- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents PHySIM, a method to combine neural network models with physical simulators for networked systems. The main idea is to replace some parts of a physical system by neural networks, while keeping the global system solved with a Newton-Raphson method. The neural networks are evaluated inside the solver, and their Jacobians are used to build the global Jacobian. The authors also propose to train the networks with a Jacobian-related loss, in order to improve the numerical behaviour of the solver.
The method is evaluated on several applications, including power systems, analog circuits and gas networks. The experiments show that PHySIM can reduce simulation time, especially when large parts of the system are replaced by neural models. The paper also studies the effect of the training loss, the choice of subsystem partitioning, and the coupling with a controller in an inverter example.
Critical strengths
The paper has a strong engineering motivation: network simulators often contain a small number of subsystem classes that are difficult to model from first principles, and the authors propose to replace them with neural surrogates. Besides, the applications are really broad: power systems, gas pipeline networks, analog circuits, which shows that the methodology is quite general.
The paper complements its empirical results with theoretical analysis on convergence, Jacobian accuracy, and partitioning, which strengthens the overall contribution.
Critical weaknesses
-- My main concern is about the novelty of the approach. Combining learned DNN into Newton or other solvers is not new, and the related work and positioning is insufficient. For example, neural constitutive laws can be injected into solvers: Vlassis & Sun (2021) :Sobolev training of thermodynamic-informed neural networks for interpretable elasto-plasticity models with level set hardening Xu, Huang & Darve (2021) : Learning constitutive relations using symmetric positive definite neural networks
-- The set of baselines is weak, they are global metamodels or simulators, while PhySIM specifically exploits modular decomposition. Many important and recent hybrid physics-ML baselines are missing, for example:
- Deep Equilibrium models or implicit layers with physics residuals
- Other augmented models with Neural ODEs like APHYNITY (Yin et al, 2021)
- Graph-based simulators designed for topology transfer, like MeshGraphNets (Pfaff et al , 2021)
-- “Exact conservation-law satisfaction” is a claim overstated: KCL is satisfied for the hybrid system, This is not conservation for the true physical system.
-- A claim is the integration of neural surrogate Jacobians into the global Newton solver. An ablation study is missing. Could the authors provide an ablation comparing: • Full PHySIM, with neural Jacobians computed by automatic differentiation • The same neural surrogates, but without their Jacobian contribution in the global Newton solve. • Finite-difference approximations of the surrogate Jacobians ?
Questions for authors
Most experiments appear to rely on synthetic or simulator-generated data. What to expect and what challenges on real-world data?
4 comments
Response to Reviewer Rodo Pt 2
Weak baselines: DEQs, APHYNITY, MeshGraphNets
We introduced three additional comparison baselines to evaluate PHySIM against learned graph models, implicit fixed-point models, and physics-constrained learning approaches: MeshGraphNets, a DEQ-style model, a physics-informed variant of it that adds the governing-equation residual to the training loss, and a physics-constrained model built on equality completion. We also clarify the distinction from APHYNITY, which is designed for forecasting ODE dynamics using additive physics-based corrections rather than solving coupled nonlinear network equations with learned device macromodels. Its formulation is therefore less directly applicable to our problem, and we discuss it in related work.
MeshGraphNets. We trained a MeshGraphNets-style model with explicit edge latents and ten processor steps on circuits with two and four transistor stages, then evaluated zero-shot transfer to circuits with 8, 16 and 32 stages (Appendix F.4). Although the model transfers to unseen sizes without retraining, its voltage error is 0.18–0.22 V, compared with 2–2.4×10⁻⁴ V for PHySIM on the same netlists. This highlights the difficulty of maintaining numerical accuracy when generalizing a learned graph representation across substantially different network sizes.
A shared setup for the remaining three. The other comparisons share one protocol (Appendix F.5), in which twelve of the thirteen governing equations are supplied to every method and the thirteenth — the constitutive law of one load — is withheld from all of them, replicating the case where a device's behavior has no first-principles model, as in load forecasting. Ground truth is an independent closed-form device law rather than PHySIM's own converged states. We report the residual as a dimensionless relative violation, separated into the five mass-balance equations and the eight constitutive relations, because the raw residual is in pressure-squared units and its magnitude reflects the units rather than the error.
DEQ. The DEQ-style network reaches 0.246% mean relative state error — the most accurate in-distribution model after PHySIM — yet leaves a mean mass imbalance of 3.2×10⁻³ at the network's nodes, worst case 6.2×10⁻³, against 1.8×10⁻¹² for PHySIM. Converging to a learned fixed point does not conserve mass.
Physics-informed variant. We implemented the variant the reviewer suggested, adding the governing-equation residual to the training objective. It reduces the mass imbalance to 1.9×10⁻³ and no further, while state accuracy is unchanged. A soft penalty moves feasibility slightly and cannot close it, because small pressure errors are strongly amplified in the governing residual and the penalty must be balanced against state-fitting error.
Equality completion (DC3-style). This is the strongest of the baselines: it predicts a device flow from the global operating conditions and solves the remaining equations directly, satisfying them by construction. What distinguishes PHySIM from it is the constitutive error under a change of network. PHySIM is flat across all three networks tested (2.3×10⁻⁴ to 4.4×10⁻⁴), while completion's grows 300×, from 3.1×10⁻⁴ to 9.8×10⁻², as the pipe geometry changes; in state error that is 0.110% rising to 16.29%, against 0.082% and 0.058% for PHySIM, and on the most distant network completion fails to converge on 22% of conditions.
Anticipating the natural objection that the network should simply be given the topology as an input, we trained that variant across 16 networks using 9,600 converged solves against our 600, and swept it over seven configurations from 8k to 524k parameters. Its training loss falls to 7×10⁻⁶ while test error rises, and it never reaches below 3.9% state error on any network — the failure is generalization in network-parameter space, not optimization.
Predicting a device flow from global operating conditions ties the learned mapping to the network it was trained on. PHySIM instead learns the device law as a function of its local boundary state, which is determined jointly with the rest of the network during the coupled solve; because that boundary state is not known before solving, the device law must participate in the solve rather than be evaluated in advance.
Taken together, these comparisons highlight the central distinction of PHySIM: it separates the known network conservation structure from the unknown device constitutive laws. The conservation equations are imposed directly during inference, while DNNs provide the device responses within the coupled Newton solve. This lets the method impose the coupled system's conservation laws without requiring exact device models, while supporting generalization across network configurations.
Response to Reviewer Rodo Pt 3
Ablation of the Jacobian contribution (full AD, no Jacobian, finite difference)
We thank the reviewer for their comment regarding providing the ablation. To demonstrate the advantage of computing the Jacobian entries of the DNN models in PHySIM, we compare it against different methods including finite difference methods (with different step sizes) and dropping the Jacobian entries (similar to Gauss-Seidel style update). We applied this is in circuit domain (BSIM-NN, 50 gate voltages) as described in Appendix F.1. In this experiment, the Newton step uses the DNN Jacobian in one of four ways:
| DNN Jacobian in the Newton step | Converged | Mean iterations |
|---|---|---|
| Full (AD), PHySIM | 100% | 4.04 |
| Finite difference, step 10⁻⁴ V | 100% | 4.10 |
| Finite difference, step 0.3 V | 60% | 9.83 |
| Finite difference, step 1.0 V | 12% | 26.17 |
| Dropped | 38.0% | 10.63 |
Dropping the DNN Jacobian clearly hurts convergence. Finite difference is more subtle: at a small step it reproduces the backpropagated Jacobian that PHySIM uses and matches it to within noise (4.10 vs 4.04 iterations), but as the step grows the difference quotient stops approximating the local slope and the Newton iteration diverges — 60% convergence at a 0.3 V step and 12% at 1.0 V. A finite-difference Jacobian can therefore fail outright through its own approximation error, and whether it does depends on a step size that is not known in advance. The full step sweep is in Appendix F.1, and the paper's finite-difference rows now state the step used.
This matters for the adaptive step-size schemes that simulators traditionally rely on. Such a scheme varies the perturbation during the solve, so it can drift into a range where the difference quotient no longer represents the local slope, and the Jacobian degrades without any error being raised. The automatic-differentiation Jacobian that PHySIM assembles has no step size to select and is exact to machine precision at every iterate, which removes this failure mode.
Question: real-world data
All experiments use simulator-generated data; we have no real-data experiment and say so in the new Limitations paragraph.
We expect the approach to carry over, for two reasons. PHySIM lets a DNN learn device behavior that first-principles models cannot capture, as the exogenous-feature experiments illustrate. And laws that any real network must satisfy, such as KCL or mass balance, remain enforced exactly by the physics-based equations rather than being learned from noisy measurements. This is an expectation, not a demonstrated result.
Dear Reviewer ,
Thank you again for your thoughtful review. As the discussion period closes soon, we wanted to check whether our rebuttal has addressed your concerns. We'd be happy to clarify anything further while there's still time.
If our responses have resolved your main points, we'd be grateful if you would consider updating your
Response to Reviewer Rodo Pt 1
We thank the reviewer for their comments. We have revised the conservation-law claim throughout the manuscript and have run the requested Jacobian ablations, added the suggested baselines, and sharpened the related-work discussion to say why the cited constitutive-law works are not comparable baselines. The specific comments and questions are addressed below:
Novelty relative to Vlassis & Sun (2021) and Xu, Huang & Darve (2021)
We thank the reviewer for these references, which we have added to the related work. Both are methods for training a neural constitutive model better: a Sobolev loss on the derivatives, and a Cholesky parameterization that makes the tangent stiffness SPD. We do not claim either idea, and we explicitly do not claim Sobolev training as a contribution.
But these still remain surrogate models whose fidelity is a property of training, and at inference cannot guarantee the learned model respect a constraint it was not trained to respect. That is the same category as the PINN baseline we already compare against, where the physics enters through the loss and the constraints hold only approximately once the operating point moves.
Our contribution is not a training method but a construction for inference. PHySIM demonstrates how DNNs can be implicitly coupled to physics-based equations inside a single Newton solve, with its Jacobian assembled as a block of the global Jacobian. Because the conservation laws are carried by the physics equations and not by the network, they hold exactly at every converged solution independent of DNN approximation error. Additionally, the coupling between physics based equations and DNNs is what gives reuse of one DNN macromodel at every instance of a device type, zero-shot transfer to unseen topologies, and reduction of internal states and runtime.
The manuscript demonstrates these advantages against similar prior works. Figure 2 in Section 4 illustrates that against the full set of trained surrogates (FC-DNN, PINN, GC-NN, FNO) under out-of-distribution operating conditions, PHySIM holds both constraint residual and state error near machine precision while the baselines degrade. In the latest revision we added a DEQ-style model, which reaches 0.246% mean state error yet leaves a mean mass imbalance of 3.2×10⁻³ at the network's nodes against 1.8×10⁻¹² for PHySIM (Appendix F.5), and an equality-completion model in the style of DC3. The MeshGraphNets-style model transfers to a new topology without retraining but at 0.18–0.22 V error against our 2–2.4×10⁻⁴ V (Appendix F.4)
The benefit and scope of this work is for network-based systems which have subsystem boundaries at which different models meet unlike continuous flows like proposed in Vlassis & Sun. For that reason we did not include it as a baseline.
"Exact conservation-law satisfaction" is overstated
We have revised the wording in the abstract, introduction, and conclusion: they now describe the coupled system’s conservation laws as imposed at inference time rather than learned, and the introduction states explicitly that the DNN’s error enters through the subsystem’s boundary response, not through the conservation structure.
The central contribution is the ability to enforce conservation laws directly during inference, independently of the accuracy of the learned device models. Approximation errors in a DNN can affect the predicted constitutive response of an individual device, but cannot introduce artificial sources or sinks of mass or charge at network nodes.
This is the same guarantee provided by conventional physics-based simulators. For example, SPICE with a BSIM compact model exactly conserves charge among the modeled device currents, even though BSIM itself approximates the behavior of a physical transistor. The same principle applies to power-flow and pipeline solvers using approximate device models. Our contribution is to extend this established simulator-level guarantee to systems incorporating learned DNN macromodels, imposing the coupled system’s conservation laws during inference without requiring the learned device models themselves to be exact.