of Length Scales in Solid State Systems. Phys. Stat. Foundations for a Systems-Based Approach for
Sol. (b), 2000. 217: p. 251-291. Materials Design. in 10th AIAA/ISSMO
Multidisciplinary Analysis and Optimization

12
 Conference. 2004. Albany, NY: AIAA MAO. Paper
No. AIAA-2004-4300.
21. Salingaros, N.A., The Structure of Pattern Languages.
Architectural Research Quarterly, 2000. 4: p. 149-161.
22. Alexander, C., S. Ishikawa, M. Silverstein, M.
Jacobson, I. Fiksdahl-King, and S. Angel, A Pattern
Language. 1977, New York: Oxford University Press.
23. Gamma, E., R. Helm, R. Johnson, and J. Vlissides,
Design Patterns - Elements of Reusable Object
Oriented Software. 2000: Addison-Wesley
Professional.
24. Eppinger, S.D., Model-based Approaches to
Managing Concurent Engineering. Journal of
Engineering Design, 1991. 2(4): p. 283-290.
25. Braha, D. and O. Maimon, A Mathematical Theory of
Design: Foundations, Algorithms, and Applications.
1998, Boston: Kluwer.
26. Buede, D.M., The Engineering Design of Systems:
Models and Methods. 2000, New York, N.Y.: John
Wiley & Sons, Inc.
27. Howard, R., Information Value Theory. IEEE
Transactions on Systems Science and Cybernetics,
1966. SSC-2(1): p. 779-783.
28. Bradley, S.R. and A.M. Agogino, An Intelligent Real
Time Design Methodology for Component Selection:
An Approach to Managing Uncertainty. Journal of
Mechanical Design, 1994. 116: p. 980-988.

13

View publication stats