acceptodds
Under review as a conference paper at ICLR 2027

Neural Constitutive Learning for Generalized Reaction-Diffusion Systems

Abstract

Generalized reaction-diffusion systems encompass diverse transport mechanisms and coupled reaction kinetics. A central question for neural PDE solvers is what should be learned so that a common interface can accommodate phase-field and degenerate transport, local reactions, and multispecies coupling. We propose the Neural Constitutive Laws–Mass-Compression-Transport (NCL-MCT) Solver, which learns PDE-specific constitutive responses while retaining temporal evolution in a shared MCT integrator. Transport is represented through mobility and thermodynamic driving force, and reaction through relative reaction rates. These constitutive responses depend on the current density rather than explicitly on the initial condition or elapsed time, motivating their reuse across different initial conditions and time horizons. The same interface supports velocity-data supervision and known-law supervision, neither of which requires time integration during training. When constitutive laws are known, supervision can be evaluated on independently sampled density fields, enabling trajectory-free constitutive learning without generating solution trajectories. Across seven systems, separately trained constitutive modules share the same interface and MCT integrator and achieve relative rollout errors of to . Tests with unseen initial-condition families and an extended time horizon assess reuse beyond training conditions, while separate experiments demonstrate trajectory-free constitutive learning. These results support constitutive responses as an effective learning target for a shared neural PDE framework.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural constitutive learning, neural PDE solvers, generalized reaction-diffusion systems, structure-preserving learning, energetic variational approach, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_VcksOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the Neural Constitutive Laws–Mass-Compression-Transport (NCL-MCT) Solver, which learns PDE-specific constitutive responses for generalized reaction-diffusion systems while keeping density evolution in one shared integrator. The density balance is ∂tρ+∇⋅(ρu)=ρr\partial t \rho + \nabla \cdot (\rho u) = \rho r. Following the energetic variational approach, transport is closed by a positive mobility and a thermodynamic driving force, u=ξfu = \xi f with ξ=ρ/η\xi = \rho / \eta and f=−∇(δE/δρ)f = -\nabla(\delta E / \delta \rho), and reaction by a relative rate r=S(ρ)/ρr = S(\rho) / \rho. These responses depend only on the current density, so the same modules can be reused across initial conditions and time horizons. The shared MCT integrator splits ρ=MI\rho = M I, advancing compression II by a conservative finite-volume step and mass MM by a semi-Lagrangian step. A convolutional neural operator predicts (ξ,f)(\xi, f) and a pointwise network predicts rr. Training does not differentiate through time integration: velocity data supervise only the product u=ξfu = \xi f, with a curl penalty on ff, while known laws supervise ξ\xi and ff separately, including on density fields that are not solution snapshots. Across seven systems, relative rollout L2L^2 errors are about 10−410^{-4} to 10−210^{-2}. Without retraining, errors on unseen initial-condition families are about 6×6\times lower than the strongest finite baseline, and over a horizon about 1.67×1.67\times the training window the errors change by at most about 1.5×1.5\times, against increases of up to about 5×5\times for the baselines. The main limitation is the assumption ρ>0\rho > 0, which is strained by vanishing density in porous-medium flow.

Critical strengths
  • The experimental design covers a wide range of generalized reaction-diffusion behavior and tests whether a learned constitutive module can be reused. Seven systems span linear diffusion, phase-field transport, degenerate diffusion, local reaction, and two-species coupling. Both velocity-data and known-law supervision are reported against F-FNO, CFO, and DOOL, with relative rollout L2L^2 errors from about 10−410^{-4} to 10−210^{-2}. The evaluation also goes beyond in-distribution rollouts: unseen initial-condition families, a horizon about 1.67×1.67\times the training window, and trajectory-free training on independently sampled density fields.
  • The method is organized around a clear split between system-specific constitutive responses and a shared evolution procedure. Transport is represented by mobility and thermodynamic driving force, and reaction by a relative rate, while the MCT integrator advances compression and mass and reconstructs the density. The two supervision modes match what the data can identify: velocity observations constrain only their product, with a curl penalty on the driving force, whereas known laws supervise the two factors separately. Training does not backpropagate through time integration, and the same interface is reused across PDEs.
Critical weaknesses
  • The experiments do not report training convergence or inference speed. There are no loss curves, and no wall-clock comparison of NCL-MCT with F-FNO, CFO, DOOL, or with the finite-difference and finite-volume solvers used to generate the data. A learned PDE solver is useful when inference is substantially faster than a classical discretization. This comparison is especially relevant here, because each step still evaluates a neural constitutive map inside a finite-volume and semi-Lagrangian update, so the cost relative to a standard scheme is not obvious from the accuracy tables.
  • The paper does not explain how the loss weights were chosen. The velocity objective adds three terms, a velocity-matching term, a curl penalty, and a reaction term, and sets the curl weight to 0.010.01. The known-law objective likewise sums mobility, force, and reaction errors. There is no sensitivity study or selection procedure for these coefficients.
  • The paper provides no approximation, stability, or rollout-error guarantee for the learned constitutive map composed with the MCT integrator. Comparable operator-learning methods do have such results. Chen and Chen (1995) give the classical universal approximation theorem for nonlinear operators. Lanthaler, Mishra, and Karniadakis (2022) prove approximation and generalization bounds for DeepONets. Kovachki, Lanthaler, and Mishra (2021) prove universal approximation for Fourier neural operators and explicit error bounds for PDE solution operators.

References:

  • Tianping Chen and Hong Chen. "Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems." IEEE Transactions on Neural Networks, 6(4):911–917, 1995.
  • Samuel Lanthaler, Siddhartha Mishra, and George Em Karniadakis. "Error estimates for DeepONets: a deep learning framework in infinite dimensions." Transactions of Mathematics and Its Applications, 6:tnac001, 2022.
  • Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. "On universal approximation and error bounds for Fourier neural operators." Journal of Machine Learning Research, 22(290):1–76, 2021.
Questions for authors

Please see the concerns in the weaknesses.

2 comments

We thank the reviewers for their constructive comments and recognition of our framework’s physical structure, accuracy, and generalization. The seven systems share a constitutive interface and time-stepper, while each system trains its own response modules. Revisions are marked in blue. The main changes are:

Summary of our response

  1. Contributions and ablations (u59j, Ds7E, qY9r, Vcks). We distinguish Vel. supervision of the mobility–force product from Law supervision of the individual responses under a specified physical convention. We add time-stepper comparisons using identical network weights, CNO–DeepONet comparisons, and curl-penalty weight ablations, distinguishing within-checkpoint comparisons from variation across training runs (Sec. 4.3; Appendices I.1, I.2, I.4).

  2. I/M material state (qY9r). We clarify how the I/M decomposition retains supplied material composition information, enabling composition-dependent dynamics. In a two-population example, we show how common compression cancels in the material derivative of the population ratio and how both MCT and direct population evolution reproduce the separation of states with equal total density (Secs. 4.2, 5.1; Appendix J.5).

  3. Constitutive identifiability (Ds7E). We show that distinct physically valid energy–dissipation laws can produce the same mobility–force product and density evolution. Thus, non-uniqueness under velocity supervision is compatible with the underlying physics (Secs. 3, 4.3; Appendix A.4).

  4. Generalization and module reuse (u59j, qY9r). We distinguish checkpoint reuse for new initial conditions from evaluation beyond a shortened training window, and clarify Schnakenberg’s state-range screening. We also reuse single-population velocity and growth networks in a two-population system without retraining (Appendices G, J.5).

  5. Conditional error analysis and statistics (u59j, Vcks). We establish a conditional finite-horizon bound between MCT rollouts with learned and numerically evaluated constitutive responses at fixed discretization. We also add sampled one-step diagnostics across seven PDEs and 95% confidence intervals for mean rollout errors over ten Schnakenberg trajectories (Appendices A.5, I.1, J.2).

  6. Training, runtime, and numerical settings (u59j, Vcks). We add training-loss curves, rollout and per-step timing breakdowns, and per-PDE factor-reset intervals. For Schnakenberg, we compare native and learned updates on the same GPU, using the same precision and batch size, including a comparison at the same time step (Appendices C, J.1).

Response to Reviewer Vcks

Q1. Can the authors provide training-loss curves and wall-clock comparisons with F-FNO, CFO, DOOL, and the reference numerical solvers to assess training convergence and inference efficiency?

A1. Thank you. We added training-loss histories for all 14 full-window NCL-MCT Solver models. From the first to the final recorded value, the losses decrease by approximately three to five orders of magnitude. These histories indicate optimization progress but do not establish convergence by the final update; Vel. and Law also optimize different objectives (Appendix J.1; Figure 24).

We clarified that the rollout timings include network evaluations and numerical updates, and added a profiled step-time breakdown. On a single NVIDIA GeForce RTX 5090, NCL-MCT Solver requires less training time than F-FNO and CFO for 100,000 optimizer steps across all seven systems. Its mean per-step rollout time is 1.94–4.10 ms, shorter than CFO’s and longer than F-FNO’s. On the three nonreactive systems, DOOL has the shortest training and rollout times but larger rollout errors than NCL-MCT Solver (Appendix J.1; Tables 1 and 16).

For Schnakenberg, we also compare numerically evaluated and learned constitutive responses within the same MCT time-stepper. Over 20,000 steps, the mean rollout time is 62.01 s with Numerical–Numerical responses and 82.00 s with Neural–Neural responses (Appendix I.3; Table 14).

Q2. How were the loss weights chosen, particularly the curl-penalty weight of 0.01, and how sensitive is performance to these choices?

A2. Thank you. We include the curl penalty to suppress spurious local rotation in the learned driving force and encourage its curl-free physical structure, f=−∇(δE/δρ)\mathbf{f}=-\nabla(\delta E/\delta\rho). On Fisher–KPP, λcurl=0.01\lambda_{\mathrm{curl}}=0.01 gives the lowest mean ErollE_{\mathrm{roll}}, Emax⁡E_{\max}, and NAsym among the tested weights, reducing mean NAsym by approximately 87% relative to a curl-penalty weight of zero (Table 12). The Schnakenberg results are provided in Appendix I.2 (Table 13).

Ablation of the curl-penalty weight on Fisher–KPP under velocity-data supervision with MCT (Appendix I.2; Table 12). Entries are mean ± SD over ten test trajectories with 500 rollout steps. NAsym is averaged over 101 evaluation times per trajectory.

λcurl\lambda_{\mathrm{curl}}ErollE_{\mathrm{roll}}Emax⁡E_{\max}NAsym
0(1.03±0.23)×10−3(1.03\pm0.23)\times10^{-3}(1.32±0.36)×10−3(1.32\pm0.36)\times10^{-3}(1.77±0.36)×10−3(1.77\pm0.36)\times10^{-3}
0.01(5.01±1.39)×10−4(5.01\pm1.39)\times10^{-4}(5.96±1.96)×10−4(5.96\pm1.96)\times10^{-4}(2.38±1.39)×10−4(2.38\pm1.39)\times10^{-4}
0.1(8.61±1.59)×10−4(8.61\pm1.59)\times10^{-4}(1.17±0.26)×10−3(1.17\pm0.26)\times10^{-3}(3.08±1.31)×10−4(3.08\pm1.31)\times10^{-4}
1(1.13±0.19)×10−3(1.13\pm0.19)\times10^{-3}(1.72±0.53)×10−3(1.72\pm0.53)\times10^{-3}(3.20±1.11)×10−3(3.20\pm1.11)\times10^{-3}

NAsym. Following the normalized Jacobian-asymmetry diagnostic in [1], we measure the skew-symmetric share of the spatial force Jacobian. NAsym lies in [0,1][0,1]; zero indicates a symmetric spatial Jacobian, while larger values indicate greater local rotation.

[1] Chao et al. On Investigating the Conservative Property of Score-Based Generative Models. ICML 2023. arXiv:2209.12753.

Q3. What approximation, stability, or rollout-error guarantees can be established for the learned constitutive maps coupled with the MCT time-stepper?

A3. Thank you. We establish a conditional finite-horizon bound on the density difference between MCT rollouts using learned and numerically evaluated constitutive responses. At a fixed grid, time step, and reset schedule, the bound holds when both rollouts remain in a common admissible set and the stated one-step discrepancy and Lipschitz assumptions are satisfied. A bound relative to the continuous PDE solution would additionally require a discretization-error estimate (Sec. 4.2; Appendix A.5).

We also evaluate one-step discrepancies and perturbation effects across seven PDEs, using 20 reinitialized reference states and 480 sampled perturbations per PDE. Maximum relative one-step discrepancies range from 3.55×10−63.55\times10^{-6} to 4.33×10−54.33\times10^{-5}, and all sampled perturbation ratios are below one. These are local density-space diagnostics at the sampled states; they do not establish uniform stability over a full rollout or under mesh refinement (Appendix J.2; Table 17).

Reviewer_qY9rOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes NCL-MCT Solver, a neural method for solving generalized reaction-diffusion PDEs by learning their underlying constitutive laws. The authors parameterize the mobility, driving force, and relative reaction rate using neural networks, which are then incorporated into the governing PDE to predict the evolution of density fields. The training loss combines velocity prediction error, curl regularization to encourage gradient fields, and reaction-rate supervision using the known source law. The networks are trained without requiring PDE integration. The approach is evaluated across several physical systems, including experiments with unseen initial conditions and extended prediction horizons, where it achieves lower rollout errors than the considered baselines.

Critical strengths

The main strength of the paper is its successful integration of neural models into the solution of generalized reaction-diffusion systems, achieving good results in terms of accuracy and generalization.

Critical weaknesses

[W1] Limited ablation analysis. The experiments do not clearly isolate the contributions of the different loss components and architectural choices. A more detailed ablation study would help identify which elements are responsible for the reported improvements and whether all the proposed components are necessary.

[W2] Limited generalization evaluation. Although the authors evaluate the method on unseen initial conditions and extended prediction horizons, these experiments consider relatively specific generalization scenarios. Additional experiments involving more diverse distribution shifts would provide stronger evidence of the method's robustness and applicability beyond the considered settings.

Overall, I find the paper's contribution interesting and its presentation well-structured. However, the work would benefit from a more comprehensive ablation study and a broader evaluation of generalization. I should clarify that my assessment focuses primarily on the machine learning aspects of the paper, as my expertise in the specific application domain is limited. Given these concerns and the limitations of my domain-specific expertise, I recommend weak rejection.

Questions for authors

None

4 comments

We thank the reviewers for their constructive comments and recognition of our framework’s physical structure, accuracy, and generalization. The seven systems share a constitutive interface and time-stepper, while each system trains its own response modules. Revisions are marked in blue. The main changes are:

Summary of our response

  1. Contributions and ablations (u59j, Ds7E, qY9r, Vcks). We distinguish Vel. supervision of the mobility–force product from Law supervision of the individual responses under a specified physical convention. We add time-stepper comparisons using identical network weights, CNO–DeepONet comparisons, and curl-penalty weight ablations, distinguishing within-checkpoint comparisons from variation across training runs (Sec. 4.3; Appendices I.1, I.2, I.4).

  2. I/M material state (qY9r). We clarify how the I/M decomposition retains supplied material composition information, enabling composition-dependent dynamics. In a two-population example, we show how common compression cancels in the material derivative of the population ratio and how both MCT and direct population evolution reproduce the separation of states with equal total density (Secs. 4.2, 5.1; Appendix J.5).

  3. Constitutive identifiability (Ds7E). We show that distinct physically valid energy–dissipation laws can produce the same mobility–force product and density evolution. Thus, non-uniqueness under velocity supervision is compatible with the underlying physics (Secs. 3, 4.3; Appendix A.4).

  4. Generalization and module reuse (u59j, qY9r). We distinguish checkpoint reuse for new initial conditions from evaluation beyond a shortened training window, and clarify Schnakenberg’s state-range screening. We also reuse single-population velocity and growth networks in a two-population system without retraining (Appendices G, J.5).

  5. Conditional error analysis and statistics (u59j, Vcks). We establish a conditional finite-horizon bound between MCT rollouts with learned and numerically evaluated constitutive responses at fixed discretization. We also add sampled one-step diagnostics across seven PDEs and 95% confidence intervals for mean rollout errors over ten Schnakenberg trajectories (Appendices A.5, I.1, J.2).

  6. Training, runtime, and numerical settings (u59j, Vcks). We add training-loss curves, rollout and per-step timing breakdowns, and per-PDE factor-reset intervals. For Schnakenberg, we compare native and learned updates on the same GPU, using the same precision and batch size, including a comparison at the same time step (Appendices C, J.1).

Response to Reviewer qY9r

Q1. How do individual loss components and architectural choices contribute to performance, and are all components necessary?

A1. Thank you. We added controlled ablations that separate the parameterization gain from the curl penalty and show that the mean-error ordering persists across the tested time-steppers. CNO further improves rollout accuracy over the tested fully connected DeepONet.

Constitutive parameterization and supervision. Direct predicts u\mathbf{u}, Vel. supervises the product u=ξf\mathbf{u}=\xi\mathbf{f}, and Law supervises ξ\xi and f\mathbf{f} separately. With λcurl=0\lambda_{\mathrm{curl}}=0, factorization reduces mean ErollE_{\mathrm{roll}} for UU from 3.40×10−23.40\times10^{-2} to 2.44×10−22.44\times10^{-2} under identical supervision (Appendices I.1–I.2; Tables 10, 13).

Curl penalty. We include the curl penalty to suppress spurious local rotation in the learned driving force and encourage its curl-free physical structure, f=−∇(δE/δρ)\mathbf{f}=-\nabla(\delta E/\delta\rho). On Fisher–KPP, λcurl=0.01\lambda_{\mathrm{curl}}=0.01 gives the lowest mean ErollE_{\mathrm{roll}}, Emax⁡E_{\max}, and NAsym among the tested weights, reducing mean NAsym by approximately 87% relative to a curl-penalty weight of zero (Table 12). The Schnakenberg results are provided in Appendix I.2 (Table 13).

Ablation of the curl-penalty weight on Fisher–KPP under velocity-data supervision with MCT (Appendix I.2; Table 12). Entries are mean ± SD over ten test trajectories with 500 rollout steps. NAsym is averaged over 101 evaluation times per trajectory.

λcurl\lambda_{\mathrm{curl}}ErollE_{\mathrm{roll}}Emax⁡E_{\max}NAsym
0(1.03±0.23)×10−3(1.03\pm0.23)\times10^{-3}(1.32±0.36)×10−3(1.32\pm0.36)\times10^{-3}(1.77±0.36)×10−3(1.77\pm0.36)\times10^{-3}
0.01(5.01±1.39)×10−4(5.01\pm1.39)\times10^{-4}(5.96±1.96)×10−4(5.96\pm1.96)\times10^{-4}(2.38±1.39)×10−4(2.38\pm1.39)\times10^{-4}
0.1(8.61±1.59)×10−4(8.61\pm1.59)\times10^{-4}(1.17±0.26)×10−3(1.17\pm0.26)\times10^{-3}(3.08±1.31)×10−4(3.08\pm1.31)\times10^{-4}
1(1.13±0.19)×10−3(1.13\pm0.19)\times10^{-3}(1.72±0.53)×10−3(1.72\pm0.53)\times10^{-3}(3.20±1.11)×10−3(3.20\pm1.11)\times10^{-3}

(Q1 cont.) Time-stepper. With network weights fixed, MCT, Forward Euler, and SSPRK2 yield similar mean ErollE_{\mathrm{roll}} and Emax⁡E_{\max} and preserve the ordering Law, Vel., then Direct for both species (Appendix I.1; Table 10). MCT explicitly evolves compression and material mass through ρ=MI\rho=MI, providing a common physical structure for incorporating richer constitutive laws and material states, and laying a foundation for more complex PDE systems (Sec. 4.2; Appendix J.5).

Constitutive parameterization, supervision, and integration on Schnakenberg (Appendix I.1; Table 10). Entries are mean ± sample SD over ten test trajectories; all rollouts use Δt=5×10−5\Delta t=5\times10^{-5}. For each constitutive model, the same trained weights are used across the three time-steppers.

Param.ObjectiveTime-stepperUU: ErollE_{\mathrm{roll}}UU: Emax⁡E_{\max}VV: ErollE_{\mathrm{roll}}VV: Emax⁡E_{\max}
u,r\mathbf{u},rLudata+Lrlaw\mathcal{L}^{\mathrm{data}}_{u}+\mathcal{L}_{r}^{\mathrm{law}}MCT(3.40±1.70)×10−2(3.40\pm1.70)\times10^{-2}(1.11±0.66)×10−1(1.11\pm0.66)\times10^{-1}(7.84±4.27)×10−3(7.84\pm4.27)\times10^{-3}(2.96±2.18)×10−2(2.96\pm2.18)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLvel\mathcal{L}^{\mathrm{vel}}MCT(2.52±1.77)×10−2(2.52\pm1.77)\times10^{-2}(8.28±8.55)×10−2(8.28\pm8.55)\times10^{-2}(6.74±3.37)×10−3(6.74\pm3.37)\times10^{-3}(2.19±1.93)×10−2(2.19\pm1.93)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLlaw\mathcal{L}^{\mathrm{law}}MCT(2.38±1.26)×10−2(2.38\pm1.26)\times10^{-2}(7.36±6.15)×10−2(7.36\pm6.15)\times10^{-2}(6.57±2.27)×10−3(6.57\pm2.27)\times10^{-3}(1.97±1.30)×10−2(1.97\pm1.30)\times10^{-2}
u,r\mathbf{u},rLudata+Lrlaw\mathcal{L}^{\mathrm{data}}_{u}+\mathcal{L}_{r}^{\mathrm{law}}Forward Euler(3.59±1.52)×10−2(3.59\pm1.52)\times10^{-2}(1.12±0.52)×10−1(1.12\pm0.52)\times10^{-1}(8.09±3.32)×10−3(8.09\pm3.32)\times10^{-3}(2.82±1.51)×10−2(2.82\pm1.51)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLvel\mathcal{L}^{\mathrm{vel}}Forward Euler(2.58±1.55)×10−2(2.58\pm1.55)\times10^{-2}(8.03±7.24)×10−2(8.03\pm7.24)\times10^{-2}(7.27±2.89)×10−3(7.27\pm2.89)\times10^{-3}(2.17±1.48)×10−2(2.17\pm1.48)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLlaw\mathcal{L}^{\mathrm{law}}Forward Euler(2.37±1.06)×10−2(2.37\pm1.06)\times10^{-2}(6.90±4.88)×10−2(6.90\pm4.88)\times10^{-2}(7.19±2.11)×10−3(7.19\pm2.11)\times10^{-3}(1.96±0.89)×10−2(1.96\pm0.89)\times10^{-2}
u,r\mathbf{u},rLudata+Lrlaw\mathcal{L}^{\mathrm{data}}_{u}+\mathcal{L}_{r}^{\mathrm{law}}SSPRK2(3.58±1.52)×10−2(3.58\pm1.52)\times10^{-2}(1.12±0.52)×10−1(1.12\pm0.52)\times10^{-1}(8.08±3.33)×10−3(8.08\pm3.33)\times10^{-3}(2.81±1.52)×10−2(2.81\pm1.52)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLvel\mathcal{L}^{\mathrm{vel}}SSPRK2(2.58±1.55)×10−2(2.58\pm1.55)\times10^{-2}(8.04±7.26)×10−2(8.04\pm7.26)\times10^{-2}(7.25±2.89)×10−3(7.25\pm2.89)\times10^{-3}(2.17±1.49)×10−2(2.17\pm1.49)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLlaw\mathcal{L}^{\mathrm{law}}SSPRK2(2.36±1.06)×10−2(2.36\pm1.06)\times10^{-2}(6.90±4.89)×10−2(6.90\pm4.89)\times10^{-2}(7.16±2.11)×10−3(7.16\pm2.11)\times10^{-3}(1.95±0.90)×10−2(1.95\pm0.90)\times10^{-2}

Choice of transport network. CNO achieves lower mean ErollE_{\mathrm{roll}} and Emax⁡E_{\max} than the fully connected DeepONet on all seven PDEs under both velocity-data and known-law supervision. The DeepONet-to-CNO mean ErollE_{\mathrm{roll}} ratio ranges from approximately 1.6 to 214 under the same training protocol and MCT rollout settings (Appendix I.4; Table 15).

I/M representation. The separately evolved factors in ρ=MI\rho=MI retain supplied material-state information. Matched pairs with identical initial total density but different compositions reach a median final reference-density separation of 2.43%, illustrating composition-dependent dynamics (Secs. 4.2, 5.1; Appendix J.5).

Q2. Can broader distribution-shift experiments strengthen the evidence for robustness and generalization?

A2. Thank you. We added tests of spatial, temporal, and composition changes. The results support broader generalization through accurate reuse of learned constitutive modules under these shifts without further adaptation.

Unseen initial conditions. We reuse the original models on ten unseen initial conditions per system across five systems, including blob, ring, and spot-lattice families. Both NCL-MCT Solver configurations achieve lower mean ErollE_{\mathrm{roll}} than the evaluated baselines (Appendix G.1; Tables 6–7).

Rollout errors from unseen initial conditions. Entries are mean ± sample SD over ten trajectories; — denotes not evaluated.

SystemMetricF-FNOCFODOOLNCL-MCT Solver (Vel.)NCL-MCT Solver (Law)
Linear diffusionErollE_{\mathrm{roll}}(1.79±0.11)×10−1(1.79\pm0.11)\times10^{-1}(1.62±0.15)×10−1(1.62\pm0.15)\times10^{-1}(5.88±0.38)×10−1(5.88\pm0.38)\times10^{-1}(4.26±0.44)×10−2(4.26\pm0.44)\times10^{-2}(4.04±0.40)×10−2(4.04\pm0.40)\times10^{-2}
Emax⁡E_{\max}(2.82±0.18)×10−1(2.82\pm0.18)\times10^{-1}(2.09±0.22)×10−1(2.09\pm0.22)\times10^{-1}(7.27±0.53)×10−1(7.27\pm0.53)\times10^{-1}(4.94±0.51)×10−2(4.94\pm0.51)\times10^{-2}(4.73±0.45)×10−2(4.73\pm0.45)\times10^{-2}
CHErollE_{\mathrm{roll}}(2.15±0.09)×10−1(2.15\pm0.09)\times10^{-1}*(2.71±0.16)×10−1(2.71\pm0.16)\times10^{-1}(4.36±0.20)×10−1(4.36\pm0.20)\times10^{-1}(1.59±0.18)×10−2(1.59\pm0.18)\times10^{-2}(1.93±0.12)×10−2(1.93\pm0.12)\times10^{-2}
Emax⁡E_{\max}(3.13±0.13)×10−1(3.13\pm0.13)\times10^{-1}*(3.84±0.21)×10−1(3.84\pm0.21)\times10^{-1}(5.96±0.30)×10−1(5.96\pm0.30)\times10^{-1}(3.05±0.45)×10−2(3.05\pm0.45)\times10^{-2}(3.86±0.61)×10−2(3.86\pm0.61)\times10^{-2}
Fisher–KPPErollE_{\mathrm{roll}}(3.56±1.99)×10−1(3.56\pm1.99)\times10^{-1}*(1.23±0.14)×10−1(1.23\pm0.14)\times10^{-1}—(5.28±0.50)×10−2(5.28\pm0.50)\times10^{-2}(4.93±0.50)×10−2(4.93\pm0.50)\times10^{-2}
Emax⁡E_{\max}(4.31±2.56)×10−1(4.31\pm2.56)\times10^{-1}*(1.54±0.20)×10−1(1.54\pm0.20)\times10^{-1}—(6.44±0.67)×10−2(6.44\pm0.67)\times10^{-2}(6.00±0.66)×10−2(6.00\pm0.66)\times10^{-2}
Reactive CHErollE_{\mathrm{roll}}(2.53±0.11)×10−1(2.53\pm0.11)\times10^{-1}(2.83±0.26)×10−1(2.83\pm0.26)\times10^{-1}—(2.94±0.80)×10−2(2.94\pm0.80)\times10^{-2}(2.48±0.39)×10−2(2.48\pm0.39)\times10^{-2}
Emax⁡E_{\max}(3.73±0.19)×10−1(3.73\pm0.19)\times10^{-1}(4.18±0.44)×10−1(4.18\pm0.44)\times10^{-1}—(6.72±2.80)×10−2(6.72\pm2.80)\times10^{-2}(5.28±1.30)×10−2(5.28\pm1.30)\times10^{-2}
Schnakenberg (UU)ErollE_{\mathrm{roll}}(2.74±1.10)×10−1(2.74\pm1.10)\times10^{-1}(1.20±0.30)×10−1(1.20\pm0.30)\times10^{-1}—(1.08±0.65)×10−2(1.08\pm0.65)\times10^{-2}(1.42±0.58)×10−2(1.42\pm0.58)\times10^{-2}
Emax⁡E_{\max}(4.84±2.17)×10−1(4.84\pm2.17)\times10^{-1}(2.87±0.53)×10−1(2.87\pm0.53)\times10^{-1}—(2.39±1.09)×10−2(2.39\pm1.09)\times10^{-2}(3.76±1.88)×10−2(3.76\pm1.88)\times10^{-2}
Schnakenberg (VV)ErollE_{\mathrm{roll}}(8.57±3.06)×10−2(8.57\pm3.06)\times10^{-2}(3.95±0.98)×10−2(3.95\pm0.98)\times10^{-2}—(3.14±1.54)×10−3(3.14\pm1.54)\times10^{-3}(4.72±1.64)×10−3(4.72\pm1.64)\times10^{-3}
Emax⁡E_{\max}(1.53±0.60)×10−1(1.53\pm0.60)\times10^{-1}(9.47±2.12)×10−2(9.47\pm2.12)\times10^{-2}—(7.18±3.01)×10−3(7.18\pm3.01)\times10^{-3}(1.27±0.53)×10−2(1.27\pm0.53)\times10^{-2}

*F-FNO results use finite trajectories only: CH 9/10; Fisher–KPP 8/10.

Extended time horizon. Models trained on the first 60% of each reference interval are evaluated over the full interval with fixed weights. Across four scalar systems, their full-window ErollE_{\mathrm{roll}} values are 0.90–1.46 times those of their full-window counterparts, evaluated over the same interval (Appendix G.2; Tables 8–9).

Composition and module reuse. The frozen velocity CNO and pointwise growth MLP trained on single-population donor data are reused in a two-population porous-medium system. MCT completes all 128 trajectories across the Basic, Composition, Spatial OOD, and Matched pairs splits without retraining. Mean total-density ErollE_{\mathrm{roll}} ranges from 9.42×10−49.42\times10^{-4} to 2.08×10−32.08\times10^{-3} across splits, supporting accurate constitutive reuse (Appendix J.5; Table 19, Figure 26).

Reviewer_Ds7EOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The central idea is clean: rather than learning the whole solution map or simply learning the PDE derivative, the paper learns the physical constitutive responses.

Critical strengths

The paper learns PDE-specific constitutive responses while retaining the shared physical evolution structure in the MCT integrator. This is a meaningful alternative to directly learning solution maps or temporal derivatives.

The method performs well on unseen initial-condition families and extended time horizons without retraining, with substantially lower errors than the evaluated baselines in these tests.

Critical weaknesses

The comparison does not fully isolate the benefit of constitutive learning. it is difficult to determine whether the reported generalization advantage comes from the proposed constitutive-learning target, the stronger physics supervision, or the MCT integrator itself.

Constitutive-law identifiability is not fully resolved because velocity-data supervision identifies only the product u=ξfu=\xi f, while the learned ξ\xi and ff may not correspond to uniquely defined physical constitutive laws.

Questions for authors

To what extent are the improved generalization and long-horizon stability attributable specifically to learning constitutive responses, rather than to the stronger physics prior and the particular MCT numerical integrator? Could the authors isolate these effects more directly?

Under velocity-data supervision, ξ\xi and ff are non-identifiable, and curl regularization does not guarantee an underlying energy functional. In what precise sense should the separately predicted ξ\xi and ff be interpreted as learned physical constitutive laws rather than an arbitrary factorization of the correct velocity?

3 comments

We thank the reviewers for their constructive comments and recognition of our framework’s physical structure, accuracy, and generalization. The seven systems share a constitutive interface and time-stepper, while each system trains its own response modules. Revisions are marked in blue. The main changes are:

Summary of our response

  1. Contributions and ablations (u59j, Ds7E, qY9r, Vcks). We distinguish Vel. supervision of the mobility–force product from Law supervision of the individual responses under a specified physical convention. We add time-stepper comparisons using identical network weights, CNO–DeepONet comparisons, and curl-penalty weight ablations, distinguishing within-checkpoint comparisons from variation across training runs (Sec. 4.3; Appendices I.1, I.2, I.4).

  2. I/M material state (qY9r). We clarify how the I/M decomposition retains supplied material composition information, enabling composition-dependent dynamics. In a two-population example, we show how common compression cancels in the material derivative of the population ratio and how both MCT and direct population evolution reproduce the separation of states with equal total density (Secs. 4.2, 5.1; Appendix J.5).

  3. Constitutive identifiability (Ds7E). We show that distinct physically valid energy–dissipation laws can produce the same mobility–force product and density evolution. Thus, non-uniqueness under velocity supervision is compatible with the underlying physics (Secs. 3, 4.3; Appendix A.4).

  4. Generalization and module reuse (u59j, qY9r). We distinguish checkpoint reuse for new initial conditions from evaluation beyond a shortened training window, and clarify Schnakenberg’s state-range screening. We also reuse single-population velocity and growth networks in a two-population system without retraining (Appendices G, J.5).

  5. Conditional error analysis and statistics (u59j, Vcks). We establish a conditional finite-horizon bound between MCT rollouts with learned and numerically evaluated constitutive responses at fixed discretization. We also add sampled one-step diagnostics across seven PDEs and 95% confidence intervals for mean rollout errors over ten Schnakenberg trajectories (Appendices A.5, I.1, J.2).

  6. Training, runtime, and numerical settings (u59j, Vcks). We add training-loss curves, rollout and per-step timing breakdowns, and per-PDE factor-reset intervals. For Schnakenberg, we compare native and learned updates on the same GPU, using the same precision and batch size, including a comparison at the same time step (Appendices C, J.1).

Response to Reviewer Ds7E

Q1. What are the respective contributions of constitutive learning, stronger physical supervision, and the MCT time-stepper to improved generalization and long-time stability, and can these effects be isolated through controlled comparisons?

A1. Thank you. We added controlled ablations to distinguish constitutive parameterization, supervision, regularization, and integration.

Constitutive parameterization and supervision. With a curl-penalty weight of zero, factorization reduces mean ErollE_{\mathrm{roll}} for UU from 3.40×10−23.40\times10^{-2} to 2.44×10−22.44\times10^{-2} under the same velocity supervision. This demonstrates an accuracy gain without curl penalty or additional factor-level supervision. Law instead supervises mobility and force separately (Sec. 4.3; Appendices I.1–I.2; Tables 10, 13).

Curl penalty. We include the curl penalty to suppress spurious local rotation in the learned driving force and encourage its curl-free physical structure, f=−∇(δE/δρ)\mathbf{f}=-\nabla(\delta E/\delta\rho). On Fisher–KPP, λcurl=0.01\lambda_{\mathrm{curl}}=0.01 gives the lowest mean ErollE_{\mathrm{roll}}, Emax⁡E_{\max}, and NAsym among the tested weights, reducing mean NAsym by approximately 87% relative to a curl-penalty weight of zero (Table 12). The Schnakenberg results are provided in Appendix I.2 (Table 13).

Ablation of the curl-penalty weight on Fisher–KPP under velocity-data supervision with MCT (Appendix I.2; Table 12). Entries are mean ± SD over ten test trajectories with 500 rollout steps. NAsym is averaged over 101 evaluation times per trajectory.

λcurl\lambda_{\mathrm{curl}}ErollE_{\mathrm{roll}}Emax⁡E_{\max}NAsym
0(1.03±0.23)×10−3(1.03\pm0.23)\times10^{-3}(1.32±0.36)×10−3(1.32\pm0.36)\times10^{-3}(1.77±0.36)×10−3(1.77\pm0.36)\times10^{-3}
0.01(5.01±1.39)×10−4(5.01\pm1.39)\times10^{-4}(5.96±1.96)×10−4(5.96\pm1.96)\times10^{-4}(2.38±1.39)×10−4(2.38\pm1.39)\times10^{-4}
0.1(8.61±1.59)×10−4(8.61\pm1.59)\times10^{-4}(1.17±0.26)×10−3(1.17\pm0.26)\times10^{-3}(3.08±1.31)×10−4(3.08\pm1.31)\times10^{-4}
1(1.13±0.19)×10−3(1.13\pm0.19)\times10^{-3}(1.72±0.53)×10−3(1.72\pm0.53)\times10^{-3}(3.20±1.11)×10−3(3.20\pm1.11)\times10^{-3}

Time-stepper. With network weights fixed, MCT, Forward Euler, and SSPRK2 yield similar mean ErollE_{\mathrm{roll}} and Emax⁡E_{\max} and preserve the ordering Law, Vel., then Direct for both species (Appendix I.1; Table 10). MCT explicitly evolves compression and material mass through ρ=MI\rho=MI, providing a common physical structure for incorporating richer constitutive laws and material states, and laying a foundation for more complex PDE systems (Sec. 4.2; Appendix J.5).

Constitutive parameterization, supervision, and integration on Schnakenberg (Appendix I.1; Table 10). Entries are mean ± sample SD over ten test trajectories; all rollouts use Δt=5×10−5\Delta t=5\times10^{-5}. For each constitutive model, the same trained weights are used across the three time-steppers.

Param.ObjectiveTime-stepperUU: ErollE_{\mathrm{roll}}UU: Emax⁡E_{\max}VV: ErollE_{\mathrm{roll}}VV: Emax⁡E_{\max}
u,r\mathbf{u},rLudata+Lrlaw\mathcal{L}^{\mathrm{data}}_{u}+\mathcal{L}_{r}^{\mathrm{law}}MCT(3.40±1.70)×10−2(3.40\pm1.70)\times10^{-2}(1.11±0.66)×10−1(1.11\pm0.66)\times10^{-1}(7.84±4.27)×10−3(7.84\pm4.27)\times10^{-3}(2.96±2.18)×10−2(2.96\pm2.18)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLvel\mathcal{L}^{\mathrm{vel}}MCT(2.52±1.77)×10−2(2.52\pm1.77)\times10^{-2}(8.28±8.55)×10−2(8.28\pm8.55)\times10^{-2}(6.74±3.37)×10−3(6.74\pm3.37)\times10^{-3}(2.19±1.93)×10−2(2.19\pm1.93)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLlaw\mathcal{L}^{\mathrm{law}}MCT(2.38±1.26)×10−2(2.38\pm1.26)\times10^{-2}(7.36±6.15)×10−2(7.36\pm6.15)\times10^{-2}(6.57±2.27)×10−3(6.57\pm2.27)\times10^{-3}(1.97±1.30)×10−2(1.97\pm1.30)\times10^{-2}
u,r\mathbf{u},rLudata+Lrlaw\mathcal{L}_{\mathbf{u}}^{\mathrm{data}}+\mathcal{L}_{r}^{\mathrm{law}}Forward Euler(3.59±1.52)×10−2(3.59\pm1.52)\times10^{-2}(1.12±0.52)×10−1(1.12\pm0.52)\times10^{-1}(8.09±3.32)×10−3(8.09\pm3.32)\times10^{-3}(2.82±1.51)×10−2(2.82\pm1.51)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLvel\mathcal{L}^{\mathrm{vel}}Forward Euler(2.58±1.55)×10−2(2.58\pm1.55)\times10^{-2}(8.03±7.24)×10−2(8.03\pm7.24)\times10^{-2}(7.27±2.89)×10−3(7.27\pm2.89)\times10^{-3}(2.17±1.48)×10−2(2.17\pm1.48)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLlaw\mathcal{L}^{\mathrm{law}}Forward Euler(2.37±1.06)×10−2(2.37\pm1.06)\times10^{-2}(6.90±4.88)×10−2(6.90\pm4.88)\times10^{-2}(7.19±2.11)×10−3(7.19\pm2.11)\times10^{-3}(1.96±0.89)×10−2(1.96\pm0.89)\times10^{-2}
u,r\mathbf{u},rLudata+Lrlaw\mathcal{L}_{\mathrm{u}}^{\mathrm{data}}+\mathcal{L}_{r}^{\mathrm{law}}SSPRK2(3.58±1.52)×10−2(3.58\pm1.52)\times10^{-2}(1.12±0.52)×10−1(1.12\pm0.52)\times10^{-1}(8.08±3.33)×10−3(8.08\pm3.33)\times10^{-3}(2.81±1.52)×10−2(2.81\pm1.52)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLvel\mathcal{L}^{\mathrm{vel}}SSPRK2(2.58±1.55)×10−2(2.58\pm1.55)\times10^{-2}(8.04±7.26)×10−2(8.04\pm7.26)\times10^{-2}(7.25±2.89)×10−3(7.25\pm2.89)\times10^{-3}(2.17±1.49)×10−2(2.17\pm1.49)\times10^{-2}
(ξ,f),r(\xi,\mathbf{f}),rLlaw\mathcal{L}^{\mathrm{law}}SSPRK2(2.36±1.06)×10−2(2.36\pm1.06)\times10^{-2}(6.90±4.89)×10−2(6.90\pm4.89)\times10^{-2}(7.16±2.11)×10−3(7.16\pm2.11)\times10^{-3}(1.95±0.90)×10−2(1.95\pm0.90)\times10^{-2}

Q2. Given that velocity supervision constrains only the product u=ξf\mathbf{u}=\xi\mathbf{f} and the curl penalty does not guarantee an energy functional, in what sense do the non-unique factors represent learned physical constitutive laws?

A2. Thank you. We have added a proof that distinct physically valid energy–dissipation laws can yield the same product u=ξf\mathbf{u}=\xi\mathbf{f} and density evolution. For example, for smooth positive solutions of ∂tρ=Δ(ρ2)\partial_t\rho=\Delta(\rho^2) on a periodic domain, energy densities ρ2\rho^2 and ρlog⁡ρ\rho\log\rho with mobilities 11 and 2ρ2\rho both give u=−2∇ρ\mathbf{u}=-2\nabla\rho, while each satisfies its own continuous energy-dissipation law. Thus, non-uniqueness under velocity supervision is compatible with the underlying physics (Appendix A.4).

Our core idea is to learn PDE-specific constitutive responses while preserving shared physical evolution in MCT. Factorization supports positive mobility and force-geometry regularization; velocity supervision constrains the product, while known-law supervision anchors individual factors (Secs. 3, 4.3).

Reviewer_u59jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NCL-MCT learns PDE-specific transport mobility, driving force, and relative reaction rates and supplies them to a shared mass-compression-transport integrator. Each PDE has separately trained modules. The paper evaluates seven systems, changed initial conditions, a modest time-horizon extension, and known-law training on independently sampled density fields. The supplementary material specifies the numerical updates, models and training protocol, reports computational costs and mass drift, and adds direct-velocity, numerical/neural modularity, and curl-weight controls.

Critical strengths
  1. The same interface and integrator accommodate seven systems, and the reported complete-solver gains on CH-type systems and Schnakenberg are substantial. Across the reported implementations in Table 10, NCL-MCT uses fewer parameters, less training memory, and less training time than F-FNO and CFO. Its inference is faster than CFO's but slower than F-FNO's. This strengthens the practical case beyond error values alone.
  2. Tables 12–14 compare direct velocity prediction, numerical and learned transport/reaction responses within the same MCT integrator, and different curl-penalty weights. These controls support the design choices and the shared interface. Numerical constitutive responses provide an appropriate reference for assessing the approximation error introduced by learned modules.
Critical weaknesses
  1. The parameterization evidence is positive but still limited. Table 12 shows lower mean errors for the factorized models than direct velocity prediction on Schnakenberg. However, the direct-velocity versus Vel. comparison also changes curl regularization, as Appendix I.4 acknowledges. The curl-weight study is on Fisher-KPP, so it does not separate the effects within the Schnakenberg comparison. The reported standard deviations across test trajectories do not quantify variation across training runs or uncertainty in the paired improvement. I would place more weight on the methodological advance if the factorization's benefit beyond regularization were established more directly.
  2. Reuse under changed initial conditions is supported by five individual examples rather than repeated tests within each changed family. Appendix G.1 further accepts the Schnakenberg example only when the reference trajectory's UU, VV, and U2VU^2V remain within the training-snapshot ranges throughout the rollout. This criterion appears designed to isolate spatial distribution shifts rather than extrapolation in state magnitude. It does not by itself demonstrate selection based on model performance or training-data leakage, but it restricts the evaluated trajectories to a relatively controlled regime. The ten-trajectory in-distribution and time-extension results remain useful but do not resolve this specific coverage issue.
Questions for authors
  1. Are paired per-trajectory differences available for Table 12? A factorized model without the curl penalty in the same Schnakenberg setting would help separate the effects of parameterization and regularization.
  2. What fraction of Schnakenberg candidate initial conditions met the reference-trajectory range criterion? A few additional, predefined initial conditions from the same changed families would help establish how representative the examples are.
  3. What TreinitT_{\mathrm{reinit}} was used? Does Appendix G's opening no-retraining statement refer only to G.1, or to the absence of further adaptation during evaluation after protocol-specific training? Section 5.2 specifies checkpoint reuse for unseen initial conditions, while G.2 describes training on the first 60% of each interval for time-horizon extension.
2 comments

We thank the reviewers for their constructive comments and recognition of our framework’s physical structure, accuracy, and generalization. The seven systems share a constitutive interface and time-stepper, while each system trains its own response modules. Revisions are marked in blue. The main changes are:

Summary of our response

  1. Contributions and ablations (u59j, Ds7E, qY9r, Vcks). We distinguish Vel. supervision of the mobility–force product from Law supervision of the individual responses under a specified physical convention. We add time-stepper comparisons using identical network weights, CNO–DeepONet comparisons, and curl-penalty weight ablations, distinguishing within-checkpoint comparisons from variation across training runs (Sec. 4.3; Appendices I.1, I.2, I.4).

  2. I/M material state (qY9r). We clarify how the I/M decomposition retains supplied material composition information, enabling composition-dependent dynamics. In a two-population example, we show how common compression cancels in the material derivative of the population ratio and how both MCT and direct population evolution reproduce the separation of states with equal total density (Secs. 4.2, 5.1; Appendix J.5).

  3. Constitutive identifiability (Ds7E). We show that distinct physically valid energy–dissipation laws can produce the same mobility–force product and density evolution. Thus, non-uniqueness under velocity supervision is compatible with the underlying physics (Secs. 3, 4.3; Appendix A.4).

  4. Generalization and module reuse (u59j, qY9r). We distinguish checkpoint reuse for new initial conditions from evaluation beyond a shortened training window, and clarify Schnakenberg’s state-range screening. We also reuse single-population velocity and growth networks in a two-population system without retraining (Appendices G, J.5).

  5. Conditional error analysis and statistics (u59j, Vcks). We establish a conditional finite-horizon bound between MCT rollouts with learned and numerically evaluated constitutive responses at fixed discretization. We also add sampled one-step diagnostics across seven PDEs and 95% confidence intervals for mean rollout errors over ten Schnakenberg trajectories (Appendices A.5, I.1, J.2).

  6. Training, runtime, and numerical settings (u59j, Vcks). We add training-loss curves, rollout and per-step timing breakdowns, and per-PDE factor-reset intervals. For Schnakenberg, we compare native and learned updates on the same GPU, using the same precision and batch size, including a comparison at the same time step (Appendices C, J.1).

Response to Reviewer u59j

Q1. A factorized model without the curl penalty in the same Schnakenberg setting would help separate the effects of parameterization and regularization.

A1. Thank you. We include the curl penalty to suppress spurious local rotation in the learned driving force and encourage its curl-free physical structure, f=−∇(δE/δρ)\mathbf{f}=-\nabla(\delta E/\delta\rho). On Fisher–KPP, λcurl=0.01\lambda_{\mathrm{curl}}=0.01 gives the lowest mean ErollE_{\mathrm{roll}}, Emax⁡E_{\max}, and NAsym among the tested weights, reducing mean NAsym by approximately 87% relative to a curl-penalty weight of zero (Table 12).

Ablation of the curl-penalty weight on Fisher–KPP under velocity-data supervision with MCT (Appendix I.2; Table 12). Entries are mean ± SD over ten test trajectories with 500 rollout steps. NAsym is averaged over 101 evaluation times per trajectory.

λcurl\lambda_{\mathrm{curl}}ErollE_{\mathrm{roll}}Emax⁡E_{\max}NAsym
0(1.03±0.23)×10−3(1.03\pm0.23)\times10^{-3}(1.32±0.36)×10−3(1.32\pm0.36)\times10^{-3}(1.77±0.36)×10−3(1.77\pm0.36)\times10^{-3}
0.01(5.01±1.39)×10−4(5.01\pm1.39)\times10^{-4}(5.96±1.96)×10−4(5.96\pm1.96)\times10^{-4}(2.38±1.39)×10−4(2.38\pm1.39)\times10^{-4}
0.1(8.61±1.59)×10−4(8.61\pm1.59)\times10^{-4}(1.17±0.26)×10−3(1.17\pm0.26)\times10^{-3}(3.08±1.31)×10−4(3.08\pm1.31)\times10^{-4}
1(1.13±0.19)×10−3(1.13\pm0.19)\times10^{-3}(1.72±0.53)×10−3(1.72\pm0.53)\times10^{-3}(3.20±1.11)×10−3(3.20\pm1.11)\times10^{-3}

NAsym. Following the normalized Jacobian-asymmetry diagnostic in [1], we measure the skew-symmetric share of the spatial force Jacobian. NAsym lies in [0,1][0,1]; zero indicates a symmetric spatial Jacobian, while larger values indicate greater local rotation.

Schnakenberg control. We also added the λcurl=0\lambda_{\mathrm{curl}}=0 control on the same Schnakenberg training dataset. It achieves lower mean ErollE_{\mathrm{roll}} than Direct for both species (Appendices I.1–I.2; Tables 10, 13).

[1] Chao et al. On Investigating the Conservative Property of Score-Based Generative Models. ICML 2023. arXiv:2209.12753.

Q2. A few additional, predefined initial conditions from the same changed families would help establish how representative the examples are.

A2. Thank you for the suggestion. We expanded the unseen-initial-condition evaluation to ten initial conditions per system across five systems, reusing the original checkpoints without retraining or retuning. The test dataset covers blob, ring, and spot-lattice families with varying spatial parameters. Both NCL-MCT Solver configurations achieve lower mean ErollE_{\mathrm{roll}} and Emax⁡E_{\max} than the evaluated baselines on all five systems (Appendix G.1; Tables 6–7).

Q3. What fraction of Schnakenberg candidate initial conditions met the reference-trajectory range criterion?

A3. Thank you. All ten Schnakenberg test cases in Appendix G.1 satisfy the reference-trajectory range criterion for UU, VV, and U2VU^2V. The original candidate acceptance fraction was not recorded.

Q4. Are paired per-trajectory differences available for the original Table 12 (Table 10 in the revision)?

A4. Thank you. Across ten paired test trajectories, mean ErollE_{\mathrm{roll}} differences are (8.80 1.10)×10−3(8.80\,1.10)\times10^{-3} for Direct−Vel. and (1.49 0.17)×10−3(1.49\,0.17)\times10^{-3} for Vel.−Law, in (U,V)(U,V) order. Vel./Law use submitted checkpoints; the retrained Direct model reproduces the original mean errors (Appendix I.1).

Q5. How does “no retraining” apply to checkpoint reuse in G.1 and truncated-horizon training in G.2?

A5. Thank you. G.1 reuses the original checkpoints without retraining. G.2 trains separate models on the first 60% of each reference interval to test extrapolation beyond a shorter training window, then evaluates the full interval with fixed weights (Appendix G).

Q6. What TreinitT_{\mathrm{reinit}} was used?

A6. Thank you. Vel. and Law share the following reinitialization settings. Resetting I=1I=1 and M=ρM=\rho limits factor distortion while preserving the reconstructed density ρ\rho (Appendix C).

SystemTreinitT_{\mathrm{reinit}}Steps per reinitialization
Linear diffusion0.01200
CH0.001100
Porous medium0.00550
Fisher–KPP0.01334
Reactive CH0.001100
Reactive porous medium0.0120
Schnakenberg0.005100

A reset occurs after ⌈Treinit/Δt⌉\lceil T_{\mathrm{reinit}}/\Delta t\rceil steps.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.