GIMLET: Generalizable and Interpretable Model Learning through Embedded Thermodynamics
Authors: Suguru Shiratori, Elham Kiyani, Khemraj Shukla, George Em Karniadakis
Organizations: Department of Mechanical Systems Engineering, Tokyo City University, 1-28-1 Tamazutsumi, Setagaya-ku, 158-8557 Tokyo, Japan · Division of Applied Mathematics, Brown University, 182 George Street, Providence, 02912, RI, USA
We develop a data-driven framework for discovering constitutive relations in models of fluid flow and scalar transport. Under the assumption that velocity and/or scalar fields are measured, our approach infers unknown closure terms in the governing equations as neural networks. The target to be discovered is the constitutive relations only, while the temporal derivative, convective transport terms, and pressure-gradient term in the governing equations are prescribed. The formulation is rooted in a variational principle from non-equilibrium thermodynamics, where the dynamics is defined by a free-energy functional and a dissipation functional. The unknown constitutive terms arise as functional derivatives of these functionals with respect to the state variables. To enable a flexible and structured model discovery, the free-energy and dissipation functionals are parameterized using neural networks, while their functional derivatives are obtained via automatic differentiation. This construction enforces thermodynamic consistency by design, guaranteeing monotonic decay of the total free energy and non-negative entropy production. The resulting method, termed GIMLET (Generalizable and Interpretable Model Learning through Embedded Thermodynamics), avoids reliance on a predefined library of candidate functions, unlike sparse regression or symbolic identification approaches. The learned models are generalizable in that functionals identified from one dataset can be transferred to distinct datasets governed by the same underlying equations. Moreover, the inferred free-energy and dissipation functions provide direct physical interpretability of the learned dynamics. The framework is demonstrated on several benchmark systems, including the viscous Burgers equation, the Kuramoto--Sivashinsky equation, and the incompressible Navier--Stokes equations for both Newtonian and non-Newtonian fluids.
Figures & tables
Fig. 1: Schematic architecture of GIMLET for data-driven discovery of constitutive models in flow and transport PDEs. The PINN predicts the solution fields ψλf from spatiotemporal inputs {x,t} , while FreeEnergyNet and DissipationNet parameterize the free-energy and dissipation functions. Automatic differentiation is used to evaluate functional derivatives, and training is performed by minimizing combined data and physics losses. The pressure field is assumed not to be measured, and the data loss is evaluated without pressure data. For the problem of fluid flow only, the scalar field ϕ and FreeEnergyNet are not used.
Equations
Dim.
Steady/Unsteady
Term discovered
g
θ
Diff. in data A/B
Burgers
1D
Unsteady
ηuxx
–
∘
Initial conditions
Kuramoto–Sivashinsky
1D
Unsteady
βϕxx+γϕxxxx
∘
–
Time regimes
Navier–Stokes
2D
Steady
∇2u/Re
–
∘
Flow types
Navier–Stokes
2D
Steady
∇⋅(η^(γ˙)D)/Re
–
∘
Reynolds numbers
Table 1: Cases for validation.
Fig. 2: Results of model discovery and its application for Burgers equation. Upper panels are for model discovery, and lower panels are for model application. (a,f): history of loss function, where the total value, Jp (GE), Jd (Data), and the test loss (Test) are separately plotted. (b,g): Spatio-temporal distribution of solution calculated by GIMLET . (c,h): Solutions by the spectral method. (d,i): Differences between predictions and data. (e,j): Selected snapshots.
Fig. 3: (a) Dissipation function θ(ux) , where the green solid line indicates the function discovered by the GIMLET and the orange dashed line the true function. (b,c) The histograms of ux in the datasets used for discovery and application, respectively. (d) Relation between trained θ and \nicefracη2(ux)2 , where the blue circles are values in the dataset A.
Fig. 4: Results of model training for Kuramoto–Sivashinsky equation. Upper panels are for model discovery, and lower panels are for model application. (a,f): History of loss function, where the total value, Jp (GE), Jd (Data), and the test loss (Test) are separately plotted. (b,g): Spatio-temporal distribution of solution calculated by GIMLET . (c,h): Solutions by the spectral method, where the time regimes shown in half-transparent are not used for training. (d,i): Differences between predictions and data. (e,j): Selected snapshots.
Fig. 5: (a,d): Free energy functions g0(ϕ) and g1(ϕx) , where the green solid lines indicate the functions discovered by the GIMLET and the orange dashed lines the true functions. (b,e): The histograms of ϕ and ϕx in the dataset A used for discovery. (c,f): The histograms in the dataset B for the generalization check. (g): Relation between trained g0 and −β/2(ϕ)2 , where the blue circles are values in the dataset A. (h): Relation between g1 and γ/2(ϕx)2 .
Fig. 6: Results of model training for dissipation function of Newtonian fluids. The upper panels are for model discovery, and the lower panels are for model application. (a,e): History of loss function, where the total value, Jp (GE), Jd (Data), test loss (Test), ℓ2 loss (L2), and loss for pressure reference (pRef) are separately plotted. (b,f): Flow field calculated by GIMLET . Streamlines (black solid lines) and velocity magnitude (color contour). (c): Reference flow field calculated by OpenFOAM . (g): Reference flow field calculated by Nektar++ . (d,h): Differences of velocity magnitude between GIMLET and reference data.
Fig. 7: (a) Comparison between true and trained dissipation functions plotted for the dataset A. (b) Dissipation function projected to one dimension (upper) and data distributions corresponding to the projected axes (middle and lower).
Fig. 8: History of coefficient for dissipation function ΓRe .
Fig. 9: (a) Architecture for multiple rotated datasets. (b) Trained dissipation functions projected to one dimension (upper), and data distributions corresponding to the projected axes (lower).
Fig. 10: Results of model training for dissipation function of non-Newtonian fluids. The upper panels are for model discovery, and the lower panels are for model application. (a,e): History of loss function, where the total value, Jp (GE), Jd (Data), test loss (Test), ℓ2 loss (L2), and loss for pressure reference (pRef) are separately plotted. (b,f): Flow field calculated by GIMLET . Streamlines (black solid lines) and velocity magnitude (color contour). (c,g): Reference flow field calculated by OpenFOAM. (d,h): Differences of velocity magnitude between GIMLET and reference data.
Fig. 11: (a) Comparison between true and trained dissipation functions plotted for the dataset A. (b) Dissipation function projected to one dimension (upper) and data distributions corresponding to the projected axes (middle and lower).
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Equations
Sec.
βℓ2
βg
βθ
βh
Nf
Nd
Adam
SSBroyden
Burgers
3.2
1\times10−11
0.0
0.0
0.0
5000
5000
1000
10000
Kuramoto–Sivashinsky
3.3
1\times10−11
1.0
0.0
0.0
30000
5000
1000
200000
Navier–Stokes (Newtonian)
3.4
1\times10−11
0.0
0.0
1.0
20000
10000
1000
100000
Navier–Stokes (non-Newtonian)
3.5
1\times10−11
0.0
0.0
1.0
30000
30000
1000
100000
Appendix
Table 2: The hyperparameters used in the validation cases.
Nh
10
20
30
Nℓ
4
1.095\times10−07
7.036\times10−11
3.803\times10−11
6
5.850\times10−10
3.103\times10−11
2.897\times10−11
8
1.913\times10−10
2.743\times10−11
2.912\times10−11
Appendix
Table 3: Final loss function for different network sizes of the PINN.
Nh
10
20
30
Nℓ
4
0.35617%
0.35673%
0.35674%
6
0.35667%
0.35675%
0.35673%
8
0.35673%
0.35674%
0.35674%
Appendix
Table 4: Root mean squared relative errors (RMSRE) in dissipation function θ for different network sizes of the PINN.
Nh
5
10
15
Nℓ
2
3.056\times10−11
3.022\times10−11
3.337\times10−11
4
2.884\times10−11
2.743\times10−11
2.764\times10−11
6
2.755\times10−11
2.866\times10−11
2.734\times10−11
Appendix
Table 5: Final loss function for different network sizes.
Nh
5
10
15
Nℓ
2
0.35675%
0.35674%
0.35674%
4
0.35674%
0.35674%
0.35672%
6
0.35674%
0.35674%
0.35673%
Appendix
Table 6: Root mean squared relative errors (RMSRE) in dissipation function θ for different network sizes.
GIMLET
PINN-SR
RMSRE
S/F (# mis-spec.)
ϵpred
ϵmis
Dataset A
noise 0%
0.356%
Fail (6)
0.01%
0.013%
noise 5%
0.366%
Fail (13)
19.9%
348%
noise 10%
0.368%
Fail (14)
29.6%
809%
Dataset B
noise 0%
0.420%
Success
0.001%
—
noise 5%
0.432%
Fail (2)
0.447%
7.42%
Appendix
Table 7: Summary of model discovery results in comparison with PINN-SR [ 10 ] for the case of Burgers equation. For the GIMLET , the accuracy is evaluated by root mean squared relative errors (RMSRE) between discovered and exact dissipation function θ . For the PINN-SR, the number of misspecified terms (# mis-spec.) is shown for failure cases. The accuracy is evaluated as relative error ϵpred in the coefficient of the diffusion term uxx . For failure cases, the root mean squares of normalized coefficients ϵmis are also evaluated. The GIMLET was also run for the noise level of 1% , and the discovered dissipation function is included in Fig. 13 ; the PINN-SR was not run for that noise level, and it is therefore not listed in this comparison.
Fig. 12: History of loss functions for the discovery of diffusion term in Burgers equation with noise of 5% . Results for Dataset A (a) and Dataset B (b).
Fig. 13: Dissipation functions θ(ux) discovered by the GIMLET for different noise levels. Results for Dataset A (a) and Dataset B (b). Gray solid lines are ground truth, whereas dashed color lines are θ(ux) discovered by the GIMLET .
We introduce Nonlinear GENERIC Informed Neural Networks (N-GINNs), a deep learning framework for discovering evolution equations of systems governed by the nonlinear GENERIC formalism (General Equation for Non-Equilibrium Reversible-Irreversible Coupling). Such systems exhibit coupled conservative and dissipative dynamics, and can be described via the superposition of a Hamiltonian flow and a generalized gradient flow. In contrast to existing approaches, our formulation incorporates generalized gradient flows via convex dissipation potentials, enabling the identification of a broader class of thermodynamically consistent dynamics, including systems with non-quadratic dissipation potentials. Thermodynamic structure is strongly enforced by construction through suitable reparameterizations of both the bivector operator and the dissipation potential, ensuring exact compliance with the first and second laws of thermodynamics. We validate the proposed approach on three representative examples: a harmonic oscillator coupled to a heat bath, an idealized chemical motor, and a one-dimensional viscoplastic model of Perzyna type. These results demonstrate the method's ability to accurately infer thermodynamically consistent models from data for systems incorporating both conservative and nonlinear dissipative dynamics.
Vojtěch Votruba, Zequn He, Weilun Qiu +2
Mathematical Institute, Faculty of Mathematics and Physics, Charles University, Prague, 18675, Czech Republic · Department of Mechanical Engineering and Applied Mechanics, University of Pennsylvania, Philadelphia, 19104, PA, United States of America
We present a physics-based neural network framework for the discovery of constitutive models in fully coupled thermomechanics. In contrast to classical formulations based on the Helmholtz energy, we adopt the internal energy and a dissipation potential as primary constitutive functions, expressed in terms of deformation and entropy. This choice avoids the need to enforce mixed convexity--concavity conditions and facilitates a consistent incorporation of thermodynamic principles. In this contribution, we focus on materials without preferred directions or internal variables. While the formulation is posed in terms of entropy, the temperature is treated as the independent observable, and the entropy is inferred internally through the constitutive relation, enabling thermodynamically consistent modeling without requiring entropy data. Thermodynamic admissibility of the networks is guaranteed by construction. The internal energy and dissipation potential are represented by input convex neural networks, ensuring convexity and compliance with the second law. Objectivity, material symmetry, and normalization are embedded directly into the architecture through invariant-based representations and zero-anchored formulations. We demonstrate the performance of the proposed framework on synthetic and experimental datasets, including purely thermal problems and fully coupled thermomechanical responses of soft tissues and filled rubbers. The results show that the learned models accurately capture the underlying constitutive behavior. All code, data, and trained models are made publicly available via https://doi.org/10.5281/zenodo.19248596.
Hagen Holthusen, Paul Steinmann, Ellen Kuhl
Institute of Applied Mechanics, University of Erlangen-Nuremberg, Egerlandstraße 5, 91058 Erlangen, Germany · Department of Mechanical Engineering, Stanford University, United States
Constitutive laws for inelastic materials must satisfy strict thermodynamic admissibility requirements, yet current data-driven approaches sacrifice interpretability, even when formal guarantees are provided by physics-encoded architectures. We propose a symbolic regression framework for the data-driven discovery of dissipation potentials governing the evolution of internal variables within the Generalized Standard Materials (GSM) formalism. Starting from the Clausius--Duhem inequality, we enforce the thermodynamic requirements, convexity and non-negativity, that the dual dissipation potential must satisfy to guarantee non-negative mechanical dissipation. These requirements are formulated in the general subdifferential setting, encompassing rate-dependent (viscoelastic) and viscoplastic dissipative mechanisms, including potentials with genuine elastic domains, within a unified framework. Candidate potentials are generated by a composition-extended convexity-preserving grammar that guarantees thermodynamic admissibility by construction. The framework is validated on synthetic datasets spanning Newtonian, power-law, and Bingham viscoplastic ground truths under process and measurement noise, and on experimental oscillatory shear measurements of a synthetic elastomer across multiple strain amplitudes and frequencies, where the discovered potentials reproduce the amplitude-dependent softening of the dynamic moduli and outperform a calibrated linear Zener baseline.