PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

Differentiable master equation solver for quantum device characterization

D. L. Craig ,1 N. Ares ,2 and E. M. Gauger 3,*

1
Department of Materials, University of Oxford, Parks Road, Oxford OX1 3PH, United Kingdom
2
Department of Engineering Science, University of Oxford, Parks Road, Oxford OX1 3PJ, United Kingdom
3
SUPA, Institute of Photonics and Quantum Sciences, Heriot-Watt University, Edinburgh EH14 4AS, United Kingdom

(Received 16 March 2024; accepted 3 October 2024; published 20 November 2024)

Differentiable models of physical systems provide a powerful platform for gradient-based algorithms, with
particular impact on parameter estimation and optimal control. Quantum systems present a particular challenge
for such characterization and control, owing to their inherently stochastic nature and sensitivity to environmental
parameters. To address this challenge, we present a versatile differentiable quantum master equation solver
facilitating direct computation of steady-state solutions, and incorporate this solver into a framework for device
characterization capable of dealing with additional nondifferentiable parameters. Our approach utilizes gradient-
based optimization and Bayesian inference to provide estimates and uncertainties in quantum device parameters.
To showcase our approach, we consider steady-state charge transport through electrostatically defined quantum
dots. Using simulated data, we demonstrate efficient estimation of parameters for a single quantum dot, and
model selection as well as the capability of our solver to compute time evolution for a double quantum dot
system. Our differentiable solver stands to widen the impact of physics-aware machine learning algorithms on
quantum devices for characterization and control.

DOI: 10.1103/PhysRevResearch.6.043175

I. INTRODUCTION efficient tomographic methods have been developed using

deep learning models as a variational ansatz [9], and deep
The challenge of characterizing quantum devices has in-
learning methods have also been applied to characterization
spired a myriad of approaches, with the goal of furthering the
of open quantum system dynamics and estimating physical
understanding and application of quantum technologies [1].
parameters describing the system and environment [10–13].
The simplest approach to characterizing aspects of a quantum
Learning an abstracted generator of quantum dynamics, such
system is phenomenological fitting, which can yield useful pa-
as in the case of deep learning, may lead to accurate prediction
rameter values with minimal resources and carefully selected
of a quantum system’s evolution but often fails to provide
measurements. Phenomenological fitting has the advantage of
direct insight into physical parameters describing the system.
being readily applied to suitable data, but neglects detailed
Capitalizing on prior knowledge about the dominant physical
structure of the model describing a physical system and often
interactions allows for more explicit and efficient charac-
fails when closed-form analytic solutions are unknown. More
terization of quantum systems than abstracted tomographic
complex methods such as quantum state tomography [2] and
models and with greater flexibility than phenomenological
quantum process tomography [3] facilitate full characteriza-
fitting [14]. In such a physically motivated paradigm, pa-
tion of the states prepared in a device or how they are changed
rameter values for a physical model could be inferred from
by device operation, but often require large amounts of data
available measurements, and other aspects of the system could
as well as access to observables providing a full tomographic
be probed by inserting inferred values into the same model.
basis set. Several approaches have been developed to improve
The success of modern deep learning methods is under-
the efficiency of these tomographic techniques, or to extend
pinned by differentiable programming. This programming
their descriptive power, by having appropriate models incor-
paradigm facilitates the automatic differentiation required
porated into the characterization of both closed [4–6] and open
for training deep neural networks with many parameters,
quantum systems [7,8].
where numerical and symbolic derivatives become severely
Deep learning methods have recently opened many promis-
impractical [15,16]. Beyond applications to deep learning,
ing avenues for characterizing quantum systems. For instance,
differentiable programming has recently been applied to
physical models in the context of physics-aware machine
learning [17–19], efficient computation of gradients with
*
Contact author: [email protected] respect to complex physical models [20,21], optimization
of nonequilibrium steady states [22], and characterization
Published by the American Physical Society under the terms of the of non-Markovian dynamics [23]. Deep learning methods
Creative Commons Attribution 4.0 International license. Further have recently been combined with differentiable models of
distribution of this work must maintain attribution to the author(s) quantum systems to enhance quantum device characteriza-
and the published article’s title, journal citation, and DOI. tion [24,25]. There have also been recent advances in using

2643-1564/2024/6(4)/043175(15) 043175-1 Published by the American Physical Society

D. L. CRAIG, N. ARES, AND E. M. GAUGER PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

FIG. 1. (a) A schematic of the process of estimating parameters of an open quantum system using gradient descent. Here, HS , HE are
the system and environment Hamiltonians, respectively, and HI is the system-environment interaction Hamiltonian. Expectation values of an
observable Ô are generated from a ground truth model, which would be experimental data in practical applications, and a differentiable
model is used to fit these values by gradient descent of a loss function. (b) Representation of the first case study of a single quantum dot with
an excited state coupled to left and right leads. Here, μL(R) is the chemical potential of the left(right) lead, L(R) is the tunnel rate between
the dot states and the left(right) lead, and E0(1) is the electrochemical potential of the dot in the ground(excited) state. (c) Representation of
the second case study of a double quantum dot coupled to left and right leads, and a phonon bath. In each case study the observable used to
inform parameter estimation is the steady-state current from left to right lead. Here EL(R) is the chemical potential of the dot with a charge on
the left(right) dot, tc is the tunnel coupling between the dots, and γ p is the coupling strength to the phonon bath. (d) An illustrative schematic of
our approach for optimizing differentiable and nondifferentiable parameters when fitting to data. Here, we denote nondifferentiable parameters
as θ vnd , differentiable parameters as θ vdiff , and optimized parameters are indicated with an asterisk.

differentiable models for simulations of quantum systems gradient-based optimization protocols and Bayesian infer-
with a focus on optimal control [26–31], a requirement for ence, and is designed to handle a mixture of differentiable
the future of pulse engineering [32]. and nondifferentiable parameters to facilitate a range of
In this article we apply the concepts of differentiable applications. In addition to gradient-based optimization of
programming to develop a versatile, fast, and differentiable parameters, Markov chain Monte Carlo methods facilitate
quantum master equation solver. We apply our solver to elec- estimation of posterior distributions from which parameter
trostatically defined quantum dots as they offer a multi-faceted uncertainties can be obtained.
platform for realizing quantum systems for which accurate To demonstrate the power and broad applicability of our
parameter estimation would benefit experiments in quantum approach, we have designed two case studies with different
computation [33–35], quantum thermodynamics [36], and levels of complexity. The features probed in our case stud-
environment-assisted transport [37,38]. The interaction of dis- ies are not suitable for phenomenological fitting, and each
crete charge states with fermionic leads and vibrational modes case study is suitably complex that typical analytic solutions
requires an open quantum systems description, for which for quantum dots are not applicable. The first case study of
weak-coupling master equations can be a suitable choice in transport through a single quantum dot with an orbital excited
many experimental device regimes [38–40]. Measurements of state is designed to highlight our approach’s capabilities in
(steady state) charge transport in single and double quantum handling differentiable and nondifferentiable parameters. The
dot systems are often used to estimate characteristics of the second case study of transport through a double quantum dot
system, such as temperature [41–43], energy spectra, [44,45], coupled to a phonon bath is designed to explore character-
and electron-phonon coupling mechanisms [38,40,46]. How- ization of the environment [13,40,52]. In this second case
ever, characterization of electrostatically defined quantum dot study, we showcase the ability of our algorithm to discriminate
devices remains challenging, not least because of the fact that between and correctly select from a set of models featuring
control via voltages applied to gate electrodes introduces an different phonon spectral densities relating to different device
abstraction between experimental control and system param- materials and geometries. To quantitatively assess the perfor-
eters [1,47,48]. Further to this abstracted control, disorder mance of our approach, these case studies utilize simulated
frustrates device characterization by causing variation in the data. A sketch of the differentiable fitting process, which is
control parameters for devices of the same design [17,49–51]. central to our approach, is shown in Fig. 1(a) along with both
To address such demanding situations, our approach com- case study scenarios in (b) and (c), respectively, and an outline
bines a differentiable quantum master equation solver with of our approach in (d).

043175-2
DIFFERENTIABLE MASTER EQUATION SOLVER … PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

This article is structured as follows: First, we discuss parameter values, so that the output will be a nb × nout array,
our implementation of a differentiable quantum master where nout is the number of outputs from the model. An
equation solver in Sec. II, before discussing the Bayesian example demonstrating the accuracy and speed of our solver
formulation of our parameter estimation algorithm in Sec. III is outlined in Appendix A.
and optimization of nondifferentiable parameters in Sec. IV. Related studies have explored differentiable master equa-
We apply our approach to case studies of single and dou- tion solvers for time dynamics [23], and incorporated
ble quantum dots in Secs. V and VI, respectively. Finally, differentiable ODE solvers for learning dynamics of quantum
in Sec. VII we demonstrate the capability of our model to systems with recurrent neural networks [24]. The effective-
calculate time evolution of a double quantum dot system. ness of differentiable models of quantum systems enhanced
by machine learning is further demonstrated in [25] in the
II. DIFFERENTIABLE MASTER EQUATION SOLVER context of photonic devices. Our approach takes a different
direction to these works by considering steady-state solutions,
The focal point of our quantum device characterization and implementing a data efficient characterization algorithm,
is a differentiable quantum master equation solver, which is which also handles nondifferentiable parameters. Another re-
implemented using TensorFlow [53]. Quantum master equa- lated paper has demonstrated differentiable computation of
tions are a method of determining the time evolution of the steady-state solutions by time-evolution using JAX [59] for
system density matrix ρ of an open quantum system [54]. In the purpose of optimizing heat transfer models.
a weak-coupling setting this evolution is governed by the Li-
ouvillian superoperator L, which contains details of both the
system Hamiltonian and the system-environment interaction, III. BAYESIAN ESTIMATION
such that ρ̇(t ) = Lρ(t ). A simple manifestation of L is given OF DIFFERENTIABLE PARAMETERS
by a Lindblad master equation, which reads
Our approach aims to fit model predictions to data by es-

ρ̇(t ) = −i[HS , ρ(t )] + i D[Ai ]ρ(t ), (1) timating values for parameters in quantum master equations.
i
We consider case studies of transport through a single quan-
tum dot with an orbital excited state, and through a double
where HS is the system Hamiltonian, i is the dissipation quantum dot coupled to a phonon bath. For both the single
rate associated with the operator Ai , which describes environ- and double quantum dot case studies, the source-drain bias
ment induced transitions in the system, and D[A]ρ = AρA† − Vbias and a single 1D current scan are used to inform the
1/2{A† A, ρ} is the Lindblad dissipator acting on the density parameter estimation. Not all parameters in a model may be
matrix. The dissipation rates are often dependent on system differentiable as discussed below, and others may have mini-
Hamiltonian parameters defining the energy difference be- mal impact on a loss function so gradients become ineffective.
tween states involved in the environment induced transitions. We shall present our approach for dealing with such parame-
Numerically solving such a quantum master equation us- ters in general terms before presenting results for each case
ing differentiable programming facilitates evaluation of the study.
gradient of an observable with respect to the master equa- We denote the nc parameters, which are controlled to
tion parameters. To compute steady states, we solve the set produce a batch of nb data points as θ c , and the remaining
of linear equations resulting from Eq. (1) under the condition nv parameters under consideration for estimation as θ v . The
ρ̇(t ) = 0, and for time evolution we use ODE solvers as dis- ground truth data is then D = {d j | j = 1 . . . nb } where d j is an
cussed in Sec. VII. A differentiable model allows gradients individual data point, and the total set of nθ = nc + nv param-
to be obtained with similar computation time to forward eval- eters is θ = [θ c , θ v ]. We assume that the ground truth data D
uation of the model using automatic differentiation, without is described by a function of the true parameters f (θ true ) with
the need for inefficient finite-element methods. Solving the Gaussian noise of standard deviation σ such that each point in
steady state directly from the Liouvillian is critical to the D is d j ∼ N ( f (θ true ), σ ). Under this assumption we formulate
speed of our solver by avoiding the many operations and the likelihood of the data given a set of parameters θ v used to
storing of gradients required for a long-time integration of an estimate the function d̂ = f (θ v ) over the domain of θ c as
ODE. Our implementation can deal with arbitrary Liouvillian
superoperators, and is thus applicable to a range of physical  nb 
j=1 (d̂ j − d j )2
systems. P(D|θ v ) ∝ exp − . (2)
Using TensorFlow allows direct use of in-built gradient- σ2
based optimization methods such as Adam [55], and Bayesian
inference methods such as Hamiltonian Monte Carlo (HMC) This likelihood can be used in Bayes’ formula P(θ v |D) ∝
[56], which can be implemented using TensorFlow probability P(D|θ v )P(θ v ) to estimate the posterior values for parame-
[57]. Several differentiable programming libraries, including ters in θ v . The prior of each parameter
nv in vθ can be chosen
v

TensorFlow, have the advantage of allowing vectorized (i.e., independently such that P(θ ) = i=1 P(θi ). The set of pa-
v

batched) inputs and computation performed on a graph, which rameters, which produce the mode of the posterior distribution
leads to significant speed up when compared to the same is the maximum a posteriori (MAP) estimate, and for uniform
computation performed using a standard library for modeling priors this becomes maximum likelihood estimation (MLE).
quantum systems, such as QuTiP [58]. Vectorized inputs refer Evaluation of the negative log probability of the posterior
to having the input to the differentiable model be an nb × nθ distribution can be used as a differentiable loss function for
array for nθ parameters and a batch of nb different sets of the gradient descent portion of the parameter fitting process

043175-3
D. L. CRAIG, N. ARES, AND E. M. GAUGER PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

discussed in the previous section. The parameters, which min- V. CASE STUDY: SINGLE QUANTUM DOT
imise this loss correspond to the MAP estimate. WITH ORBITAL EXCITED STATE
Further to this, a Bayesian approach allows the use of
A. Model
Markov chain Monte Carlo (MCMC) methods to estimate
the posterior distribution for each parameter. We specifically We model a single quantum dot (SQD) with an orbital
consider Hamiltonian Monte Carlo (HMC) implemented in excited state as shown in Fig. 1(b) such that the set of charge
TensorFlow probability [57] to generate a set of ns posterior states correspond to a maximum of one excess charge on the
samples for the parameter values S = {θ vj | j = 1 . . . ns }, using dot. This restriction is motivated by the large charging energy
the MAP estimate as a seed to reduce burn-in. These posterior of the quantum dot compared to the bias window and the
samples allow for an estimate of uncertainty and correlations energy splitting between the ground and first-orbital excited
in the parameter values. The posterior distribution can also be state. The available charge states are |0, |G, and |E , where
normalized and used to update the prior distribution for use |0 has an integer M charges on the dot, |G and |E  both have
with further data. M + 1 charges on the dot. The single excess charge exists
in the ground state for |G, and an orbital excited state for
|E . The Hamiltonian describing the coupling between these
IV. OPTIMIZING NONDIFFERENTIABLE PARAMETERS charge states and the leads is given by
The output of a model is a function of the parameters,
f (θ), which may not be differentiable with respect to certain H = H0 + Hl + Hel , (3)
elements of θ. In the case of current traces in electrostatically
defined quantum dots, such parameters are the definition of H0 = E0 |GG| + (E0 + δ)|E E |, (4)
the bias window in θ c due to the abstraction between gate 
voltages and the energy of the dot electrochemical potential Hl = (k − μl )dkl† dkl + (k − μr )dkr
†
dkr , (5)
relative to the lead chemical potentials. In other situations, k

f (θ) may not have well-defined gradients across the whole Hel = (tl dkl† c j + tr dkr
†
c j + H.c.), (6)
parameter space, which renders gradient descent methods im- k j=g,e
practical.
To deal with a mixture of differentiable and nondiffer- where E0 is the energy of the ground state, and δ is the
entiable parameters, we design an optimization procedure, energy splitting between the ground and orbital excited state.
which takes advantage of the fast differentiable model. The Hl describes the leads with μl (r) being the chemical potential
idea is to find the nondifferentiable parameters for which the of the left(right) leads, and dkl† (r) (dkl (r) ) are the fermionic
remaining differentiable parameters provide the best fit. A creation(annihilation) operators for the left(right) lead for the
Nelder-Mead optimization routine [60] finds appropriate val- mode of energy k . Hel describes the coupling between the dot
ues for the nondifferentiable parameters, where the function states and the leads where cg(e) is the fermionic annihilation
being optimized is a fit of the model to ground truth data operator acting on the ground(excited) dot, tl (r) are the tunnel
using the remaining differentiable parameters. This fit first couplings between the leads and the dot states, and H.c. de-
involves a log-spaced grid search of the differentiable ele- notes the Hermitian conjugate. For simplicity, we assume the
ments of θ v where the evaluation of f (θ) with the lowest loss same tunnel rates to the leads for both the ground and excited
with respect to the ground truth data is selected as a starting charge state.
point for a short gradient descent optimization using the Adam To model the steady-state current flowing through the SQD
optimizer. The evaluation of the function being optimized by at a given source-drain bias Vbias we solve a master equation of
the Nelder-Mead routine is simply the final loss after this the form
gradient descent. Finally, using the optimal nondifferentiable
parameters, conducting a longer gradient descent optimization ρ̇ = −i[H0 , ρ] + Lleads ρ, (7)
of the differentiable parameters provides an estimate of all where ρ is the density matrix of the SQD system and Lleads is
parameters in θ. A schematic of this optimization procedure is the dissipator associated with the leads. In the weak-coupling
shown in Fig. 1(d), and discussion of its scaling can be found regime Lleads takes the form
in Appendix B.
The model fitting algorithm discussed above is general to 
Lleads = {(Ws j + Wd j )D[| j0|]
the situation when some nondifferentiable parameters must
j=G,E
be optimized before optimizing differentiable parameters. For
our case studies, the nondifferentiable parameters define the + (W d j + W s j )D[|0 j|]}. (8)
energy scale of the measurement axis, which in turn defines
how parameters such as tunnel rates change the profile of the The tunnel rates from a lead onto the dot are given by WL(R) j =
fit. This is a result of dot electrochemical potentials being L(R) fL(R) (E j ) and from the dot onto the lead W L(R) j =
controlled by gate voltages in experiments, which do not L(R) [1 − fL(R) ( j )], where L(R) = π gL(R) |tl (r) |2 with gL(R)
have well-defined values relative to the bias window. In other the constant density of states in the leads in the wide
experimental settings, it may be the case that the measurement band approximation, and fL(R) () = [e(−μl (r) )/kB T + 1]−1 is
axis is easily quantifiable (e.g., the frequency of an oscillating the Fermi-Dirac distribution for the left(right) lead.
signal) and so the order in which differentiable and nondiffer- With e denoting the electronic charge and ρss the steady-
entiable parameters are optimized becomes less important. state solution of Eq. (7), the steady-state current flowing

043175-4
DIFFERENTIABLE MASTER EQUATION SOLVER … PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

TABLE I. Ground truth, optimal, and posterior means with

standard deviation errors for the example shown in Fig. 2. The
nondifferentiable parameters are not part of the HMC sampling and
therefore do not have posterior means and uncertainties.

Parameter Unit Ground truth Optimal Posterior mean ± Std.

μl pixels 15.4 15.3
μr pixels 96.6 96.4
δ meV 0.084 0.085
L MHz 18.1 18.0 18.4 ± 0.8
R MHz 183.1 180.6 171.3 ± 4.3
T mK 55.9 53.4 59.4 ± 1.8

method is generally effective. There are a small number of

outliers with large errors, but upon further inspection we find
FIG. 2. An example of our approach applied to a single quantum that most of these outliers have no discernible second peak in
dot with an orbital excited state. (a) A Coulomb diamond from which the current profile (see insets of Fig. 3, specifically Outliers
the current trace in (b) is taken, indicated by the black-dashed line. A, C, and D). This means the location of the excited state
(b) The example ground truth simulated current data with 100 fA cannot be determined at the first optimization step, which im-
Gaussian noise used to fit parameters, with the bias set to 0.109 mV. pacts the estimation of further parameters. The outliers have
Nominal locations of the points on the E0 axis defined by the non- a lower signal to noise ratio than typical samples, which may
differentiable parameters [μl , μr , δ] are indicated by vertical dashed further impact the parameter estimation, as evident in Outlier
lines. (c) The results of gradient descent parameter optimization on B, which does have an excited state peak.
[L , R , T ] after the nondifferentiable parameters have been fitted. From Fig. 3, we observe that parameter estimates using
(d) The HMC posterior samples of the current values generated using the posterior mean from HMC samples has a more narrow
500 posterior samples of the parameters [L , R , T ]. Ground truth distribution around zero-percentage error, and is therefore
and optimal parameter values, as well as posterior means with stan- generally better than the gradient descent optimal values. The
dard deviation errors for relevant parameters, are shown in Table I. additional information contained in the posterior distribution
such as spread relative to the mean, multimodal structures, and
through the SQD is correlations between parameters provide further insight into
 results of our approach. In Appendix C, we use the additional
I=e W jR Tr(ρss | j j|) − W jR Tr(ρss |00|). (9) information provided by posterior distributions to filter results
j=E ,G to remove uncertain estimates and ensure a higher accuracy in
accepted results.
B. Parameter estimation
We consider the model presented above as a case study VI. CASE STUDY: DOUBLE QUANTUM DOT
to assess the performance of each part of the parameter esti- A. Model
mation process. This model has a single controlled parameter
The second case study considers a double quantum dot
θ c = [E0 ], which is swept through the bias window. The vari-
(DQD) coupled to left and right leads and a phonon bath, as
able parameters θ v = [μl , μr , δ, L , R , T ] can be separated
shown in Fig. 1(c). The states describing the DQD are |0,
into differentiable parameters [L , R , T ] and nondifferen-
|L, and |R where |0 has (M, N ) charges on the dot, |L
tiable parameters [μl , μr , δ]. The left and right chemical
has (M + 1, N ) charges, and |R has (M, N + 1) charges. The
potentials are nondifferentiable as they define the energy scale
Hamiltonian describing the coupling between these charge
of the data before the model is computed, and the orbital
states, the leads, and the phonon bath is given by
excited state energy splitting δ has zero gradient for large areas
of the parameter space and is thus optimized as a nondifferen- H = H0 + Hl + Hp + Hel + Hep , (10)
tiable parameter.

An example of our parameter estimation algorithm is H0 = σz + tc σx , (11)
shown in Fig. 2 with parameter estimates shown in Table I. 2

In this example, the largest absolute percentage error in Hl = (k − μl )dkl† dkl + (k − μr )dkr
†
dkr , (12)
optimized nondifferentiable values is 1.2%, and in optimal
differentiable values is 4.5%. Posterior means and standard
k

Hp = ωk ak† ak , (13)
deviations drift slightly from the optimal parameter values,
but retain a good fit to the data. In Fig. 3 we consider the  k

distributions of percentage errors for each parameter for a set Hel = (tl dkl† cl + tr dkr
†
cr + H.c.), (14)
of 100 iterations of our parameter estimation algorithm on k

randomly generated data. The distributions are all peaked at Hep = σz λk (ak† + ak ), (15)
zero percentage error, indicating that our parameter estimation k

043175-5
D. L. CRAIG, N. ARES, AND E. M. GAUGER PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

FIG. 3. Distributions of percentage errors for parameter estimation over a test set of 100 randomly generated current traces for a single
quantum dot with an orbital excited state. Distributions displayed are: (a)–(c) optimal values of nondifferentiable parameters [μl , μr , δ],
(d)–(f) optimal values of differentiable parameters [L , R , T ], and (g)–(i) posterior means for the differentiable parameters. All distributions
are peaked at zero error, with a small number of outliers. Insets show the target data (red points) and optimal fit (blue line) for selected outliers.
Summary statistics of these distributions can be found in Appendix C.

where σμ are the Pauli operators in the subspace |L and |R, where the tunnel rates from a lead onto its neighboring dot
 = EL − ER is the detuning between the dot energy levels, follow equivalent definitions as for the previous case study.
and tc is the interdot tunnel rate. Hl describes the leads with The phonon dissipator is most conveniently expressed in
μl (r) being the chemical potential of the left(right) leads, and terms of the eigenstates of H0 , which for positive detuning are
dkl† (r) (dkl (r) ) are the fermionic creation(annihilation) operators
|+ = cos(θ /2)|L − sin(θ /2)|R, (18)
for the left(right) lead for the mode of energy k . Hel describes
the coupling between the dots and the leads where cl (r) is
|− = sin(θ /2)|L + cos(θ /2)|R, (19)
the fermionic annihilation operator acting on the left(right)

dot, and tl (r) are the tunnel couplings between the leads and with energy h̄ω± = ±  2 /4 + tc2 . The energy splitting
the dots. The phonon bath is described by Hp where ak† (ak ) between the eigenstates is h̄ω p = h̄(ω+ − ω− ), and θ =
are the creation(annihilation) operators for the phonon mode arctan(2tc /). The weak-coupling phonon dissipator [39] is
of angular frequency ωk . The electron-phonon interaction is then
described by Hep where λk is the coupling constant between
the dots and the phonons. Lphonons = γ (ω p )([1 + n(ω p )]D[|−+|]
To model the steady-state current flowing through the DQD + n(ω p )D[|+−|]), (20)
at a given source-drain bias Vbias = μl − μr we solve a master
equation of the form where γ (ω) = sin2 θ J (ω) is the rate at which the phonon bath
induces interdot charge transitions in the DQD, dictated by the
ρ̇ = −i[H0 , ρ] + Lleads ρ + Lphonons ρ, (16) phonon spectral density
where ρ is the density matrix of the DQD system, and Lleads 
J (ω) = 2π |λk |2 δ(ω − ωk ). (21)
and Lphonons are the dissipators associated with the leads and
k
phonon bath, respectively. In the weak-coupling regime Lleads
takes the form The functional form of the phonon spectral density can be
modelled as
Lleads = WL D[|L0|] + W L D[|0L|]  
ω n−α ω
e−ω /2ωa ,
2 2
+ WR D[|R0|] + W R D[|0R|], (17) J (ω) = J0 1 − sinc (22)
ωd ωd

043175-6
DIFFERENTIABLE MASTER EQUATION SOLVER … PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

where ωd = 2πc d
and ωa = 2πc a
, with d the center-to-center
separation of the dots, a the linear extent of the dots, and c
the speed of sound in the crystal, and J0 is a scale factor. We
also have n as the dimension of the system, and α = 0 for
piezoelectric coupling and α = 2 for deformation potential
coupling to phonons. The complete set of spectral densities
for varying dimensions and coupling mechanism is shown in
Appendix E.
The Bose-Einstein occupation of the phonon bath n(ω) =
[eω/kB T − 1]−1 produces a temperature dependence on the
phonon emission and absorption processes that are described
by D[|−+|] and D[|+−|], respectively. As the DQD
system is at low temperature (kB T  ω p ), phonon emission
processes dominate.
With e denoting the electronic charge, using the steady-
state solution of Eq. (16), ρss , the steady-state current flowing
FIG. 4. An example of our approach applied to a double quantum
through the DQD is dot coupled to a phonon bath. (a) A bias triangle from which the
current trace in (b) is taken, indicated by the black dashed line.
I = e[W R Tr(ρss |RR|) − WR Tr(ρss |00|)]. (23) (b) The example ground truth simulated current data with 100 fA
Gaussian noise used to fit parameters, with the bias set to 0.241 mV.
B. Model selection Nominal locations of the points on the E0 axis defined by the non-
differentiable parameters [μl , μr ] are indicated by vertical dashed
We now consider the model presented above to assess the lines. (c) The results of gradient descent parameter optimization on
performance of our approach on a more complex system. This [L , R , tc , J0 , T ] after the non-differentiable parameters have been
model has a single controlled parameter θ c = [], which is fitted. (d) The HMC posterior samples of the current values generated
varied such that each dot energy level is swept through the using 500 posterior samples of the parameters [L , R , tc , J0 , T ].
bias window. We choose the variable parameters to be θ v = This example considers a 3-dimensional piezoelectric phonon spec-
[μl , μr , L , R , tc , J0 , T ], which can be separated into differ- tral density with fixed parameters are cs = 3000 ms−1 , a = 20 nm,
entiable parameters [L , R , tc , J0 , T ] and nondifferentiable and d = 100 nm. Ground truth and optimal parameter values, as
parameters [0 , bias ], which are the points where  = 0 and well as posterior means with standard deviation errors for relevant
 = μL − μR , respectively. The remaining parameters are set parameters, are shown in Table III.
to representative fixed values of cs = 3000 ms−1 , a = 20 nm,
and d = 100 nm. We choose these parameters to be fixed as
they can all be estimated from the device design, but they mation potential, 2D and 3D piezoelectric) on datasets with
could also be included in the set of differentiable parameters. different ground truth spectral densities. We use the negative
The dot radius a in particular has minimal impact on the fit log likelihood, − log P(D|θ c ), as a metric for our fit. As the
for the energy scales probed by typical experimental measure- data has Gaussian noise, the value of the negative log like-
ments. The problem of estimating parameters for this model lihood indicates the number of noise standard deviations a
from a single current trace is expected to be underdetermined given fit is from the true value. In Fig. 5(a) we display the
because of the number of parameters defining the system. We frequency of fits, which are within two standard deviations
find fitting to data points is accurate, with an example shown of the true value on average. Having the largest values along
in Fig. 4. Discussion of parameter estimation for this model the diagonal indicates that our fit is only successful when the
can be found in Appendix D. correct model is used, providing evidence that the functional
Capitalizing on the accuracy of fitting to data, we use our form of the spectral density adds a significant constraint on
approach to characterise other aspects of a device hosting the fit. The failed fits, which are on average more than five
these dots. We focus on model selection for the interaction standard deviations from the true value are shown in Fig. 5(b),
of the double quantum dot with phonons, using our fitting and in this case the diagonal has the lowest values indicating
process to perform model selection on the phonon spectral that the fitting process does not fail often when using the cor-
density. The phonon spectral density is determined by ma- rect model. We can also observe from this plot that the fitting
terial properties and dimensionality of the semiconductor in process is notably worse when fitting piezoelectric spectral
which the dots are defined, as well as the geometry of the densities to data with deformation potential spectral densities,
dots, as outlined in Appendix E. Accurately distinguishing and also performs poorly when fitting deformation potential
between piezoelectric and deformation potential phonon cou- spectral densities to data generated using piezoelectric spec-
pling would yield information about the device in which tral densities in 3D. Fits, which are between two and five
the dots are formed. Determining the dimensionality of the standard deviations of the ground truth could still be classed as
phonon spectral density would yield information about the successful but may be more subjective, so we do not consider
confinement of phonon modes in the device and help identify them useful when assessing our fitting of different spectral
any possible waveguide effects. densities.
In Fig. 5 we present the results of performing our fitting We find that our approach provides useful insight into the
process with each phonon spectral density (2D and 3D defor- electron-phonon coupling mechanism in a double quantum

043175-7
D. L. CRAIG, N. ARES, AND E. M. GAUGER PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

FIG. 5. The occurrence of fits, which have a negative log like-

lihood of (a) less than 2, and (b) greater than 12.5 in a test
set of 50 for each ground truth model. In each ground truth
model a different phonon spectral density is used, with “3D:
D” being three-dimensional deformation potential, “3D: P” three-
dimensional piezoelectric, “2D: D” two-dimensional deformation
potential, and “2D: P” two-dimensional piezoelectric. With Gaussian
noise, − log P(D|θ c ) < 2 indicates that the average error is less than FIG. 7. The time evolution of a double quantum dot system
two noise standard deviations from the ground truth, which we con- coupled to left and right leads and a phonon bath with a time-
sider to be a good fit. Fits with − log P(D|θ c ) > 12.5 indicates that independent Hamiltonian. The upper plots display the current as a
the average error is more than five noise standard deviations, repre- function of detuning at several points along the evolution, with the
senting a failure of fitting. We fit the nondifferentiable parameters steady-state current (red) shown for comparison. The evolution is
[0 , bias ] and the differentiable parameters [L , R , tc , J0 , T ]. Fixed computed as a batch of nb = 250 detuning values and evolved for
parameters are cs = 3000 ms−1 , a = 20 nm, and d = 100 nm. 15 ns with a step-size of 0.01 ns. The initial state√for the entire
batch is the pure state, |ψ (0) = (|0 + |L + |R)/ 3. Parameter
values used are, Vbias = 0.2 mV, L = R = 500 MHz, tc = 4 GHz
and T = 100 mK. The phonon coupling follows a 3D piezoelectric
dot system, even when individual model parameters cannot be spectral density with cs = 3000 ms−1 , a = 20 nm, and d = 50nm,
accurately determined. In order to apply this strategy to real and J0 = 1 GHz.
data where the ground truth is not available, using a batch of
size nbatch detuning traces from different bias triangles allows
a decision to be made when selecting a model based on the
frequencies of fit outcomes displayed in Fig. 5. Defining a
score function Sbatch = (nsuccess − nfailure )/nbatch where nsuccess VII. TIME EVOLUTION
is the number of fits in a batch with − log P(D|θ c ) < 2 and
nfailure is the number of fits in a batch with − log P(D|θ c ) > Having considered parameter estimation from steady-state
12.5 provides reliable insight into the correct phonon coupling solutions to quantum master equations using our differentiable
mechanism for simulated test data as shown in Fig. 6, which solver, this section briefly explores time-evolution functional-
considers the case of nbatch = 7. ity and its limitations. Dynamic control of qubits is a natural
motivation for studying time evolution, and thus we consider
the model of a double quantum dot coupled to fermionic
leads and a phonon bath outlined above as charge qubits have
been implemented in such systems [61,62]. A Lindblad master
equation can be recast as a set of coupled ordinary differential
equations (ODEs) for time evolution solutions, which means
standard ODE solvers can be used. Our solver implements
a fourth-order Runge-Kutta method [63] to compute time
dynamics. In Appendix A, we compare performance of our
ODE solver with direct computation of the steady state as
used for previous results. To verify that our solver is accu-
rate, Fig. 7 shows that the steady-state solution is reached
after a suitable evolution time under the influence of a time-
independent Hamiltonian. This figure also demonstrates batch
computation of time evolution, whereby a batch of nb = 250
FIG. 6. The score of fitting to a batch of size nbatch = 7 for each detuning values is√evolved from an initial state of |ψ (0) =
phonon spectral density considered in Fig. 5. The batch score Sbatch = (|0 + |L + |R)/ 3 over 15 ns for 1500 time steps in 0.38s
(nsuccess − nfailure )/nbatch for seven independent batches is plotted for using the same hardware as discussed in Appendix A.
each fitted spectral density. A positive score indicates that the spectral In addition to evolution under a time-independent Hamil-
density fit succeeds more than it fails within the batch, in accordance tonian, our solver can handle time-dependent Hamiltonians.
to the definitions of success and failure discussed in the text. We see We demonstrate this functionality by considering a Rabi pulse
that the best scores for each batch correlate with the ground truth sequence on the DQD system operating as a charge qubit,
spectral density. with the details and results of this pulse sequence shown in

043175-8
DIFFERENTIABLE MASTER EQUATION SOLVER … PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

from its capability to compute a batch of outputs in a single

evaluation of a model. This speed in turn facilitates more
involved algorithms requiring many evaluations, which would
be impractical using standard solvers. An example of such an
algorithm is the characterization approach we develop, which
employs both gradient-based and gradient-free optimization,
and MCMC methods.
Our characterization algorithm has been applied to case
studies of transport through a single quantum dot with an
orbital excited state, and a double quantum dot coupled to
a phonon bath. We have found our approach to be success-
ful for dealing with parameter estimation in the case of a
single quantum dot using a single current trace, although
using an equivalent trace only provided relatively poor accu-
racy if applied to parameter estimation for a double quantum
FIG. 8. Demonstration of a Rabi pulse sequence on a DQD dot. However, we anticipate that our approach is capable
charge qubit computed using our quantum master equation solver. of successfully characterizing more complex systems when
(a) Outline of the sequence whereby the DQD is initialized to be presented with a broader range of input data; in the case of
|ψ (0) = |L and after a time τi = 1 ns a pulse is applied to the a double quantum dots this would, for instance, be a set of
right dot to vary the detuning by δ for a duration of τ p , and cur- sequential traces based on knowledge of the system structure
rent is measured for a time τm = 1 ns. (b) Schematics of the DQD as discussed in Appendix D.
in the initialization and measurement regime (top and the pulsed To highlight the utility of our algorithm beyond param-
regime (bottom). Rabi chevrons without (c) and with (d) coupling eter estimation, we have also performed model selection
to a phonon bath. We define the pulse detuning  p = |δ| − |0 | to determine the electron-phonon coupling mechanism in
and the measured current is taken to be the mean current integrated a double quantum dot. Our approach is successful in re-
during the measurement period. Parameter values used are, L = liably determining the ground truth spectral density. Our
R = 500 MHz, tc = 4 GHz and T = 100 mK. The phonon coupling approach could be directly applied to experimental studies
follows a 3D piezoelectric spectral density with cs = 3000 ms−1 , of the spectral density, such as in Ref. [40]. Decoher-
a = 20 nm, and d = 50 nm, and J0 = 1 GHz. ence from electron-phonon interactions impacts both charge
and spin qubits, therefore characterizing the details of this
coupling mechanism is important for understanding qubit
Fig. 8. We initialize the DQD such that the charge is on the dynamics.
left dot, with its electrochemical potential remaining below Our differentiable master equation solver and its appli-
the source and drain chemical potentials for the duration of cation to parameter estimation contributes to the field of
the sequence. The electrochemical potential of the right dot is characterizing quantum systems using differentiable models
pulsed by an energy δ for a duration τ p , which defines the [23–25]. Our approach differs from existing methods by uti-
Rabi pulse. We then simulate the integration of current for a lizing significantly fewer data points and by dealing with a
duration of τm and output the mean current during this period. combination of differentiable and nondifferentiable param-
Successfully reproducing the expected charge qubit dynamics eters. These attributes are key for the characterization of
under time-independent and time-dependent Hamiltonians is more general systems where experimental data may be time
a demonstration that our fourth-order Runge-Kutta solver is consuming or difficult to acquire. Physics-aware machine
sufficient for the phenomena considered; however, more so- learning is a notable area in which differentiable models have
phisticated solvers could make computation of dynamics more been applied previously [17–19], and our solver stands to
efficient [63]. We note that while batch computation of time widen the impact of such approaches, particularly in the field
evolution is facilitated by our solver, providing a speed im- of quantum transport.
provement over conventional solvers, gradients with respect
to pulse parameters would require further work to implement The code used in this paper is available online at Zenodo
a differentiable model for optimal quantum control. [64].

VIII. CONCLUSIONS ACKNOWLEDGMENTS

We have presented a fast and differentiable quantum master We acknowledge M. Cygorek, G. A. D. Briggs, S. C.
equation solver, and demonstrated its capability to compute Benjamin, and C. Schönenberger for useful discussions dur-
steady-state solutions and time evolution of open quantum ing the development of this manuscript. This work was
systems. Our solver can be applied to any system described supported by Innovate UK Grant No. 10031865, the Royal
by a Lindblad master equation, and minimal changes would Society (URFR1191150), the European Research Council
permit application to a more general time-local approach in- (Grant Agreement No. 948932), EPSRC Platform Grant ( No.
volving fewer approximations. The speed of our solver comes EP/R029229/1), and EPSRC Grant No. EP/T01377X/1.

043175-9
D. L. CRAIG, N. ARES, AND E. M. GAUGER PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

APPENDIX A: DIFFERENTIABLE SOLVER TEST CASE

To demonstrate the effectiveness of our differentiable
solver, we consider the case of a single quantum dot (SQD)
coupled to left and right leads. Considering the basis states
|0 for N electrons on the dot, and |1 for N + 1 electrons on
the dot, the total Hamiltonian is
H = H0 + Hl + Hel , (A1)

H0 = |11|, (A2)

Hl = (k − μl )dkl† dkl + (k − μr )dkr
†
dkr , (A3)
k

Hel = ti (dki† c + dki c† ), (A4) FIG. 9. The steady-state current and its derivatives using (a) ana-
k i=l,r lytic solutions and (b) the differentiable solver. Parameters are set
as L = 150 Hz, R = 200 Hz, and T = 100 mK. A quantitative
where  is the chemical potential of the dot, dkl† (r) and dkl (r) comparison of the numerical outputs and analytic solutions is shown
are the creation and annihilation operators for the kth mode in in (c) using the logarithm of the relative absolute error. All plots share
the left(right) lead, and tl (r) is the tunnel coupling between the the same legend.
dot and modes in the left(right) lead. The operators c = |01|
and c† = |10| act on the quantum dot occupation. To model
the steady-state current flowing through the SQD at a given where
source-drain bias Vbias = μl − μr we solve a master equa- 
∂ fl (r) ()  − μl (r)  − μl (r)
tion of the form = cosh−2 . (A12)
∂T kB T 2 2kB T
ρ̇ = Lρ = −i[H0 , ρ] + Lleads ρ, (A5)
where ρ is the density matrix of the SQD system, L is the Li- Values of Iss and its derivatives computed using the dif-
ouvillian superoperator, and Lleads is the dissipator associated ferentiable solver are compared with the analytic results in
with the leads. In the weak-coupling regime Lleads takes the Fig. 9, displaying good agreement in relative error across the
form domain of dot energy . We use 32 bit precision for real valued
elements of the differentiable solver, with complex values
Lleads ρ = −(WL D[c† ]ρ + W L D[c]ρ + WR D[c† ]ρ
defined using two 32 bit floats.
+ W R D[c]ρ), (A6) Having established that our differentiable solver returns the
desired values, we now investigate the speed of the model
where the Lindblad superoperators act according to D[A]ρ =
and how computation time scales with the batch size, nb .
AρA† − 1/2{A† A, ρ}. The energy-dependent tunnel rates
The numerical results presented in Fig. 9 were performed by
from a lead onto the dot are given by WL(R) = L(R) fl (r) () and
sending a batch of nb = 200 parameter values with varying
from the dot onto the lead W L(R) = L(R) [1 − fl (r) ()], where dot energy  to our solver. We compare the evaluation of
 is the electrochemical potential of the dot and fl (r) () = the output on the same hardware [Intel(R) Core(TM) i5-9500
[e(−μl (r) )/kB T + 1]−1 is the Fermi-Dirac distribution for the CPU] with QuTiP, an established library for solving quantum
left(right) lead. master equations [58]. As shown in Fig. 10(a), our differen-
With e denoting the electronic charge and ρss the steady- tiable model computes batch outputs significantly faster than
state solution of Eq. (A5), the steady-state current flowing QuTiP (where batch computation scales linearly) across all
though the SQD is computed as tested batch sizes. Tracking and evaluating gradients adds an
Iss = e W R Tr(ρss |11|) − WL Tr(ρss |00|) . (A7) overhead to computation time, which is much more noticeable
This model has the benefit of a simple analytic solution for at low batch sizes. Considering the computation time per out-
the steady-state current, which can be used to directly com- put, Fig. 10(b) demonstrates the advantage of batched inputs
pare with the output and gradients generated by our solver. with our solver where the time per output decreases to less
The steady-state current is than 10 µs. These fast computation times make our model
ideal for computing a vector of outputs and evaluating the
WLW R − W LWR
Iss = e (A8) desired gradients.
L + R We also consider how the computation of the steady state
with gradients with respect to L , R , and T being scales with the dimension of the Hilbert space. This scaling
∂Iss eR2 is shown in Fig. 11, where we find that our solver scales with
= [ fl () − fr ()], (A9) Hilbert space size Hdim as ∼Hdim 4
. This is as expected from
∂L (L + R )2
the system of equations being solved in ρ̇(t ) = 0 growing as
∂Iss eL2 Hdim
2
. Given the timescales shown in Fig. 11, we expect our
= [ fl () − fr ()], (A10)
∂R (L + R )2 characterization approach to be effective in practical times
for systems of up to six dimensions; however, larger systems
∂Iss eL R  ∂ fl () ∂ fr ()  could be considered if characterization time is not a constraint.
= − , (A11)
∂T L + R ∂T ∂T For the quantum dot systems considered in our case studies,

043175-10
DIFFERENTIABLE MASTER EQUATION SOLVER … PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

FIG. 10. Batch size scaling of direct computation of the steady FIG. 12. Batch size scaling of computing the steady state using
state. The total time (a) and the time per output (b) for computations an ODE solver. The total time (a) and the time per output (b) for
on a range of batch sizes using the differentiable master equation for computations on a range of batch sizes using the differentiable master
the single quantum dot model described in Eq. (A5). Each point is an equation for the single quantum dot model described in Eq. (A5)
average of 100 evaluation times and the time per output is simply the using an RK4 ODE solver. The ODE is solved for an initial state
total time divided by the batch size. Output refers to the time taken of |1 for 10 ns with 1000 time steps. Each point is an average
to compute the output of a batch of input parameters. To compute of 100 evaluation times and the time per output is simply the total
a gradient by automatic differentiation, a variable must be tracked time divided by the batch size. Output refers to the time taken to
during the output computation, and once the output is available the compute the output of a batch of input parameters. Both plots share
gradient can be evaluated. Both plots share the same legend. The the same legend. The resource intensive nature of this computation
black lines corresponding to QuTiP execution times are extrapolated leads to unexpected scaling beyond a batch size of ∼102 , and the
from the time taken for a single computation. last two points in each series corresponding to gradient tracking and
associated output are absent as the CPU resources were exhausted
a six-dimensional system would equate to a chain of five for these batch sizes.
quantum dots in the single excitation manifold.
As our results consider steady-state solutions, we compare
APPENDIX B: ALGORITHM SCALING
the speed of finding the steady state using direct computation,
as shown in Fig. 10, with a long-time integration to the steady We display the scaling of each constituent part of our
state using an ODE solver. As discussed in the main text, algorithm in Fig. 13, namely the grid search over variables
we implement a fourth-order Runge-Kutta (RK4) method in and final gradient descent optimization. The grid search and
TensorFlow for time dynamics for which the scaling with gradient descent each take times of a similar magnitude;
batch size is shown in Fig. 12. It is clear from the comparison however, the the grid search scales less favorably as can be
of batch size scaling in Figs. 10 and 12 that direct computation expected from a simple grid search. Our algorithm has several
of the steady state is much faster and more memory efficient hyperparameters, which can be tuned depending on require-
than using an ODE solver. More advanced ODE solvers may ments for speed and accuracy. The most relevant parameters
find some improvement over our implementation; however, for scaling are the number of points in the grid search over
because of the necessarily larger number of operations in each variable, and the number of gradient descent steps in both
evolving an ODE it is unlikely that they would achieve equiv- the grid search and final gradient descent optimization. The
alent performance to direct computation of the steady state. overall algorithm is then subject to the scaling of the Nelder-
Mead optimizer. We also note that Bayesian optimization is
a valid alternative to a Nelder-Mead optimizer. We selected

FIG. 11. Hilbert space scaling of the direct computation of the

steady state. (a) Demonstrates the same outputs as in Fig. 10 and FIG. 13. Time taken to perform (a) gradient descent optimization
how their evaluation time scales with Hilbert space size with a batch of variables and (b) the grid search over variables for the model
size of 100. (b) Displays the output time scaling with batch size and presented in Sec. VI. Each point uses the same algorithm hyper-
Hilbert space size. Each point in the average of 100 repetitions in parameters as for the results in Sec. VI and is the average of 10
(a) and 10 repetitions in (b). The Hilbert space dimension is increased repetitions. The relevant hyperparameters are: in (a), 2500 gradient
by padding the basis vectors defining the single dot system described descent steps, and in (b), seven points in the grid search for each
in Eq. (A5). variable and 400 gradient descent steps to compute the loss.

043175-11
D. L. CRAIG, N. ARES, AND E. M. GAUGER PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

TABLE II. Summary statistics (mean and standard deviation)

for the distributions displayed in Fig. 3. As the standard deviation
is sensitive to outliers, we remove values greater than 3 standard
deviations away from the mean before recomputing the mean and
standard deviation presented here. The number of outliers identified
is 1−3% for each distribution.

% Error (mean ± standard deviation)

Parameter Optimal Posterior mean
μl −0.01 ± 0.64 FIG. 14. (a) The distribution of percentage errors over a test
μr −0.00 ± 0.08 dataset using posterior mean estimates as a function of threshold
δ −0.72 ± 3.59 value t used for filtering results. For each parameter, the dashed line
L −0.60 ± 18.83 −0.87 ± 16.59 is the mean and the shaded area represents the mean plus/minus the
R −1.61 ± 12.62 −1.76 ± 12.24 standard deviation. (b) The percentage of accepted results using the
T 0.96 ± 8.28 −0.09 ± 5.40 θ
filtering condition |σpost /μθpost |  t for all differentiable parameters.

APPENDIX D: DOUBLE QUANTUM DOT

PARAMETER ESTIMATION
Nelder-Mead over Bayesian optimization as the more reliable
approach for our application. If one performs sequential esti- Table III displays the parameter estimation results for the
mates using Bayesian updates to priors using posteriors, the example shown in Fig. 4. The bias is set to 0.241 mV, and
approach would scale linearly with the number of updates. the remaining parameters are set to representative fixed values
Gradient-free optimization protocols, such as Nelder-Mead of cs = 3000 ms−1 , a = 20 nm, and d = 100 nm. The per-
and Bayesian optimization, do not scale well with number of formance of this example is representative of our approach’s
parameters, which may limit aspects of our approach to more parameter estimation for this model. Our approach is success-
complex systems. As discussed in Appendix D, we anticipate ful in identifying the values of 0 and bias in this example,
that breaking more complex systems into smaller character- and the percentage error in these values across the case study
ization steps with fewer parameters would circumvent this test set is 3.48% for 0 and −0.55% for bias . Identifying these
poor scaling. points accurately and automatically is of benefit to experimen-
tal applications where the bias window must be known for
pulsing qubits [33,61,62].
The combination of an accurate fit and inaccurate param-
eter estimates indicate that this problem is underdetermined
APPENDIX C: ERRORS AND POSTERIOR FILTERING using the available data, as expected. To improve the per-
formance of parameter estimation additional data would be
The mean and standard deviation values for the distribu-
required to constrain the fit. An example of such additional
tions displayed in Fig. 3 relating to the first case study are
data could be a current scan along the base of the bias trian-
shown in Table II. This table highlights that our parameter
gle, perpendicular to the detuning scan shown in Fig. 4. Our
estimates are not strongly biased and have good consistency.
Bayesian formulation then allows for posterior distributions
The posterior mean errors have a lower standard deviation
of values for parameters from one fit to be used as priors for
than optimal value errors, particularly in the case of tempera-
the subsequent fit.
ture.
While the percentage error distributions for L and R
have standard deviations, which are not unreasonable for
parameters, which can vary by orders of magnitude in an ex-
TABLE III. Ground truth, optimal, and posterior means with
perimental setting, the precision can be improved. To achieve
standard deviation errors for the example shown in Fig. 4. The
this improvement, we capitalise on the extra information con-
nondifferentiable parameters are not part of the HMC sampling and
tained in the posterior distributions to reject samples with high therefore do not have posterior means and uncertainties.
parameter uncertainty. For each parameter posterior distribu-
tion, P(θ |D), we use the mean μθpost and standard deviation Parameter Unit Ground truth Optimal Posterior mean ± Std.
θ
σpost to accept a result if all parameters fulfill the condi-
θ 0 pixels 10.4 10.3
tion |σpost /μθpost |  t, where t is a threshold value as shown bias pixels 45.5 45.6
in Fig. 14. By choosing t = 0.05, the error distributions are L MHz 854.1 127.0 126.2 ± 5.4
significantly narrowed such that L : −0.81 ± 3.72%, R : R MHz 166.4 529.5 540.9 ± 38.2
1.58 ± 6.05%, and T : 0.17 ± 2.68%, while accepting 59% tc GHz 1.49 0.91 0.94 ± 0.04
of results. This is a useful approach when estimation accuracy J0 GHz 0.98 3.18 2.96 ± 0.3
and precision are prioritized, similar to threshold choices in T mK 143.7 136.5 126.1 ± 24.7
classifiers to optimize true-positive rates [17].

043175-12
DIFFERENTIABLE MASTER EQUATION SOLVER … PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

TABLE IV. A summary of the phonon spectral density functions for deformation potential and piezoelectric interactions with a double
quantum dot in varying dimensions. A similar table of spectral densities can be found in Ref. [39].

Phonon spectral density: J (ω)

Dimension Deformation potential Piezoelectric
2 2
3 − ω2 − ω2
3D D2 h̄
2π 2 μc2 d 3
( ωω ) (1 − sinc ωω )e 2ωa P2 h̄ ω
2π 2 μc2 d ωd
(1 − sinc ωω )e 2ωa
d d d
2 2
2 − ω2 − ω2
2D D2 h̄
4π 2 μc2 d 2
( ωω ) [1 − J0 ( ωω )]e 2ωa P2 h̄
4π 2 μc2
[1 − J0 ( ωω )]e 2ωa
d d d
2 2
D2 h̄ ω ω − ω2 P2 h̄d ωd ω − ω2
1D 8π 3 μc2 d ωd
(1 − cos ωd
)e 2ωa
8π 3 μc2 ω
(1 − cos ωd
)e 2ω a

APPENDIX E: SPECTRAL DENSITIES where J0 (·) is the zeroth-order Bessel function of the first
kind. Changing variables to x = cos θ we have
The following discussion provides a brief derivation of
the spectral density functions considered in this article. In a    4
 
ψ̃ ω 1 − x 2 , ω x
1
spin-boson model such as that described by Hsb = H0 + Hp + vω2 
F (ω) = 2 3  
Hep , and assuming the bosonic environment is in thermal π cs 0  cs cs
equilibrium, the environment is completely ωd 
 described by the × 1 − J0 1−x 2 dx. (E5)
spectral density function J (ω) = 2π k |λk |2 δ(ω − ωk ) as cs
defined in (21).
The interaction Hamiltonian Hep has a diagonal (σz ) system Assuming a sharply peaked electron density simplifies the
operator by considering well defined quantum dots with min- integral as |ψ̃ (κ, kz )|4 ≈ 1. In the context of a double quantum
imal overlap of the dot wavefunctions, i.e., L|R  1. Using dot system this is a reasonable assumption as the dots are
the single particle wavefunctions, the phonon coupling matrix usually confined to a narrow region of space in all directions.
element λk is defined as The finite extent of the dot wavefunctions, assumed to be
λk = Mk (L|eik·r |L − R|eik·r |R) Gaussian ψ (r) ∝ e−r /2a , introduces a Gaussian decay from
2 2

 the Fourier transform of the dot wavefunction in E 2.

z √
= Mk dr[ψL (r) − ψR (r)]eik·r (E1) Finally, we make use of the integral 0 J0 ( z2 − x 2 )dx =
sin z to evaluate the angular form factor of the spectral density,
where Mk is a matrix element, which depends on the material
properties of the system.  2
v ω ω −ω2 /2ωa2
Adopting the notation k = (κ, kz ) and r = (ρ, rz ), and as- F (ω) = 1 − sinc e , (E6)
π cd ωd
2 2 ωd
suming the dots exist in the x-y plane with identical, radially
symmetric wavefunctions separated by a vector d such that
ψL (ρ, rz ) = ψ (ρ, rz ) ψR (ρ, rz ) = ψ (ρ − d, rz ) it follows that where ωd = cs /d and ωa = cs /a. A similar derivation can be
performed in 2D and 1D by reducing the dimensionality of the
λk = Mk (1 − eiκ·ρ )|ψ̃ (κ, kz )|2 . (E2)
angular integral.
Substituting this result into (21) we can derive the form The prefactor Mk depends on the type of phonon interac-
factor for the spectral density J (ω). We use a linear dispersion tion [65]. The total interaction matrix element is the sum of
relation, ωk = cs |k| = cs k where cs is the speed of sound in the deformation potential and piezoelectric interaction,
crystal, and omit the prefactor Mk to find the angular form
factor F (ω) based on the dot geometry,  −1/2
 1
Mk = (D|k| + iP), (E7)
F (ω) = δ(ω − ck)|ψ̃ (κ, kz )|4 |1 − eiκ·d |2 . (E3) 2Mωk
k

 We convert  the sum over wavevectors to a 3D integral, where M is the average mass of the unit cell, and D and P are
k → v
(2π )3
dk. Using the Jacobi-Anger expansion and per- the deformation potential and piezoelectric constants. Being
forming the radial and azimuthal integrals, it follows that out of phase, the contributions do not interfere to second
 π  4 order. The form of the respective spectral densities in (22) fol-
vω2  ω ω 
F (ω) =  ψ̃ sin θ , cos θ  lows directly from these results. A summary of the functional
2π 2 cs3 0  cs cs forms of the spectral densities for deformation potential and
 piezoelectric phonon coupling with a double quantum dot in
ωd
× 1 − J0 sin θ sin θ dθ , (E4) varying dimensions is shown in Table IV.
cs

[1] V. Gebhart, R. Santagati, A. A. Gentile, E. M. Gauger, D. Craig, [2] U. Leonhardt, Quantum-state tomography and discrete Wigner
N. Ares, L. Banchi, F. Marquardt, L. Pezzè, and C. Bonato, function, Phys. Rev. Lett. 74, 4101 (1995).
Learning quantum systems, Nat. Rev. Phys. 5, 141 (2023). [3] M. Mohseni, A. T. Rezakhani, and D. A. Lidar, Quantum-

043175-13
D. L. CRAIG, N. ARES, AND E. M. GAUGER PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

process tomography: Resource analysis of different strategies, [22] R. A. Vargas-Hernández, R. T. Chen, K. A. Jung, and P. Brumer,
Phys. Rev. A 77, 032322 (2008). Fully differentiable optimization protocols for non-equilibrium
[4] J. H. Cole, S. G. Schirmer, A. D. Greentree, C. J. Wellard, D. steady states, New J. Phys. 23, 123006 (2021).
K. L. Oi, and L.C. L. Hollenberg, Identifying an experimental [23] S. Krastanov et al., Unboxing quantum black box models:
two-state Hamiltonian to arbitrary accuracy, Phys. Rev. A 71, Learning non-Markovian dynamics, arXiv:2009.03902.
062312 (2005). [24] É. Genois, J. A. Gross, A. DiPaolo, N. J. Stevenson, G.
[5] S. J. Devitt, J. H. Cole, and L. C. L. Hollenberg, Scheme for Koolstra, A. Hashim, I. Siddiqi, and A. Blais, Quantum-tailored
direct measurement of a general two-qubit Hamiltonian, Phys. machine-learning characterization of a superconducting qubit,
Rev. A 73, 052317 (2006). PRX Quantum 2, 040355 (2021).
[6] J. Zhang and M. Sarovar, Quantum Hamiltonian identification [25] A. Youssry, Y. Yang, R. J. Chapman, B. Haylock, F. Lenzini, M.
from measurement time traces, Phys. Rev. Lett. 113, 080401 Lobino, and A. Peruzzo, Experimental graybox quantum sys-
(2014). tem identification and control, npj Quantum Inf. 10, 9 (2024).
[7] M. R. Jørgensen and F. A. Pollock, Exploiting the causal tensor [26] F. Schäfer, M. Kloc, C. Bruder, and N. Lörch, A differentiable
network structure of quantum processes to efficiently simulate programming method for quantum control, Mach. Learn.: Sci.
non-Markovian path integrals, Phys. Rev. Lett. 123, 240602 Technol. 1, 035009 (2020).
(2019). [27] L. Coopmans, D. Luo, G. Kells, B. K. Clark, and J. Carrasquilla,
[8] G. A. White, F. A. Pollock, L. C. Hollenberg, K. Modi, Protocol discovery for the quantum control of Majoranas by
and C. D. Hill, Non-Markovian quantum process tomography, differentiable programming and natural evolution strategies,
PRX Quantum 3, 020344 (2022). PRX Quantum 2, 020332 (2021).
[9] G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko, [28] N. Wittler, F. Roy, K. Pack, M. Werninghaus, A. S. Roy, D.
and G. Carleo, Neural-network quantum state tomography, J. Egger, S. Filipp, F. K. Wilhelm, and S. Machnes, Integrated
Nat. Phys. 14, 447 (2018). tool set for control, calibration, and characterization of quantum
[10] L. Banchi, E. Grant, A. Rocchetto, and S. Severini, Modelling devices applied to superconducting qubits, Phys. Rev. Appl. 15,
non-Markovian quantum processes with recurrent neural net- 034080 (2021).
works, New J. Phys. 20, 123030 (2018). [29] I. Khait, J. Carrasquilla, and D. Segal, Optimal control of quan-
[11] M. J. Hartmann and G. Carleo, Neural-network approach to tum thermal machines using machine learning, Phys. Rev. Res.
dissipative quantum many-body dynamics, Phys. Rev. Lett. 122, 4, L012029 (2022).
250502 (2019). [30] L. Coopmans, S. Campbell, G. De Chiara, and A. Kiely, Opti-
[12] E. Flurin, L. S. Martin, S. Hacohen-Gourgy, and I. Siddiqi, Us- mal control in disordered quantum systems, Phys. Rev. Res. 4,
ing a recurrent neural network to reconstruct quantum dynamics 043138 (2022).
of a superconducting qubit from physical observations, Phys. [31] F. Sauvage and F. Mintert, Optimal control of families of quan-
Rev. X 10, 011006 (2020). tum gates, Phys. Rev. Lett. 129, 050507 (2022).
[13] J. Barr, G. Zicari, A. Ferraro, and M. Paternostro, Spectral den- [32] Z. Liang, H. Wang, J. Cheng, Y. Ding, H. Ren, Z. Gao, Z. Hu,
sity classification for environment spectroscopy, Mach. Learn.: D. S. Boning, X. Qian, S. Han et al., Variational quantum pulse
Sci. Technol. 5, 015043 (2024). learning, in 2022 IEEE International Conference on Quantum
[14] R. C. Kurchin, Using Bayesian parameter estimation to learn Computing and Engineering (QCE), Broomfield, CO, USA
more from data without black boxes, Nat. Rev. Phys. 6, 152 (IEEE, Piscataway, NJ, 2022), pp. 556–565.
(2024). [33] R. Hanson, L. P. Kouwenhoven, J. R. Petta, S. Tarucha,
[15] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning and L. M. Vandersypen, Spins in few-electron quantum dots,
representations by back-propagating errors, Nature (London) Rev. Mod. Phys. 79, 1217 (2007).
323, 533 (1986). [34] G. Burkard, T. D. Ladd, A. Pan, J. M. Nichol, and J. R. Petta,
[16] A. G. Baydin, B. A. Pearlmutter, A. A. Radul, and J. M. Siskind, Semiconductor spin qubits, Rev. Mod. Phys. 95, 025003 (2023).
Automatic differentiation in machine learning: A survey, J. [35] A. Chatterjee, P. Stevenson, S. De Franceschi, A. Morello, N. P.
Mach. Learn. Res. 18, 5595 (2017). de Leon, and F. Kuemmeth, Semiconductor qubits in practice,
[17] D. L. Craig et al., Bridging the reality gap in quantum devices Nat. Rev. Phys. 3, 157 (2021).
with physics-aware machine learning, Phys. Rev. X 14, 011001 [36] M. Josefsson, A. Svilans, A. M. Burke, E. A. Hoffmann, S.
(2024). Fahlvik, C. Thelander, M. Leijnse, and H. Linke, A quantum-
[18] A. Pachalieva, D. O’Malley, D. R. Harp, and H. Viswanathan, dot heat engine operating close to the thermodynamic efficiency
Physics-informed machine learning with differentiable pro- limits, Nat. Nanotechnol. 13, 920 (2018).
gramming for heterogeneous underground reservoir pressure [37] G. Granger et al., Quantum interference and phonon-mediated
management, Sci. Rep. 12, 18734 (2022). back-action in lateral quantum-dot circuits, Nat. Phys. 8, 522
[19] L. G. Wright et al., Deep physical neural networks trained with (2012).
backpropagation, Nature (London) 601, 549 (2022). [38] T. R. Hartke, Y.-Y. Liu, M. J. Gullans, and J. R. Petta,
[20] H.-J. Liao, J.-G. Liu, L. Wang, and T. Xiang, Differentiable Microwave detection of electron-phonon interactions in a
programming tensor networks, Phys. Rev. X 9, 031041 (2019). cavity-coupled double quantum dot, Phys. Rev. Lett. 120,
[21] Y. Hu, L. Anderson, T.-M. Li, Q. Sun, N. Carr, J. Ragan-Kelley, 097701 (2018).
and F. Durand, DiffTaichi: Differentiable programming for [39] T. M. Stace, A. C. Doherty, and S. D. Barrett, Population
physical simulation, In International Conference on Learning inversion of a driven two-level system in a structureless bath,
Representations (ICLR, 2020). Phys. Rev. Lett. 95, 106801 (2005).

043175-14
DIFFERENTIABLE MASTER EQUATION SOLVER … PHYSICAL REVIEW RESEARCH 6, 043175 (2024)

[40] A. Hofmann, C. Karlewski, A. Heimes, C. Reichl, W. [51] I. Seidler et al., Tailoring potentials by simulation-aided design
Wegscheider, G. Schon, K. Ensslin, T. Ihn, and V. F. Maisi, of gate layouts for spin-qubit applications, Phys. Rev. Appl. 20,
Phonon spectral density in a GaAs/AlGaAs double quantum 044058 (2023).
dot, Phys. Rev. Res. 2, 033230 (2020). [52] M. J. Gullans, J. M. Taylor, and J. R. Petta, Probing
[41] C. W. Beenakker, Theory of Coulomb-blockade oscillations electron-phonon interactions in the charge-photon dynamics of
in the conductance of a quantum dot, Phys. Rev. B 44, 1646 cavity-coupled double quantum dots, Phys. Rev. B 97, 035305
(1991). (2018).
[42] J. Prance, Z. Shi, C. B. Simmons, D. E. Savage, M. G. Lagally, [53] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro,
L. R. Schreiber, L. M. K. Vandersypen, M. Friesen, R. Joynt, S. G. S. Corrado, A. Davis, J. Dean, M. Devin et al., TensorFlow:
N. Coppersmith, and M. A. Eriksson, Single-shot measurement Large-scale machine learning on heterogeneous distributed sys-
of triplet-singlet relaxation in a Si/SiGe double quantum dot, tems, arXiv:1603.04467.
Phys. Rev. Lett. 108, 046808 (2012). [54] H.-P. Breuer and F. Petruccione, The Theory of Open Quantum
[43] F. Borsoi, N. W. Hendrickx, V. John, M. Meyer, S. Motz, F. van Systems (Oxford University Press, New York, 2002).
Riggelen, A. Sammak, S. L. de Snoo, G. Scappucci, and Menno [55] D. P. Kingma and J. Ba, Adam: A method for stochastic opti-
Veldhorst, Shared control of a 16 semiconductor quantum dot mization, arXiv:1412.6980.
crossbar array, Nat. Nanotechnol. 19, 21 (2023). [56] R. M. Neal, MCMC using Hamiltonian dynamics, in Handbook
[44] E. B. Foxman, P. L. McEuen, U. Meirav, N. S. Wingreen, Y. of Markov Chain Monte Carlo (Chapman and Hall/CRC, New
Meir, P. A. Belk, N. R. Belk, M. A. Kastner, and S. J. Wind, Ef- York, 2011).
fects of quantum levels on transport through a Coulomb island, [57] J. V. Dillon, I. Langmore, D. Tran, E. Brevdo, S. Vasudevan, D.
Phys. Rev. B 47, 10020 (1993). Moore, B. Patton, A. Alemi, M. Hoffman, and R. A. Saurous,
[45] L. P. Kouwenhoven, C. M. Marcus, P. L. McEuen, S. Tarucha, TensorFlow distributions, arXiv:1711.10604.
R. M. Westervelt, and N. S. Wingreen, Electron transport in [58] J. R. Johansson, P. D. Nation, and F. Nori, QuTiP: An
quantum dots, in Mesoscopic Electron Transport, edited by L. L. open-source Python framework for the dynamics of open
Sohn, L. P. Kouwenhoven, and G. Schön, NATO Science Series quantum systems, Comput. Phys. Commun. 183, 1760
E (Springer, New York, 1997), Vol. 345, pp. 105–214. (2012).
[46] J. K. Sowa, J. A. Mol, G. A. D. Briggs, and E. M. [59] J. Bradbury et al., JAX: composable transformations of
Gauger, Environment-assisted quantum transport through Python+NumPy programs (2018), http://github.com/google/jax
single-molecule junctions, Phys. Chem. Chem. Phys. 19, 29534 [60] J. A. Nelder and R. Mead, A simplex method for function
(2017). minimization, Comput. J. 7, 308 (1965).
[47] N. Ares, Machine learning as an enabler of qubit scalability, [61] J. Gorman, D. G. Hasko, and D. A. Williams, Charge-qubit
Nat. Rev. Mater. 6, 870 (2021). operation of an isolated double quantum dot, Phys. Rev. Lett.
[48] J. P. Zwolak and J. M. Taylor, Colloquium: Advances in au- 95, 090502 (2005).
tomation of quantum dot devices control, Rev. Mod. Phys. 95, [62] K. D. Petersson, J. R. Petta, H. Lu, and A. C. Gossard, Quantum
011006 (2023). coherence in a one-electron semiconductor charge qubit, Phys.
[49] E. Chatzikyriakou, J. Wang, L. Mazzella, A. Lacerda-Santos, Rev. Lett. 105, 246804 (2010).
M. C. da Silva Figueira, A. Trellakis, S. Birner, T. Grange, [63] J. C. Butcher, Numerical Methods for Ordinary Differential
C. Bäuerle, and X. Waintal, Unveiling the charge distribution Equations (John Wiley & Sons, Hoboken, NJ, 2016).
of a GaAs-based nanoelectronic device: A large experimental [64] D. L. Craig, N. Ares, and E. M. Gauger, Code for the paper
dataset approach, Phys. Rev. Res. 4, 043163 (2022). “Differentiable master equation solver for quantum device char-
[50] G. J. Percebois and D. Weinmann, Deep neural networks for acterisation” (v1.0.0), Zenodo (2024), https://doi.org/10.5281/
inverse problems in mesoscopic physics: Characterization of the zenodo.10783608.
disorder configuration from quantum transport properties, Phys. [65] G. D. Mahan, Many-Particle Physics (Springer Science & Busi-
Rev. B 104, 075422 (2021). ness Media, 2000).

043175-15