acceptodds
Under review as a conference paper at ICLR 2027

Multiphysics Post-Training: From Decoupled Modules to Coupled Equilibria

24 views

Abstract

Multiphysics simulation is essential for systems governed by interacting physical processes, classically relying on partitioned analysis to couple interdisciplinary solvers through interface exchange. Neural simulation can inherit this modular architecture by assembling independently pretrained physical modules. However, directly assembling modules pretrained to learn isolated single physics introduces an objective mismatch, as they are never optimized for the target coupled equilibrium. To address this, we derive module-to-system error bounds, proving that coupled accuracy is governed by cross-interface operator coupling rather than module fidelity on isolated single physics alone. This motivates multiphysics post-training, a learning paradigm that adapts pretrained modules via a system-level loss on their joint predictions while keeping each module's inputs, outputs, and exchange structure unchanged. Under this paradigm, we provide two complementary implementations, Finite Unrolling (Unrolled-) for low-latency inference and Equilibrium Post-Training (EquiPT), which achieves activation memory and high fidelity via an implicit adjoint. Across neutron–thermal, reaction–diffusion, and fluid–structure interaction benchmarks, EquiPT reduces system error by 48.7%–96.1% over frozen assembly. These gains stem from training at the coupled equilibrium rather than from coupled data alone, as EquiPT also reduces system error by 34.2%–74.1% relative to module-wise continued training on the same coupled data. Pretrained initialization further improves accuracy under the same post-training budget. Furthermore, post-training shifts the measured coupling spectral radii toward lower values, lowering the ceiling on interface error amplification along each eigenmode.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multiphysics, post training, data-driven simulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_abY2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes multiphysics post-training: jointly adapting separately pretrained single-physics surrogates while preserving their physical interfaces. It introduces EquiPT, which supervises coupled equilibria via implicit differentiation, and Unrolled-K, which trains through a fixed number of field exchanges. A local error analysis characterizes how module errors propagate through coupling. Experiments on neutron–thermal coupling, Gray–Scott reaction–diffusion, and fluid–structure interaction show improved accuracy over frozen assembly and module-wise training, plus analyses of deployment behavior and computational trade-offs.

Critical strengths

The paper tackles a practical problem, namely the mismatch between training modules in isolation and deploying them coupled, and offers a modular solution that preserves physical interfaces, giving a useful way to adapt existing single-physics surrogates to coupled prediction. The evidence is solid: results across three distinct physical systems, including matched-data module-wise continued-training controls, show gains beyond just more data exposure, and the deployment-depth and computational analyses make clear when equilibrium training or finite unrolling is the better choice.

Critical weaknesses

EquiPT’s methodological novelty beyond established DEQ training and related work such as DEQ-MPI could be clearer. The validation/test split also needs clarification: although training and held-out trajectories are separated, it is unclear whether validation and test sets are trajectory-disjoint, given the overlapping GS windows. Finally, the spectral diagnostics do not disentangle module-bias reduction from changes in directional error amplification.

Questions for authors

See weaknesses

4 comments
AuthorsOpenReview ↗

Official Response to Reviewer abY2 (1)

We thank the reviewer for the positive assessment. We appreciate the reviewer's recognition of our modular solution to the training–deployment mismatch, which preserves physical interfaces and adapts existing single-physics surrogates, backed by solid evidence across three physical systems. The reviewer's valuable concerns on methodological novelty, the data split, and the spectral diagnostics helped us strengthen the paper, and we address each below, with revisions highlighted in blue in the revised manuscript.

Weakness 1: EquiPT’s methodological novelty beyond established DEQ training and related work such as DEQ-MPI could be clearer.

Response:

We thank the reviewer for the comment. EquiPT indeed adopts the idea of differentiating through fixed points from Deep Equilibrium Models (DEQs) [1] along with existing root solvers. We do not claim this as our contribution either, as detailed below.

Standard DEQs are monolithic networks trained from random initialization, where the equilibrium is defined over internal latent features, which means existing single-physics models cannot be reused, the equilibrium has no physical ground truth for supervision, and errors cannot be decomposed across physical fields. We evaluated its most favorable variant on our problem (removing only module-wise pre-training while keeping all other components identical to EquiPT) and found that it performs markedly worse than EquiPT (Table 1 and Appendix D.11 of the revised manuscript). Meanwhile, DEQ-MPI [2] addresses the simplest setting for equilibrium methods, where a single unknown field with a known fixed system matrix is solved throughout the iterations. In contrast, our target multiphysics coupling involves multiple interacting unknown fields, where each field operator is learned with inherent approximation bias, and these modular biases mutually amplify across coupled interfaces.

As shown above, existing DEQ related methods are not designed for multiphysics coupling. Applying them directly yields poor performance and lacks theoretical characterization of cross-field error propagation. EquiPT's methodological novelty lies in embedding equilibrium differentiation as a tool within our proposed multiphysics post-training paradigm, while redesigning three core elements of equilibrium training specifically for coupled multiphysics including the equilibrium state definition, the training initialization, and the supervision objective:

  1. Equilibrium state: defined directly over physical fields. The fixed-point map is partitioned by physical field, where each block is a module preserving its physical inputs and outputs, and the equilibrium state consists of the physical fields themselves. Because modules exchange only physical quantities across interfaces, heterogeneous network architectures can be directly coupled and existing single-physics models can be reused.
  2. Training initialization: starting from pretrained modules. Each module is first pretrained on its isolated single physics, allowing equilibrium training to begin from modules that are already accurate on individual fields, which allows higher data efficiency.
  3. Supervision objective: grounded in theory. Because the equilibrium resides in the physical state space, it can be directly supervised against physical ground truth. Our theorem 1 of the original manuscript shows that coupled error arises from modular bias amplified by cross-field sensitivity, indicating that improving isolated single-physics accuracy alone is insufficient, and a system-level loss must be applied directly to the coupled predictions.

We have added the discussion and relevant references to Section 2 of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer abY2 (2)

Weakness 2: The validation/test split also needs clarification: although training and held-out trajectories are separated, it is unclear whether validation and test sets are trajectory-disjoint, given the overlapping GS windows.

Response:

We thank the reviewer for pointing this out. After careful check, we find that there is indeed a mistake of overlapping frames between validation and test sets of the GS dataset in our original submission. To completely eliminate this, we designed a new split and re-ran all training, checkpoint selection, and evaluation for GS.

In the new split of the 25 GS trajectories, the training data remain unchanged (16 trajectories for pre-training and 4 for post-training). The remaining 5 held-out trajectories are partitioned into 2 validation trajectories and 3 test trajectories. Then all training, validation, and test sets are mutually disjoint and share no overlapping frames. For all methods, checkpoints are selected strictly based on validation error, and the test set is used once only at evaluation.

Under this new split, we retrained all GS methods in Table 1 of the revised manuscript while keeping all other configurations identical to our previous setup, again across 3 random seeds (GenCP and M2PDE: single runs). The updated results are shown in Table Rebuttal D1, with relative sample standard deviations in parentheses. These results confirm that the trajectory-disjoint split does not change any of our conclusions. Across all methods, the system errors remain within roughly 7% of the original values, and EquiPT still achieves the lowest error across all metrics. Specifically, EquiPT reduces system error by 50.2% relative to Frozen assembly (vs. 48.7% originally) and by 52.2% relative to MW-CT trained on the same coupled data (vs. 54.2% originally).

In the revised manuscript, we have replaced the GS columns in Table 1 and the corresponding results with these retrained numbers. Due to time constraints, we have retrained only the GS results in Table 1 and the numbers derived from them. Because the new split leaves these results essentially unchanged (all within about 7%), it does not affect any analysis built on them, so the remaining GS analyses keep the original split and will be updated in the next revision. We thank the reviewer again for catching this issue and helping us improve the manuscript.

Table Rebuttal D1. GS errors under the trajectory-disjoint split. Means over three seeds (GenCP and M2PDE: single runs), relative standard deviations in parentheses.

uvSystem
Frozen0.0207 (±27.27%)0.0461 (±22.74%)0.0334 (±24.14%)
GenCP0.01990.05430.0371
M2PDE0.26010.64990.455
EquiPT w/o pre-training0.0189 (±3.49%)0.0467 (±5.63%)0.0328 (±4.98%)
MW-CT0.0213 (±19.14%)0.0483 (±15.58%)0.0348 (±16.61%)
Unrolled-10.0140 (±11.22%)0.0251 (±3.61%)0.0195 (±1.94%)
Unrolled-150.0112 (±14.04%)0.0283 (±12.79%)0.0197 (±13.13%)
EquiPT0.0096 (±5.85%)0.0237 (±4.56%)0.0166 (±4.93%)
AuthorsOpenReview ↗

Official Response to Reviewer abY2 (3)

Weakness 3: Finally, the spectral diagnostics do not disentangle module-bias reduction from changes in directional error amplification.

Response:

We agree with the reviewer's point that the spectral radius alone cannot disentangle these two effects, which is why we treat it strictly as an accompanying diagnostic in the paper.

In fact, h=Ab+O(∥b∥22)h = Ab + \mathcal{O}(\|b\|_2^2) in Theorem 1 of the original manuscript separates the two effects directly, so we explicitly measured both. For each method, we compute three quantities: the modular bias B=∥b∥2B = \|b\|_2, evaluated using reference partner fields as inputs; the equilibrium error H=∥h∥2H = \|h\|_2, evaluated under each method's coupled deployment; and the amplification ratio H/BH/B. The results are reported in Table Rebuttal D2.

On GS and FSI, BB actually increases after EquiPT fine-tuning (by 77% on GS and 6% on FSI), yet EquiPT lowers the overall error by suppressing coupled error amplification (H/BH/B); on NT, both BB and H/BH/B decrease. It should be noted that a smaller modular bias does not by itself guarantee a smaller coupled error. Take NT as an example, MW-CT continues training on the same coupled data using the module-wise objective, thereby achieving a smaller modular bias than EquiPT, yet suffering from much stronger error amplification (1.611.61 vs. 0.300.30), resulting in a larger overall error. This behavior aligns with what each objective supervises: the module-wise objective never evaluates coupled predictions, whereas post-training directly supervises the coupled prediction itself. This is precisely the mismatch between module-wise objectives and coupled prediction that motivates our post-training paradigm.

At the same time, spectral diagnostics still provide valuable complementary insight. First, the spectral radius characterizes the coupling sensitivity of the learned system, and Corollary 2 of the original manuscript links it directly to error amplification by showing that the modular bias along each eigenmode is amplified by at most (1−ρ)−1(1 - \rho)^{-1}. Post-training substantially lowers this ceiling: on NT, the per-sample median drops from 10.13 for Frozen assembly to 3.03 for EquiPT, consistent with the empirical suppression of amplification observed above. The two diagnostics thus serve distinct roles. The spectral radius depends solely on the learned coupled system and, via Corollary 2 of the original manuscript, provides a reference-free upper bound on the gain of any individual eigenmode, whereas H/BH/B measures the actual amplification that the coupled system induces on the modular bias, which also depends on which modes are excited. In short, the former bounds how much the system can amplify a single mode in the worst case, while the latter measures how much it actually amplifies the realized bias. Consistent with our discussion in Section 3.2 of the original manuscript and Section 5.4 of the revised manuscript, the spectral radius is best interpreted as a supporting diagnostic alongside empirical prediction errors.

We have added this analysis to Section 5.4 and included the corresponding table in Appendix D.12 of the revised manuscript.

Table Rebuttal D2. Modular bias BB, equilibrium error HH, and amplification H/BH/B. BB uses reference partner fields as inputs; HH uses each method's deployment solver.

NTGSFSI
BHH/BBHH/BBHH/B
Frozen27.729.51.060.53.77.371.654.362.64
MW-CT0.851.371.610.503.456.841.644.442.71
EquiPT2.490.760.30.891.611.811.751.751.00

References

[1] Bai, Shaojie, et al. "Deep equilibrium models." NeurIPS (2019)

[2] Güngör, Alper, et al. "DEQ-MPI: A deep equilibrium reconstruction with learned consistency for magnetic particle imaging." IEEE Transactions on Medical Imaging (2024)

Thank you for the clarifications and additional analyses, which are helpful in assessing the paper. I maintain my original rating of Weak Accept.

Reviewer_XAUsOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies post-training of independently pretrained neural modules for partitioned multiphysics simulation. The modules are assembled through physical interfaces, and their parameters are jointly adapted using a system-level objective. The paper considers both differentiation through a finite number of coupling iterations and EquiPT, which evaluates the objective at a coupled fixed point and differentiates implicitly through that equilibrium. Experiments examine prediction accuracy, deployment-depth sensitivity, and computational costs across three multiphysics benchmarks.

The problem is relevant, and the experiments provide useful evidence that accurate isolated modules—or accurate finite-depth predictions—need not yield accurate coupled equilibria. My main concern is that the paper does not sufficiently distinguish its methodological contribution from closely related work on coupled fine-tuning, compositional fixed-point simulation, and implicit neural operators.

Critical strengths
  1. A relevant deployment problem. Independently trained physical modules can encounter different inputs after coupling. Adapting them under the assembled system’s objective is a meaningful problem, particularly when preserving existing module interfaces is desirable.

  2. Useful analysis of training–deployment mismatch. The distinction between supervising a finite coupling iterate and supervising the converged equilibrium is experimentally consequential. For example, the reported NT error of Unrolled-15 increases from 0.0096 at its supervised depth to 0.6477 when the same modules are solved to a substantially smaller coupling residual. This illustrates why residual convergence alone does not establish predictive accuracy.

  3. Informative controls. Comparisons against frozen modules, finite-depth training, training from scratch, and approximate equilibrium gradients help identify which aspects of the framework contribute to its performance. The phantom-gradient comparison is particularly useful for separating the equilibrium objective from the choice of backward solver.

Critical weaknesses
  1. The scope of the post-training novelty claim could be clarified. SamudrACE [1] independently pretrains atmosphere and ocean emulators, couples them through physical interfaces, and jointly fine-tunes their weights. The manuscript cites this work primarily in the context of autoregressive rollout fine-tuning, but its connection to the proposed pretraining–coupling–adaptation workflow deserves further discussion. EquiPT makes a meaningful distinction by optimizing a coupled consistency equilibrium rather than a temporal rollout. Qualifying the “first post-training framework” claim around this distinction would make the contribution more precise.

  2. The relationship to existing equilibrium methods deserves a more explicit comparison. Time Without Timesteps [2] composes learned subsystem operators through a self-consistency fixed point and differentiates through that equilibrium, although its optimization experiments concern physical coupling parameters rather than pretrained network weights. FNO-DEQ [3] likewise establishes the use of fixed-point formulations and implicit differentiation in neural operators. I see value in bringing these ideas to the adaptation of pretrained multiphysics modules, particularly given the paper’s empirical analysis of deployment behavior. However, I would appreciate a clearer account of which aspects require new methodological development and which constitute an application or integration of established techniques. This would help readers better appreciate the contribution without relying on an overly broad interpretation of its novelty.

  3. Positioning relative to jointly trained neural operators could be strengthened. A dedicated discussion of structured learning of coupled fields would provide useful context, including Xiao et al. [4], Rahman et al. [5], and Jing et al. [6]. These works incorporate cross-variable interactions through coupled kernels, codomain attention, and shared or variable-specific operator components, illustrating that joint training and structured coupling are compatible choices.

Against this background, I would find it helpful to distinguish the coupling architecture, the pretraining and adaptation protocol, and the state at which supervision is applied. My understanding is that the paper’s particular contribution lies in adapting independently pretrained modules at their coupled equilibrium while retaining their physical interfaces. Explaining this relationship more explicitly would clarify when the proposed approach is advantageous relative to directly learning a coupled neural operator. This is primarily a concern about positioning and interpretation; it does not diminish the useful empirical evidence presented for equilibrium post-training.

References

[1] Duncan et al. SamudrACE: Fast and Accurate Coupled Climate Modeling with 3D Ocean and Atmosphere Emulators. Geophysical Research Letters, 2026; preprint, 2025.

[2] Zerihun and Lee. Time Without Timesteps: Simulating Coupled Dynamical Systems via Self-Consistency. arXiv:2609.03358, 2026. Link

[3] Marwah et al. Deep Equilibrium Based Neural Operators for Steady-State PDEs. NeurIPS, 2023.

[4] Xiao et al. Coupled Multiwavelet Neural Operator Learning for Coupled Partial Differential Equations. 2023.

[5] Rahman et al. Pretraining Codomain Attention Neural Operators for Solving Multiphysics PDEs. NeurIPS, 2024.

[6] Jing et al. Extending Fourier Neural Operators for Modeling Parameterized and Coupled PDEs. ICLR, 2026.

Questions for authors

Stated in the weakness section.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer XAUs (1)

We thank the reviewer for the insightful review. We appreciate that the reviewer recognizes our work addresses a meaningful problem, highlights our useful analysis of the training–deployment mismatch with illustrative examples, and acknowledges our informative controls supported by comprehensive evaluations and detailed ablation studies. The reviewer's valuable suggestions on positioning relative to prior equilibrium methods and coupled neural operators helped us sharpen our claims, and we address each below with corresponding revisions to the revised manuscript.

Weakness 1: The scope of the post-training novelty claim could be clarified. SamudrACE [1] independently pretrains atmosphere and ocean emulators, couples them through physical interfaces, and jointly fine-tunes their weights. The manuscript cites this work primarily in the context of autoregressive rollout fine-tuning, but its connection to the proposed pretraining–coupling–adaptation workflow deserves further discussion. EquiPT makes a meaningful distinction by optimizing a coupled consistency equilibrium rather than a temporal rollout. Qualifying the “first post-training framework” claim around this distinction would make the contribution more precise.

Response:

We thank the reviewer for pointing out this important work, SamudrACE [1]. While we acknowledge that SamudrACE applies post-training to the coupled atmosphere–ocean problem, its formulation relies heavily on the timescale separation, which is specific to atmosphere–ocean dynamics, thereby addressing a simplified coupling regime that is difficult to transfer to other multiphysics domains. In contrast, our approach is not restricted by domain-specific physical properties and targets at general multiphysics coupling. Because our contribution is a post-training paradigm for general multiphysics coupling, our “first” claim, as scoped in the revised Introduction, remains unaffected.

We have added a discussion in Section 2 of the revised manuscript contrasting SamudrACE with our work from this post-training perspective. To clarify the exact scope of our novelty, the Introduction of the revised manuscript specifies our framework as jointly adapting pretrained modules for distinct physics that are bidirectionally coupled and iterated to mutual consistency within each step, supervised by a system-level loss on their coupled predictions. This formulation corresponds to the strongest and most general form of multiphysics coupling, of which a single explicit exchange per step is only a simplified special case.

AuthorsOpenReview ↗

Official Response to Reviewer XAUs (2)

Weakness 2: The relationship to existing equilibrium methods deserves ...... would help readers better appreciate the contribution without relying on an overly broad interpretation of its novelty.

Response:

We thank the reviewer for recognizing the value of bringing equilibrium formulations to multiphysics post-training and for the constructive suggestions.

To better understand our contributions, we start from a fundamental, long-standing pain point in partitioned multiphysics simulation: even when pretrained modules are individually accurate, their coupled solution at joint inference can still suffer from severe error. That is why we focus on this challenging partitioned setting where each physical field is predicted by a separately pretrained module, and at deployment, modules bidirectionally exchange predictions and iterate to convergence. Under this setting, existing partitioned learning approaches suffer from a training–inference mismatch: module-wise pre-training optimizes each single-field prediction in isolation, whereas the coupled state returned at deployment is jointly determined by all modules. Consequently, high accuracy on isolated modules does not guarantee an accurate coupled solution. Resolving this exact mismatch is the core goal of our paper.

In fact, we rely on only two established techniques: (i) implicit differentiation at fixed points (via the Implicit Function Theorem and adjoint methods from DEQs [7]), used for backpropagation in EquiPT; and (ii) classical root solvers (such as Anderson acceleration and Picard iteration), used to solve the coupled equilibrium in the forward pass and the linear adjoint system in the backward pass. Much like standard backpropagation in neural networks, these serve purely as computational tools.

Neither of the two existing works the reviewer mentioned targets this mismatch, and directly transferring them does not resolve the issue:

  • FNO-DEQ [3] is not designed for coupled multiphysics problems. It is a monolithic network for solving a single PDE whose equilibrium is defined over latent features, with no notion of modular decomposition. Adapting it directly to multiphysics introduces several limitations: architecturally, it lacks physical interfaces for modular assembly, preventing the reuse of existing single-physics models; representationally, a latent-space equilibrium lacks physical ground-truth targets, making physical errors difficult to measure explicitly or analyze theoretically; and in terms of inductive bias, losing the explicit physical structure of cross-field coupling can hurt data efficiency during learning.
  • Time Without Timesteps [2] assembles learned subsystem operators to solve for a self-consistent equilibrium, but as the reviewer noted, it optimizes physical coupling parameters while keeping the network weights frozen. Because the modules themselves are never trained at coupled equilibrium, the mismatch remains unresolved.

Since existing works leave this mismatch unaddressed, we developed three new components to solve it:

  1. Modeling and theory. We formulate the coupled inference of pretrained modules as a fixed-point problem directly over physical fields, aligning the learned equilibrium with physical ground truth. Theorem 1 of the original manuscript shows that coupled error arises from modular bias amplified by cross-field sensitivity. This reveals the root cause of the mismatch that conventional module-wise training only constrains modular bias without controlling the amplification induced by coupled assembly, so training individual modules only is never enough.
  2. Training paradigm. Motivated by this insight, we propose a general multiphysics post-training paradigm that introduces a stage between pre-training and deployment, applying a system-level loss directly to the coupled equilibrium predictions to jointly update all module weights. Because modules preserve their physical inputs and outputs and exchange only physics at interface, networks with different architectures can be directly coupled and existing single-physics models can be reused.
  3. Unified framework. Post-training is a paradigm-level framework rather than a single algorithm. Its implementation spans a broad design space across state supervision targets, gradient backward rules, and solver types, allowing practitioners to trade off accuracy, memory, and computational cost according to deployment needs and resource budgets. For example, finite unrolling and EquiPT are two variants within this framework. On NT and GS, the former offers lower inference latency, whereas the latter achieves higher accuracy; the latter also keeps training memory independent of the number of exchanges.

We have added a discussion of both works to Section 2 of the revised manuscript and explicitly outlined these three design choices in Contribution 3 of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer XAUs (3)

Weakness 3: Positioning relative to jointly ...... coupling are compatible choices.

Against this background, I would find it helpful to ...... for equilibrium post-training.

Response:

We thank the reviewer for recommending these references and for the constructive suggestions. The reviewer's summary is very insightful. Following the three dimensions the reviewer outlined, we distinguish our work from Xiao et al. [4], Rahman et al. [5], and Jing et al. [6] as follows:

  • Coupling architecture. These methods use a single network to directly learn the coupled solution, embedding inter-field interactions internally within architecture (via multiwavelet coefficients in Xiao et al., cross-variable attention in Rahman et al., and coupled spectral convolutions in Jing et al.). In contrast, our framework assigns a dedicated module to predict each physical field and these modules exchange physics bidirectionally across interfaces and iterate to mutual consistency within each step.
  • Pre-training and adaptation protocol. The key difference lies in what is learned during pre-training and what is updated during adaptation. Xiao et al. and Jing et al. do not perform pre-training, instead learning all physical fields from scratch directly on coupled data. While CoDA-NO (Rahman et al.) pre-trains a shared attention backbone before fine-tuning on coupled data, its pre-training objective is representations rather than the operator of any physical process. Prediction heads are attached only during fine-tuning, meaning both individual field evolution and cross-field interactions must be learned from coupled samples. By contrast, our pre-training is conducted separately for each physical field so that each module learns the solution operator of its own physical process while preserving physical inputs and outputs. Post-training then uses a system-level loss solely to correct the mismatch of these modules at the coupled state.
  • State at which supervision is applied. These prior approaches supervise the direct output of a single forward pass on coupled data, with no internal coupled equilibrium to solve. In contrast, we supervise the coupled equilibrium reached after iterating the modules to mutual consistency—which is the exact state returned at deployment.

These distinctions determine when our approach is advantageous over learning the coupled solution with a single network.

First, when constituent physical processes differ markedly in spatiotemporal scales, stiffness, or discretizations, or when mature single-physics models have already been developed independently, a unified network struggles to accommodate them or reuse existing solvers. Our modules can be heterogeneous and are assembled directly.

Second, when coupled data is expensive to generate with full multiphysics solvers, our framework needs coupled data only to correct the mismatch at the coupled equilibrium, since each module has already learned its single-physics operator. Removing module-wise pre-training while keeping the same coupled data and budget yields 2.0 to 9.8 times EquiPT’s system error.

Third, when the reliability of coupled predictions is required to be analyzed or diagnosed, our framework provides it by embedding the fixed-point structure of partitioned coupling as a physical inductive bias. Each module approximates its respective physical operator and exchanges physics across interfaces, so the coupling Jacobian directly corresponds to the sensitivity of one field with respect to another. Building directly on this structure, Theorem 1 of the original manuscript establishes error bounds in terms of cross-field sensitivities, and Corollary 2 of the original manuscript further establishes the spectral radius of coupling Jacobian as a diagnostic for error amplification. Jointly trained methods, by contrast, lack explicit physical structure therefore precluding such theoretical analysis.

We have added a discussion of Xiao et al., Rahman et al., and Jing et al. in Section 2 of the revised manuscript.

References

[1] Duncan, James P. C., et al. "SamudrACE: Fast and accurate coupled climate modeling with 3D ocean and atmosphere emulators." Geophysical Research Letters (2026)

[2] Zerihun, Liyu, and Mark Shinyoung Lee. "Time Without Timesteps: Simulating coupled dynamical systems via self-consistency." arXiv preprint arXiv:2609.03358 (2026)

[3] Marwah, Tanya, et al. "Deep equilibrium based neural operators for steady-state PDEs." NeurIPS (2023)

[4] Xiao, Xiongye, et al. "Coupled multiwavelet neural operator learning for coupled partial differential equations." arXiv preprint arXiv:2303.02304 (2023)

[5] Rahman, Ashiqur, et al. "Pretraining codomain attention neural operators for solving multiphysics PDEs." NeurIPS (2024)

[6] Jing, Cheng, et al. "Extending Fourier neural operators for modeling parameterized and coupled PDEs." ICLR (2026)

[7] Bai, Shaojie, et al. "Deep equilibrium models." NeurIPS (2019)

Reviewer_NE3kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the mismatch between separately trained physical surrogate modules and their deployment as a coupled multiphysics system. Each module is pretrained with reference partner fields, but at deployment it must respond to other modules’ imperfect predictions. The proposed multiphysics post-training framework jointly adapts these modules using a loss on their coupled output while preserving their inputs, outputs, and exchange structure.

The theoretical contribution is a local perturbation analysis relating module bias to system error through the coupling resolvent (I−C)−1(I-C)^{-1}, where CC is the coupling Jacobian. It also characterizes modal amplification, the relationship between unrolled and equilibrium gradients, and solver independence under stated regularity and convergence assumptions. Two implementations are developed: Unrolled-KK, which differentiates through a fixed number of exchanges, and EquiPT, which optimizes a numerically solved equilibrium using an implicit adjoint. EquiPT avoids retaining the forward iteration trajectory, giving constant activation-memory complexity with respect to solver iteration count.

Experiments cover neutron–thermal coupling, Gray–Scott reaction–diffusion, and fluid–structure interaction. Comparisons include frozen assembly, generative composition baselines, module-wise continued training on the same coupled data, and equilibrium training without pretrained initialization. The post-training and training-control results are averaged over three runs. EquiPT reduces system error by 48.7%–96.1% versus frozen assembly and 34.2%–74.1% versus module-wise continued training. Additional studies examine inference depth, root solvers, memory, runtime, adjoint approximations, and coupling strength. They show that finite-unrolling models can lose accuracy when solved to equilibrium, whereas equilibrium-trained models retain accuracy across tested solvers reaching the same solution. Reduced measured coupling spectral radii accompany the gains, although they do not alone determine system error.

Critical strengths

The system-level training objective is supported by meaningful experimental controls. The paper explains why accurate individual modules can still produce an inaccurate coupled prediction: interface errors propagate through the assembled system’s resolvent. This motivates joint adaptation while preserving the modules’ existing inputs, outputs, and exchange structure. The module-wise continued-training control is particularly useful because it uses the same coupled data and pretrained initialization but retains separate module objectives. EquiPT reduces system error by 34.2%–74.1% against this control across three benchmarks. These results provide evidence that the improvement depends on how the assembled system is trained, beyond simply providing additional coupled data. This is consequential because it directly supports the central motivation for multiphysics post-training.

Critical weaknesses
  1. The equilibrium-based interpretation needs stronger numerical-convergence controls. Some important comparisons combine model accuracy with substantially different levels of solver convergence. In Table 9, the NT model trained without pretraining has a relative residual of 3.77×10−23.77\times10^{-2}, compared with 1.38×10−41.38\times10^{-4} for pretrained EquiPT; MW-CT has a residual of 6.39×10−36.39\times10^{-3}. These measurements establish performance under the prescribed budgets, but do not determine how much of the accuracy gap remains after each learned system is solved more accurately. The NT comparison across multiple solvers is helpful for Frozen and EquiPT, but does not resolve this issue for the scratch and MW-CT controls. Moreover, Appendix B.3 shows that gradient accuracy depends on both forward and adjoint residuals and conditioning, yet the experiments do not provide corresponding training-time residual distributions or a systematic tolerance-sensitivity study. This matters because the claimed benefits of pretraining and equilibrium optimization depend partly on distinguishing numerical approximation from the accuracy of the learned equilibrium. Convergence-controlled evaluations and backward-solve diagnostics could materially change that interpretation.

  2. The validation/test split is insufficiently specified for the temporal benchmarks. Section 4 states that held-out coupled data are randomly halved into validation and test sets. Appendix C.1 describes Gray–Scott samples as fifteen-frame windows extracted with a ten-frame stride, so neighboring windows overlap. It is unclear whether the final validation/test division is performed by trajectory or by individual window. If it is the latter, overlapping temporal content from the same trajectory could enter both sets. This would make checkpoint selection and final evaluation statistically dependent, weakening the evidence for generalization to independent trajectories. The initial separation of training and held-out trajectories does not resolve this subsequent partitioning question. Likewise, repeating training with three seeds on one fixed split measures optimization variability without establishing robustness to trajectory selection. Since the empirical gains are central to the contribution, the split unit should be made explicit and, if necessary, results should be confirmed using trajectory-disjoint validation and test sets.

Questions for authors
  1. Validation/test independence. Is the final held-out split performed by trajectory or by individual sample for Gray–Scott and FSI? Gray–Scott windows overlap because they contain fifteen frames but are extracted with stride ten. Can you confirm whether validation and test sets share trajectories or time frames? If they do, are results from a trajectory-disjoint evaluation available? This distinction matters because the validation set is used for checkpoint selection.

  2. Convergence of the training controls. In Table 9, the NT scratch and MW-CT controls have substantially larger residuals than EquiPT. When their learned equations are solved with tighter tolerances or larger iteration budgets, how do their prediction errors change? Does the accuracy advantage persist at comparable numerical convergence? The original fixed-budget comparison remains useful, but this check would distinguish numerical solve error from differences between the learned equilibria.

  3. Accuracy of the implicit gradient. What algorithm, tolerance, and iteration limit are used for the adjoint solve? How often do forward or backward solves reach their iteration caps during training? Are residual distributions or gradient checks available to assess the effects of inexact solves, including the different Jacobian evaluation states described in Appendix C.3?

3 comments
AuthorsOpenReview ↗

Official Response to Reviewer NE3k (1)

We thank the reviewer for the thoughtful review. We are pleased that the reviewer finds our system-level training objective supported by meaningful experimental controls, and that the strong performance in reducing prediction error offers direct evidence for our core motivation. The reviewer also considers our module-wise continued-training control strategy particularly useful. The reviewer's valuable concerns on numerical convergence, the data split, and gradient accuracy helped us strengthen the evidence, and we address each below, with new experiments included in the revised manuscript.

Weakness 1: The equilibrium-based interpretation needs stronger numerical-convergence controls. Some important ...... could materially change that interpretation.

Response:

We thank the reviewer for raising this point. The concern is whether our advantage over the baselines relies on their inferior results under limited solver budgets, and whether this gap could be closed if the baselines were solved to lower residuals at inference. We already discussed this issue in the original manuscript (Section 5.3 and Appendix D.5). For all our method and the baselines, coupled prediction accuracy depends on two factors: the quality of model training and the precision with which the fixed-point equations are solved at inference. Training quality is the prerequisite for accurate prediction, whereas inference solver precision only determines how fully that trained capability is realized. For the baselines, the mismatch between training and inference leaves their learned coupled equations inferior to ours, so high solver precision at inference cannot make up for it. Stronger solvers only help methods whose errors stem from unconverged coupling iterations.

To specifically examine the larger relative residuals the reviewer pointed out in Table 9 of the original manuscript for NT MW-CT (6.39×10−36.39 \times 10^{-3}) and EquiPT w/o pre-training (3.77×10−23.77 \times 10^{-2}), we conducted additional experiments on both models with larger iteration budgets, tighter tolerances, and alternative root finders. The results are summarized in Table Rebuttal B1.

As shown, even when MW-CT's residual is driven down to 4.65×10−44.65 \times 10^{-4} (within the same order of magnitude as EquiPT), its error remains at 0.0114, showing almost no further improvement and confirming that its predictions had already converged. For EquiPT w/o pre-training, lacking the single-physics initialization from pre-training makes reducing the residual considerably harder. That is why 500 steps of Picard iteration and Anderson-200 only bring the residual down to 2.98×10−12.98 \times 10^{-1} and 2.12×10−22.12 \times 10^{-2}, respectively. Although Newton–Krylov drives the residual down to 1.72×10−41.72 \times 10^{-4}, the prediction error remains high at roughly 0.0641, still far higher than our 0.0077.

Clearly, the errors of neither baseline stem from incomplete coupling iterations. Strengthening the solver through larger iteration budgets, tighter tolerances, or different root finders therefore cannot narrow their gaps, which are intrinsic to the learned models themselves. Again, solvers cannot alter the learned equilibrium, only training can, which is precisely the motivation behind our post-training paradigm.

We have added Table Rebuttal B1 to Appendix D.11 and incorporated this discussion into Section 5.2 of the revised manuscript.

Table Rebuttal B1. Convergence-controlled comparison on NT. System error and relative exchange residual of the re-solved controls; "Table 9 setting" refers to Table 9 of the original manuscript.

MethodDeployment settingSystem errorRelative exchange residual
EquiPTAnderson-50 (Table 9 setting)0.00771.38E-04
MW-CTPicard-200 (Table 9 setting)0.01176.39E-03
MW-CTPicard-5000.01176.47E-03
MW-CTAnderson-500.01751.04E-03
MW-CTAnderson-200, tol 10⁻⁸0.01144.65E-04
EquiPT w/o pre-trainingAnderson-50 (Table 9 setting)0.07233.77E-02
EquiPT w/o pre-trainingAnderson-200, tol 10⁻⁸0.06712.12E-02
EquiPT w/o pre-trainingPicard-2000.21252.94E-01
EquiPT w/o pre-trainingPicard-5000.21382.98E-01
EquiPT w/o pre-trainingNewton–Krylov, tol 10⁻⁸0.06411.72E-04

For the training-time residuals and gradient checks, see our response to Question 3.

AuthorsOpenReview ↗

Official Response to Reviewer NE3k (2)

Weakness 2: The validation/test split is insufficiently specified for the temporal benchmarks. Section 4 states that held-out coupled data are randomly halved into validation and test sets. Appendix C.1 describes Gray–Scott samples as fifteen-frame windows extracted with a ten-frame stride, so neighboring windows overlap. It is unclear whether the final validation/test division is performed by trajectory or by individual window. If it is the latter, overlapping temporal content from the same trajectory could enter both sets. This would make checkpoint selection and final evaluation statistically dependent, weakening the evidence for generalization to independent trajectories. The initial separation of training and held-out trajectories does not resolve this subsequent partitioning question. Likewise, repeating training with three seeds on one fixed split measures optimization variability without establishing robustness to trajectory selection. Since the empirical gains are central to the contribution, the split unit should be made explicit and, if necessary, results should be confirmed using trajectory-disjoint validation and test sets.

Response:

We thank the reviewer for pointing this out. In all three systems, the held-out cases or trajectories are disjoint from the training data, and held-out samples are assigned to validation and test at the sample level. For NT and FSI, no time frames are shared: NT samples are independent simulations, and FSI samples are non-overlapping windows. For GS, after careful check, we find that there is indeed a mistake of overlapping frames between validation and test sets of the GS dataset in our original submission. To completely eliminate this, we designed a new split and re-ran all training, checkpoint selection, and evaluation for GS.

In the new split of the 25 GS trajectories, the training data remain unchanged (16 trajectories for pre-training and 4 for post-training). The remaining 5 held-out trajectories are partitioned into 2 validation trajectories and 3 test trajectories. Then all training, validation, and test sets are mutually disjoint and share no overlapping frames. For all methods, checkpoints are selected strictly based on validation error, and the test set is used once only at evaluation.

Under this new split, we retrained all GS methods in Table 1 of the revised manuscript while keeping all other configurations identical to our previous setup, again across 3 random seeds (GenCP and M2PDE: single runs). The updated results are shown in Table Rebuttal B2, with relative sample standard deviations in parentheses. These results confirm that the trajectory-disjoint split does not change any of our conclusions. Across all methods, the system errors remain within roughly 7% of the original values, and EquiPT still achieves the lowest error across all metrics. Specifically, EquiPT reduces system error by 50.2% relative to Frozen assembly (vs. 48.7% originally) and by 52.2% relative to MW-CT trained on the same coupled data (vs. 54.2% originally).

In the revised manuscript, we have replaced the GS columns in Table 1 and the corresponding results with these retrained numbers. Due to time constraints, we have retrained only the GS results in Table 1 and the numbers derived from them. Because the new split leaves these results essentially unchanged (all within about 7%), it does not affect any analysis built on them, so the remaining GS analyses keep the original split and will be updated in the next revision. We thank the reviewer again for catching this issue and helping us improve the manuscript.

Table Rebuttal B2. GS errors under the trajectory-disjoint split. Means over three seeds (GenCP and M2PDE: single runs), relative standard deviations in parentheses.

uvSystem
Frozen0.0207 (±27.27%)0.0461 (±22.74%)0.0334 (±24.14%)
GenCP0.01990.05430.0371
M2PDE0.26010.64990.455
EquiPT w/o pre-training0.0189 (±3.49%)0.0467 (±5.63%)0.0328 (±4.98%)
MW-CT0.0213 (±19.14%)0.0483 (±15.58%)0.0348 (±16.61%)
Unrolled-10.0140 (±11.22%)0.0251 (±3.61%)0.0195 (±1.94%)
Unrolled-150.0112 (±14.04%)0.0283 (±12.79%)0.0197 (±13.13%)
EquiPT0.0096 (±5.85%)0.0237 (±4.56%)0.0166 (±4.93%)

Question 1: Validation/test independence. Is the final held-out split ...... used for checkpoint selection.

Response:

See Weakness 2.

Question 2: Convergence of the training controls. In Table 9, the NT ...... between the learned equilibria.

Response:

See Weakness 1.

AuthorsOpenReview ↗

Official Response to Reviewer NE3k (3)

Question 3: Accuracy of the implicit gradient. What algorithm, tolerance, and iteration limit are used for the adjoint solve? How often do forward or backward solves reach their iteration caps during training? Are residual distributions or gradient checks available to assess the effects of inexact solves, including the different Jacobian evaluation states described in Appendix C.3?

Response:

We thank the reviewer for these questions. Taking the NT benchmark as an example, we provide supplementary numerical experiments and clarifications below to address the reviewer's concerns, and have incorporated these results into the revised manuscript. Specifically, we first clarify the solver settings, then report the forward and adjoint residuals during training, the gradient checks, and the effect of where the two Jacobians are evaluated.

  • Solver settings. Both the forward and adjoint solves use Anderson acceleration (via TorchDEQ [1]) with an iteration cap of 50 and a residual norm tolerance of 10−510^{-5}. If a solve reaches the iteration cap, it returns the iterate with the lowest residual. The adjoint system (I−JG(y)⊤)λG=g(I - J_G(y)^\top)\lambda_G = g in Eq. (28) of the original manuscript is solved using vector-Jacobian products (VJPs). Appendix C.3 of the revised manuscript now specifies these settings.
  • Residuals and iteration counts during training. We logged both solves across 64 NT training batches under our standard training setup. The adjoint solve reaches the 10−510^{-5} tolerance on every batch, requiring a median of 16 iterations and a maximum of 26, without ever hitting the iteration cap. In contrast, the forward solve reaches the iteration cap on every batch. The residual decreases over roughly 40 iterations and then plateaus, returning a state with a residual of 2.6×10−42.6 \times 10^{-4} (our tests further confirm that this plateau comes from the TF32 convolutions, and this numerical error does not affect the results). See Appendix D.10 of the revised manuscript for these measurements.
  • Gradient check. We carried out both checks the reviewer suggested. Because finite differences can only evaluate one directional derivative at a time rather than the full gradient of a network directly, we structured this verification in two steps. First, we verified that the converged adjoint yields the exact gradient via finite differences. After solving to convergence, the directional derivative computed by the adjoint along a random parameter direction matches central finite differences with a relative error of 5×10−55 \times 10^{-5}. Second, we measured the cosine similarity between the gradients used during training and the converged adjoint gradients, finding a median of 0.994 and values above 0.96 on every batch. Taken together, these two steps confirm that our training solves preserve the true gradient direction. Both checks have been added to Appendix D.10 of the revised manuscript.
  • Jacobian evaluation points. We conducted experiments confirming that whether the two Jacobians are evaluated at the same state does not change the gradients. Specifically, we recomputed the gradients with both Jacobians evaluated at ww and with both evaluated at y=G(w)y = G(w). Both variants achieve a cosine similarity of at least 0.9997 with our current implementation. This matches our expectation that since yy and ww differ only by a tiny forward residual, the two Jacobians are evaluated at nearly the same point, and the extra term derived in our gradient error analysis in Appendix B.3 of the original manuscript is directly proportional to this distance. Appendix D.10 of the revised manuscript also includes this comparison.

References

[1] Geng, Zhengyang, and J. Zico Kolter. "TorchDEQ: A library for deep equilibrium models." arXiv preprint arXiv:2310.18605 (2023)

Reviewer_1RCyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the mismatch between independently training single-physics neural modules and deploying them as a coupled multiphysics system. It proposes jointly fine-tuning pretrained modules using a loss on their assembled predictions while retaining the existing module interfaces. Two implementations are considered: differentiation through a fixed number of exchanges (Unrolled-K), and implicit differentiation through a numerically solved coupled equilibrium (EquiPT).

The paper provides local error-propagation bounds involving the coupling resolvent, alongside derivations relating unrolled and equilibrium gradients. Experiments cover neutron–thermal transport, Gray–Scott reaction–diffusion, and fluid–structure interaction. EquiPT improves over frozen assembly and module-wise continued training on the same coupled data. Additional experiments examine deployment depth, solver choice, memory consumption, approximate gradients, and coupling spectra.

Critical strengths

The controlled experiments substantiate the benefit of system-level training. Starting from the same pretrained modules, EquiPT improves over module-wise continued training using the same Stage II data and update budget on all three benchmarks. System errors decrease from 0.0123 to 0.0081 on NT, 0.0350 to 0.0161 on GS, and 0.1650 to 0.0428 on FSI. These comparisons are more consequential than the gains over frozen assembly because they address whether additional coupled data alone explains the improvement.

The study provides useful evidence about deployment consistency and computational trade-offs. The depth and solver experiments demonstrate that reducing a learned model’s residual does not necessarily improve its physical prediction. Equilibrium training addresses this mismatch while avoiding storage of the full forward iteration history. The paper also reports the associated latency trade-offs rather than presenting EquiPT as uniformly faster. Together, these findings offer practical guidance for composing learned physical modules.

Critical weaknesses

The discussion of prior modular neural simulation omits a closely related training strategy. Neural Modular Physics for Elastic Simulation (Li et al., arXiv:2512.15083, ICML 2026), Section 3.3, independently trains physically meaningful neural modules and subsequently fine-tunes them jointly using intermediate and final prediction losses. Consequently, separate module pretraining followed by joint adaptation is not itself new. The authors should cite and distinguish this work, clarifying that their contribution concerns adaptation at a coupled multiphysics equilibrium and its associated analysis and differentiation, rather than the general two-stage modular training paradigm. NMP composes constitutive and integration modules sequentially, so it does not subsume the proposed equilibrium formulation.

The methodological novelty is limited, and the incremental benefit of the full adjoint is modest. EquiPT applies established equilibrium differentiation to a physically structured collection of pretrained modules; the main error analysis is a local fixed-point perturbation argument. Moreover, Table 6 shows that a one-step phantom gradient retains most of the improvement: its errors are 0.0084/0.0196/0.0424 versus 0.0077/0.0161/0.0430 for the full adjoint on NT/GS/FSI. On NT, it also uses less memory and training time. This does not undermine system-level post-training, but it limits the additional contribution attributable to the full implicit adjoint and makes the application-level evidence central to acceptance.

Numerical evidence for the implemented equilibrium gradients is incomplete. Equation (28) evaluates the state Jacobian at y = G(w) but the parameter Jacobian at w. These coincide at an exact equilibrium, and Appendix B.3 correctly identifies an additional error term otherwise. However, the paper does not clearly specify the backward linear solver, its tolerance and iteration cap, or report adjoint residuals and gradient checks. Table 9 reports nonzero forward residuals, particularly for some controls, so the exact-equilibrium assumptions are not automatically satisfied by the finite-budget implementation. This leaves uncertainty about gradient accuracy and the sensitivity of comparisons to solver budgets.

Questions for authors
  1. Numerical clarification and validation: What backward solver, stopping criterion, iteration cap, and failure handling are used for Equation (28)? Please report forward and adjoint residuals during training and a gradient check against finite differences or a sufficiently converged differentiated reference on a manageable case. Does evaluating both Jacobians at the same state materially change the results?

  2. Comparison clarification: How sensitive are the principal comparisons to tighter solve budgets, especially NT MW-CT and EquiPT without pretraining, whose reported relative residuals are 6.39 × 10⁻³ and 3.77 × 10⁻²? Are their accuracy gaps preserved when residuals are brought closer to those of pretrained EquiPT?

  3. Contribution clarification: Does the full adjoint consistently outperform the phantom rule across repeated seeds and matched wall-clock training budgets? Table 6 suggests that the cheaper approximation captures most of the benefit. Establishing when the additional adjoint computation is worthwhile would materially affect my assessment of EquiPT’s incremental contribution.

  4. Evaluation-protocol clarification: Is the validation/test halving performed by trajectory or by individual sample? For GS, consecutive windows overlap, so a sample-level split could place shared frames in validation and test. Please clarify the split unit and whether trajectory-disjoint evaluation changes the conclusions.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer 1RCy (1)

We thank the reviewer for the detailed review. We appreciate the reviewer's recognition of our work in addressing the training–deployment mismatch of prior modular approaches and offering practical guidance for composing learned physical modules with strong performance and well-considered computational trade-offs. The reviewer's valuable questions on novelty, gradient accuracy, and the evaluation protocol helped us strengthen the paper, and we address each below, with revisions highlighted in blue in the revised manuscript.

Weakness 1: The discussion of prior modular neural simulation omits a closely related training strategy. Neural Modular Physics for Elastic Simulation (Li et al., arXiv:2512.15083, ICML 2026), Section 3.3, independently trains physically meaningful neural modules and subsequently fine-tunes them jointly using intermediate and final prediction losses. Consequently, separate module pretraining followed by joint adaptation is not itself new. The authors should cite and distinguish this work, clarifying that their contribution concerns adaptation at a coupled multiphysics equilibrium and its associated analysis and differentiation, rather than the general two-stage modular training paradigm. NMP composes constitutive and integration modules sequentially, so it does not subsume the proposed equilibrium formulation.

Response:

We thank the reviewer for pointing out Neural Modular Physics (NMP) [1]. In fact, NMP focuses solely on a single elastic body rather than a multiphysics system, representing a degenerate case of the coupled multiphysics problem studied in our paper.

While we agree that the general sequence of pre-training followed by fine-tuning is widespread across deep learning, introducing this paradigm into a new domain involves distinct transfer challenges and meaningfully advances in the field. Specifically within partitioned multiphysics simulation, we are indeed the first to establish a general two-stage post-training paradigm. Formulating and adapting this modeling framework for coupled equilibria is non-trivial, and its novelty is not diminished by NMP applying pre-training and fine-tuning to a single elastic body. This is just like how the emergence and impact of post-training in LLMs are not diminished by earlier post-training practices in other areas. We have cited NMP and added this discussion to the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer 1RCy (2)

Weakness 2: The methodological novelty is limited, and the incremental benefit of the full adjoint is modest. EquiPT ...... makes the application-level evidence central to acceptance.

Question 3: Contribution clarification: Does the full adjoint consistently outperform ...... would materially affect my assessment of EquiPT’s incremental contribution.

Response:

We thank the reviewer for the thoughtful comments. The reviewer suggests that because our method adopts established equilibrium differentiation and the one-step phantom rule retains most of the accuracy gains, the incremental contribution attributable to the full implicit adjoint is limited. However, this inference does not hold since the differentiation rule itself is never claimed as our contribution. We adopt it as an established technique or computational tool, just like backpropagation in standard neural networks. Consequently, the results in Table 6 of the original manuscript neither weaken our contributions nor diminish our novelty, as detailed below.

A differentiation rule only computes the gradient once the loss is fixed; our contributions determine where that loss should be applied and why. Through local perturbation analysis, Theorem 1 of the original manuscript shows that coupled error stems from modular bias amplified by cross-field sensitivity. Hence, simply training individual modules to higher isolated accuracy is insufficient, and the loss must directly supervise coupled predictions. Motivated by this, we propose the multiphysics post-training paradigm and formulate a unified framework structured around three design choices: coupled state, gradient rule, and numerical solver. The full adjoint and the one-step phantom rule simply represent two options within the design space of the gradient rule. The former computes the gradient of the system-level loss via the adjoint equation, whereas the latter is a low-cost approximation of that gradient (Appendix D.2 of the revised manuscript). Therefore, the fact that the phantom rule retains most of the improvements actually confirms our central contribution: optimizing the equilibrium objective in EquiPT resolves the mismatch between decoupled pre-training and coupled inference, while the backward rule is merely the tool used to optimize this objective.

We have also conducted additional multi-seed and equal-budget experiments as the reviewer's advice to expand Table 6 of the original manuscript, as shown in Table Rebuttal A1.

Under matched update steps, the full adjoint achieves lower error by roughly 9% across every random seed on NT and reduces error by 18% on GS, while the two rules perform comparably on FSI. The full adjoint also aligns much more closely with the converged equilibrium gradient. Take NT as instance, its median cosine similarity against a converged reference gradient is 0.99, compared to 0.75 for the phantom rule (Appendix D.10 of the revised manuscript). Theoretically, the full adjoint gradient expands into a Neumann series, whereas the one-step phantom rule retains only the zeroth-order term (Appendix B.3 of the original manuscript).

Under a matched wall-clock budget, still taking NT as an example, where each full-adjoint update takes 1.4 times the runtime and 2 times the peak memory of the phantom rule (Appendix D.2 of the revised manuscript), both rules achieve almost identical error. Under the same wall-clock training time, the phantom rule reaches 0.0074 after 2,800 updates and the full adjoint reaches 0.0077 after 2,000 updates (Table Rebuttal A2). This indicates that when compute budget is fixed, the phantom rule can trade its lower per-step cost for more updates to maintain performance. Accordingly, we adopt the full adjoint as the default in our main method since it most closely tracks the gradient of the equilibrium objective, while recommending the phantom rule when per-step runtime or GPU memory is the primary constraint.

We have updated Table 6 to report three-seed means and added the matched-budget results and practical selection guidelines to Appendix D.2 of the revised manuscript.

Table Rebuttal A1. One-step phantom rule vs. full adjoint. System error, mean ± standard deviation over three seeds.

NTGSFSI
One-step phantom0.0089 ± 0.00040.0195 ± 0.00120.0425 ± 0.0001
Full adjoint0.0081 ± 0.00040.0161 ± 0.00050.0428 ± 0.0026

Table Rebuttal A2. One-step phantom rule vs. full adjoint at matched wall-clock time on NT. Same seed, both rules trained for the same wall-clock time (about 6,600 s); peak memory of one forward/backward pass at batch size two.

RuleUpdatesPeak memory (GiB)System error
One-step phantom2,8003.200.0074
Full adjoint2,0006.400.0077
AuthorsOpenReview ↗

Official Response to Reviewer 1RCy (3)

Weakness 3: Numerical evidence for the implemented equilibrium gradients is incomplete. Equation (28) evaluates the state Jacobian at y = G(w) but the parameter Jacobian at w. These coincide at an exact equilibrium, and Appendix B.3 correctly identifies an additional error term otherwise. However, the paper does not clearly specify the backward linear solver, its tolerance and iteration cap, or report adjoint residuals and gradient checks. Table 9 reports nonzero forward residuals, particularly for some controls, so the exact-equilibrium assumptions are not automatically satisfied by the finite-budget implementation. This leaves uncertainty about gradient accuracy and the sensitivity of comparisons to solver budgets.

Response:

See Question 1 and Question 2.

Question 1: Numerical clarification and validation: What backward solver, stopping criterion, iteration cap, and failure handling are used for Equation (28)? Please report forward and adjoint residuals during training and a gradient check against finite differences or a sufficiently converged differentiated reference on a manageable case. Does evaluating both Jacobians at the same state materially change the results?

Response:

We thank the reviewer for these questions. Taking the NT benchmark as an example, we provide supplementary numerical experiments and clarifications below to address the reviewer's concerns, and have incorporated these results into the revised manuscript. Specifically, we first clarify the solver settings, then report the forward and adjoint residuals during training, the gradient checks, and the effect of where the two Jacobians are evaluated.

  • Solver settings. Both the forward and adjoint solves use Anderson acceleration (via TorchDEQ [2]) with an iteration cap of 50 and a residual norm tolerance of 10−510^{-5}. If a solve reaches the iteration cap, it returns the iterate with the lowest residual. The adjoint system (I−JG(y)⊤)λG=g(I - J_G(y)^\top)\lambda_G = g in Eq. (28) of the original manuscript is solved using vector-Jacobian products (VJPs). Appendix C.3 of the revised manuscript now specifies these settings.
  • Residuals and iteration counts during training. We logged both solves across 64 NT training batches under our standard training setup. The adjoint solve reaches the 10−510^{-5} tolerance on every batch, requiring a median of 16 iterations and a maximum of 26, without ever hitting the iteration cap. In contrast, the forward solve reaches the iteration cap on every batch. The residual decreases over roughly 40 iterations and then plateaus, returning a state with a residual of 2.6×10−42.6 \times 10^{-4} (our tests further confirm that this plateau comes from the TF32 convolutions, and this numerical error does not affect the results). See Appendix D.10 of the revised manuscript for these measurements.
  • Gradient check. We carried out both checks the reviewer suggested. Because finite differences can only evaluate one directional derivative at a time rather than the full gradient of a network directly, we structured this verification in two steps. First, we verified that the converged adjoint yields the exact gradient via finite differences. After solving to convergence, the directional derivative computed by the adjoint along a random parameter direction matches central finite differences with a relative error of 5×10−55 \times 10^{-5}. Second, we measured the cosine similarity between the gradients used during training and the converged adjoint gradients, finding a median of 0.994 and values above 0.96 on every batch. Taken together, these two steps confirm that our training solves preserve the true gradient direction. Both checks have been added to Appendix D.10 of the revised manuscript.
  • Jacobian evaluation points. We conducted experiments confirming that whether the two Jacobians are evaluated at the same state does not change the gradients. Specifically, we recomputed the gradients with both Jacobians evaluated at ww and with both evaluated at y=G(w)y = G(w). Both variants achieve a cosine similarity of at least 0.9997 with our current implementation. This matches our expectation that since yy and ww differ only by a tiny forward residual, the two Jacobians are evaluated at nearly the same point, and the extra term derived in our gradient error analysis in Appendix B.3 of the original manuscript is directly proportional to this distance. Appendix D.10 of the revised manuscript also includes this comparison.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all four reviewers (1RCy, NE3k, XAUs, abY2) for their careful reading and constructive feedback. We are grateful that the reviewers recognized the training–deployment mismatch as a relevant problem (XAUs, abY2), found our error analysis informative (NE3k), and considered our controlled comparisons convincing (1RCy, NE3k, XAUs, abY2). We also thank the reviewers for the raised questions and weaknesses, and we have revised the manuscript accordingly.

To frame our responses, we first restate the problem and our contributions. We study partitioned multiphysics simulation, where independently pretrained modules, one per physical field, exchange predictions bidirectionally and iterate to a coupled equilibrium within each step. Module-wise pre-training never optimizes this coupled equilibrium, so individually accurate modules can still yield inaccurate coupled solutions. To resolve this mismatch, we make three contributions:

  1. Modeling & theory. We formulate coupled inference as a fixed-point problem over physical fields. Theorem 1 of the original manuscript shows that coupled error is modular bias amplified by cross-field sensitivity, so higher isolated module accuracy alone is insufficient.
  2. Multiphysics post-training paradigm. A stage between pre-training and deployment applies a system-level loss to coupled predictions, jointly adapting all modules while preserving their physical interfaces.
  3. Unified framework & validation. A method is specified by three design choices in our framework: the supervised state, the gradient rule, and the solver. Across three distinct coupled systems, our EquiPT cuts system error by 50.2% to 96.1% over frozen assembly and 34.2% to 74.1% over MW-CT.

Positioning Clarified (Reviewers 1RCy, XAUs, abY2)

Several reviewers asked how these contributions relate to prior work. We clarify their positioning as follows.

  • Novelty. We adopt only fixed-point differentiation rules and classical root solvers. All three contributions above were newly developed for this mismatch. Prior works noted by the reviewers do not target this mismatch and cannot resolve it directly.
  • Scope of the "first" claim. The revised Introduction scopes the claim to pretrained modules for distinct physical fields that are bidirectionally coupled, iterated to mutual consistency within each step, and jointly adapted by a system-level loss. This is the most general form of multiphysics coupling. Handling this form naturally covers the remaining forms as well, making our post-training paradigm generally applicable.
  • Advantages over jointly trained coupled neural operators (XAUs). Our framework reuses heterogeneous single-physics modules, needs coupled data only to correct their mismatch at the coupled state, and its explicit coupling structure supports error bounds and reference-free diagnostics.

New Experiments & Analyses (All Supporting Our Conclusions in original manuscript)

The remaining concerns asked for stronger empirical and numerical evidence. We added the following experiments and analyses, all of which support our original conclusions.

  1. GS trajectory-disjoint split & retraining (1RCy, NE3k, abY2). Acknowledging overlapping frames in the original GS validation/test set, we redesign a split and retrained all methods on GS in Table 1. EquiPT stays lowest and all methods remain within ~7% of original values, leaving conclusions unchanged.
  2. Fairness under matched solver convergence (1RCy, NE3k). Solving MW-CT and EquiPT w/o pre-training down to residuals within the same order of magnitude as EquiPT still leaves errors far higher than EquiPT. Therefore, stronger solvers cannot close these gaps, which are intrinsic to the learned modules.
  3. Disentangling modular bias vs. error amplification (abY2). Measuring both quantities via Theorem 1 of the original manuscript shows that on NT, MW-CT has smaller modular bias than EquiPT yet far stronger error amplification (1.61 vs. 0.30), yielding higher coupled error, indicating that lowering modular bias alone does not reduce coupled error. Besides, our proposed spectral diagnostics remain a useful reference-free supporting tool.
  4. Numerical verification of implicit adjoint gradients (1RCy, NE3k). Converged adjoint directional derivatives match central finite differences within 5×10−55\times 10^{-5} relative error. Training gradients achieve 0.994 median cosine similarity with converged reference gradients. Besides, evaluating both Jacobians at the same state leaves gradients virtually unchanged.
  5. Full adjoint vs. one-step phantom rule (1RCy). Across 3 seeds, full adjoint cuts error by ~9% on NT and ~18% on GS compared to phantom, and matching it on FSI. Appendix D.2 of the revised manuscript adds equal-wall-clock results and guidelines (full adjoint by default, phantom under tight compute/memory).

Please see our point-by-point responses below for complete details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.