acceptodds
Under review as a conference paper at ICLR 2027

PHySIM: Neural-Augmented Physical Simulation of Networked Systems

Abstract

Network-based systems, including power grids, circuits, and gas pipelines, consist of interconnected subsystems governed by coupled equations. Simulating these systems requires solving algebraic and differential subsystem equations simultaneously while enforcing their shared physical boundary variables. Physics-based solvers exploit the subsystem structure but are slow to converge when subsystems have severe nonlinearities or behaviors driven by exogenous factors (weather, occupancy) that first-principles models cannot capture. Black-box approaches offer low-cost approximations that capture hidden behavior but cannot enforce conservation laws at inference time. These methods also need training data covering the network's full state space and struggle to generalize across topology changes. To address these limitations, we introduce PHySIM, a flexible framework for modeling network-based systems through subsystem partitioning, where each subsystem can be represented by either a physics-based model or a DNN coupled at inference time through shared state variables at subsystem boundaries. PHySIM implicitly captures interactions across coupled subsystems while combining the predictive capabilities of DNNs to capture hidden behaviors seamlessly with the explainability and accuracy of physics-based models. The coupled system is solved inside the Newton-Raphson solver loop that extracts the DNN's sensitivities via backpropagation and assembles them into the same Jacobian matrix as physics-based terms, thereby solving the coupled system simultaneously rather than sequentially. The inference-time coupling inherits the benefits of both physics-based and DNN models: exact conservation-law satisfaction; superlinear state-space scaling from DNN reuse; data requirements that scale with device complexity rather than network size; zero-shot topology generalization without retraining; and provable Newton stability linked directly to a Sobolev Jacobian training loss. We validate PHySIM across power systems, analog circuits, and gas pipeline networks, and demonstrate a capstone electrothermal grid simulation that correctly identifies voltage-collapse onset where physcs-based and black-box simulation baselines fail.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physics-based simulation, hybrid machine learning, DNN macromodeling, gray-box modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RodoOpenReview ↗
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
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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 stepConvergedMean iterations
Full (AD), PHySIM100%4.04
Finite difference, step 10⁻⁴ V100%4.10
Finite difference, step 0.3 V60%9.83
Finite difference, step 1.0 V12%26.17
Dropped38.0%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

Reviewer_1VkyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This work presents PHySIM, a framework for modeling network-based systems. PHySIM combines predictive DNNs and physics-based models, resulting in a hybrid framework that absorbs advantages from both sides.

Critical strengths
  1. Experimental results show that PHySIM is a good balance between physics-based and DNN approaches, benefiting from both sides.
Critical weaknesses
  1. The research gaps in introduction should be further justified. The three targeted structural problems are only associated with DNNs. In addition, the first claimed gap (no conversation mechanism in DNN) is too broad: for example, projection (simple as it may be) is widely used in enforcing hard constraints during inference; the third claimed gap (structural change needs retraining) is not always true, especially for GNNs -- Dynamic GNNs, even an ordinary message-passing GNN can avoid having to retrain after every topology change.
  2. It is not clear whether the baselines are properly trained using same information as PHySIM.
  3. There is no clear illustrative overview of PHySIM's architecture and key components.
Questions for authors

Please see the critical weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1Vky Pt 1

We thank the reviewer for their comments on the the research gaps and the fairness of the comparisons. We have rewritten parts of the manuscript clarifying the baselines and limiting the scope of PHySIM. Additionally, we added an explicit statement of training information, added an architecture overview figure, and tested the topology-transfer claim with a MeshGraphNets-style baseline. Specific reviewer comments and questions are addressed below:

Research gaps are only about DNNs; projection enforces hard constraints; GNNs can transfer across topologies without retraining

The introduction now acknowledges hard-constraint and projection layers (e.g., DC3), which can enforce a known constraint set but still require a model of the whole network's behavior. It also acknowledges that message-passing GNNs can be applied to a new graph without retraining.

"The gaps are only about DNNs." The methodology is not tied to the DNN architecture. PHySIM is a way to structurally integrate any ML macromodel with physics-based equations inside one solver, and the macromodel's architecture can be changed without changing the coupling. Each localized macromodel can also be trained with a PINN-style loss on its own device equations. We have not run that variant, so this is a statement about what the framework allows, not a result.

"Projection enforces hard constraints." Projection is not a cheap or simple fix, for two reasons.

  • Methods that guarantee the equality constraints at inference (exact projection, or equality completion as in DC3) must solve the nonlinear network equations, which is the physics solve with its convergence difficulties. Exact projection onto nonlinear equality constraints is a nonconvex problem whose KKT system contains those same equations. Cheaper alternatives (penalty terms, linearized one-step projection) satisfy the constraints only approximately.
  • A projection operator must know all of the network's physics-based equations, including those of the complex devices. It therefore inherits the limitations that motivate PHySIM: (1) nonlinear, high-dimensional device models with many internal states degrade Newton convergence, whereas a macromodel absorbs those states (the state-space reduction grows from 3.9% to 38.4% as the power network scales, and replacing a subtree of 200 coupled devices with a macromodel gives a 50.9× speedup; Section 4 and the additional-experiments appendix); and (2) behavior driven by exogenous factors such as weather or time of day has no first-principles equation to project onto, whereas a DNN macromodel can learn it (exogenous-feature experiments).
  • Projection also still requires a model of the whole network's behavior to project, so it does not remove the data and retraining costs listed in the introduction.

"GNNs transfer across topologies without retraining." Dynamic or message-passing GNNs are trained to minimize a loss. They therefore do not guarantee that physical constraints such as KCL are met, even when their error on a new topology is small. For engineering use, a constraint violation is not acceptable even when the average error is low. PHySIM satisfies the coupled system's conservation laws by construction at every converged solution. GNNs also learn from whole-network examples, so covering a variety of topologies needs far more training data than PHySIM, which trains only on local device data. We did not measure GNN data requirements separately; the data-scaling argument follows the training setups described in Section 4 and Appendix D.7. The test below checks accuracy, not constraint satisfaction, and in it the GNN's accuracy also degrades.

We then tested the GNN directly (Appendix F.4). We built a MeshGraphNets-style model (explicit edge latents, 10 processor steps), trained it on circuits with 2 and 4 transistor stages, and evaluated it zero-shot on 8, 16 and 32. While MeshGraphNets fit its training sizes well (mean absolute error about 0.02 V), the zero-shot error to changes in topology is 0.18–0.22 V, where as PHySIM achieves an error ranging 2–2.4×10⁻⁴ V for the same changes in the netlists.
So a GNN can run on a new graph without retraining, but in our test its accuracy degrades during topology transfer.

AuthorsOpenReview ↗

Response to Reviewer 1Vky Pt 2

It is not clear whether baselines are trained with the same information as PHySIM

We have clarified this in the manuscript (Appendix D.7, with a pointer in Section 4); the details are below.

  • The black-box baselines (FC-DNN, PINN, GC-NN, FNO) are trained on full-system data: operating conditions mapped to the complete network state. PHySIM is trained only on local device-boundary data and never sees full-system state data. If anything, the baselines have richer per-sample information.
  • PHySIM is additionally given the network's governing equations at inference. This is the modeling assumption of the method, not extra training data. On the other hand, the baseline PINN methods use the equations as a soft penalty, and the GC-NN is given the graph.
  • The MeshGraphNets comparison is matched by construction (same ground-truth generator, split and operating-point distributions).

No clear overview of the architecture

We have added an overview figure (Appendix D.1, referenced at the start of Section 3). The figure shows how we partition the network into physics and device subsystems. This hybrid model with one DNN per device type is solved via Newton–Raphson, and PHySIM's proposed method to compute each Newton step is illustrated under each panel.

We placed it in the appendix to keep the main text within the page limit, but we agree it is central to understanding the method. If the reviewer would prefer, we are happy to move it into the main text.

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

Thank you for your rebuttal. My main concerns are addressed and I am raising my score.

Reviewer_q5zpOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PHySIM, a hybrid framework for simulating networked physical systems by integrating neural subsystem models into a Newton–Raphson solver. The framework partitions a network into physics-based and neural components, with neural networks approximating local subsystem behavior and their Jacobians incorporated into the global Newton system. The authors further introduce Jacobian-constrained training and a subsystem partitioning strategy. Experiments cover power systems, analog circuits, and gas pipeline networks.

Critical strengths
  • The integration of neural subsystem models into a unified physics-based solver is practically interesting, particularly for complex nonlinear systems with partially unknown dynamics.

  • The framework preserves explicit network coupling constraints while allowing trained device-level models to be reused across different network configurations.

  • Experiments cover multiple application domains, and the supplementary material provides additional investigations of subsystem partitioning, numerical stability, and runtime efficiency.

Critical weaknesses

Clarification on Theoretical Stability and Optimality Guarantees

  • Newton Convergence: Jacobian nonsingularity (Appendix B.2) ensures matrix invertibility but does not guarantee convergence of the nonlinear Newton iterations.
  • Overly Strong Assumptions: Generalizing Jacobian error bounds (Appendix B.3) requires stronger regularity conditions than the stated Lipschitz continuity.
  • Evaluation Cost Assumption: Partition optimality assumes constant neural evaluation costs, yet distinct device instances still require separate forward passes at their respective boundary states.

Inconsistent Runtime Advantages and Profiling

  • Offset Speedups on Small Systems: State-space reduction benefits are largely neutralized by neural model evaluations and Jacobian computations on 14-bus and 118-bus systems (Appendix E.3).
  • Missing Overhead Profiling: A systematic analysis of end-to-end runtime, Jacobian computation overhead, peak memory, and Newton iteration counts is needed to clearly delineate when the method outperforms conventional numerical solvers.
Questions for authors
  • Can PHySIM be applied to synchronization networks of coupled chaotic systems, such as Lorenz oscillators? How would the framework handle sensitive dependence on initial conditions, long-term error accumulation, and synchronization stability?

  • How does the proposed monolithic Newton solver compare with fully converged alternating or operator-splitting approaches, such as block Gauss–Seidel and waveform relaxation, under identical subsystem models and convergence criteria?

  • Does the stability analysis guarantee convergence of the nonlinear iterations, or only nonsingularity of the hybrid Jacobian? What additional conditions are required for convergence?

  • How does the total computational cost scale with the number of neural subsystem instances, including repeated neural evaluations and Jacobian assembly?

4 comments
AuthorsOpenReview ↗

Response to Reviewer q5zp Pt 1

We thank the reviewer for a close reading of the theory and comments on the runtime experiments. In summary, we have narrowed the theoretical claims and have added end-to-end runtime comparisons as well as a per-iteration profile. We address each of the reviewer’s point below.

Newton convergence: Jacobian nonsingularity does not guarantee convergence of the nonlinear iterations (also Question 3)

Theorem 1 shows only that the hybrid Jacobian stays nonsingular, so each Newton step is well defined. It does not prove convergence of the nonlinear iteration, which also depends on the initial guess, damping and the nonlinearity of the physics.

We have therefore narrowed the claim. Section 3 now states the nonsingularity-only guarantee, and a new scope remark after Theorem 1 says that convergence of the nonlinear iteration is not proved.

We have also strengthened the theorem, whose guarantee now rests on an explicit, checkable condition, ε_p + ε_q + (L_H + L_J)δ < 1/μ, with the assumptions stated (see the next response).

Finally, we added empirical evidence on convergence. Appendix F.1 shows that by extracting the Jacobian values via backpropagation of localized DNN models, Newton converged on 100% of 315 gas cases (6.80 iterations on average) and 100% of 50 circuit cases (4.04), against 86.7% (36.40) and 38.0% (10.63) when the Jacobian term is not included. Appendix E.3 shows Sobolev training raising Newton convergence from 73.6±6.7% to 100% on a small device model.

Overly strong assumptions: the Jacobian error bounds need more than Lipschitz continuity

We thank the reviewer for the catch. The neighborhood bound in B.3 relates the DNN's Jacobian at an unsampled point to its Jacobian at a nearby training point. That needs the Jacobian to be Lipschitz continuous (g_nn in C^{1,1}), which Assumption 4 did not state.

Assumption 4 now requires a bounded and Lipschitz Jacobian (constants L, L_H). We give explicit bounds for MLPs with smooth activations and bounded weights, and note that the guarantee does not apply to ReLU networks. The macromodels in the paper use tanh, sigmoid or softplus.

We also fixed a second gap. The old Lemma 1 applied an average training loss to arbitrary Newton iterates. The error ε_p is now the worst-case Jacobian residual on the training set (measurable after training), the lemma holds at training points, and the unobserved-variable term ε_q is defined explicitly.

Theorem 1 now rests on a checkable condition, ε_p + ε_q + (L_H + L_J)δ < 1/μ, while the iterates remain in a δ-covered region.

Evaluation cost: partition optimality assumes constant neural evaluation cost, but each instance needs its own forward pass

The cost of evaluating the neural network is indeed not exactly consistent across the experiments. Our runtimes now include the cost of evaluating and performing the backpropagation as well as the time spent in each NR iteration (Appendix F.3). We found that multiple instances of one device type can be evaluated as a batch to evaluate the right hand side vector and Jacobian values at each NR iteration. We verified agreement with per-instance evaluation to about 1e-7 relative error.

We found that batched evaluation is 3.2× faster than serially evaluating each model. Inside a full Newton solve on N uncoupled gas-network blocks, the whole-solve speedup grows from 2.4× to a plateau of about 27× (N≥100). When the number of instances increases to N=1000, the DNN evaluation and Jacobian extraction time falls from 25.4 s to 0.035 s, while physics stamping and the sparse linear solve are unchanged.

Q1: Lorenz / chaotic synchronization networks

The scope of PHySIM is network-based systems whose governing equations split into physics-based subsystems and subsystems that a DNN macromodel replaces, coupled through shared boundary variables and solved with a Newton method. We evaluate it on steady-state problems (circuit operating points, power flow, gas flow) and on implicitly time-stepped transients, where each time step is a nonlinear solve of the same form (Appendix D).

Lorenz and chaotic synchronization networks are outside this scope. Their defining property is sensitive dependence on initial conditions, so a small macromodel error grows exponentially over a long horizon and a pointwise trajectory error is not a meaningful measure. Assessing them would need statistical criteria such as attractor geometry and Lyapunov exponents, and an analysis of how macromodel error propagates over time, which our per-step theory does not provide. Synchronization stability is a further question our analysis does not reach: whether coupled oscillators lock depends on the Lyapunov exponents transverse to the synchronization manifold, which is a property of the coupled trajectory over time and not of any single Newton step.

AuthorsOpenReview ↗

Response to Reviewer q5zp Pt 2

Offset speedups on small systems; missing overhead profiling

This was the most concrete gap, and we have added both the end-to-end measurement and the crossover.

Per-iteration overhead. Appendix F.2 profiles one gas-network instance into five phases. DNN evaluation (forward plus Jacobian) takes 57.0% of per-iteration compute, and the backpropagation call that produces the Jacobian alone takes 49.6%. This is the overhead a macromodel must offset, and it is why aggregation does not pay on the smallest networks.

End-to-end runtime, circuit. The same netlist and the same 50-point gate-voltage sweep run with the physics BSIM-CMG compact device and with the BSIM-NN macromodel: 25.23 ms per operating point against 3.60 ms, a 7.0× end-to-end reduction, both converging to the same solution in the same 4.04 Newton iterations on average. The system dimension is unchanged here, so the whole gain is evaluation cost.

End-to-end runtime, power, scaled up. We also scaled the power example to a much larger network: an 800-bus feeder with 9,950 unknowns, where one aggregate macromodel replaces a downstream subtree. Complete-solve times, medians of five runs:

Buses replacedStates removedNewton itersSolve timeSpeedupMax bus-voltage error
— (full physics)0%215.71 s1.00×—
526.5%215.31 s1.08×0.0001%
10313.4%213.90 s1.46×0.0009%
20324.7%212.66 s2.15×0.0008%
39748.7%31.39 s4.10×0.0005%
71690.2%20.16 s36.6×0.0008%

This is the same behaviour the gas sweep already showed (near parity at n=10 rising to 50.9× at n=200) and for the same reason: the gain comes from the states the macromodel removes, not from the presence of a DNN. Each macromodel is trained only on its own subtree's boundary response, and the maximum voltage error at every retained bus stays below 0.001% of nominal against the full physics solution. Both sweeps are now in the same table in the appendix.

Q2: Monolithic Newton vs. block Gauss–Seidel or waveform relaxation

The dropped-Jacobian variant in Appendix F.1 is a block Gauss–Seidel (Picard) iteration. The DNN output is held fixed within each linear solve of the physics equations and re-evaluated at the next iteration, so the DNN does not contribute to the Jacobian. We compare it with the PHySIM Newton step that couples DNNs and physics based equations in the Jacobian where we observe PHySIM provides robust convergence to changes in initial conditions as opposed to block Gauss-Seidel and, on average, converges in fewer iterations.

DomainSolverConvergedMean iterations
Gas (315 points)Monolithic Newton (PHySIM)100%6.80
Block Gauss–Seidel (dropped Jacobian)86.7%36.40
Circuit (50 points)Monolithic Newton (PHySIM)100%4.04
Block Gauss–Seidel (dropped Jacobian)38.0%10.63

The convergence requirements of block Gauss–Seidel are much stricter than those of Newton–Raphson. Gauss–Seidel is a fixed-point iteration, so it converges only if the coupling between the blocks is weak enough for the iteration map to be a contraction (the spectral radius of the block coupling must be below one), and even then it converges only linearly. Newton–Raphson needs only a nonsingular Jacobian near the solution and then converges quadratically, because the Jacobian carries the coupling between the DNN and the physics explicitly. This is what the results show: alternating between the DNN and the physics converges on fewer cases and, on the gas points where it does converge, needs 5.4× more iterations. The failing circuit points are concentrated where the device current is most nonlinear in its terminal voltages (v₁ ≥ 0.38 V).

Q4: Cost scaling with the number of neural instances

Because all instances of a device type are one shared function, the Jacobian assembly for N instances is a single batched call, and its cost stops growing with N. The consequence for total cost is that the DNN term flattens while physics stamping and the sparse linear solve do not, so those bound the whole-solve speedup at about 27× (N≥100) on this benchmark. We demonstrate this in Appendix F.3.

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

Thank you for your detailed response. It addressed most of my concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.