acceptodds
Under review as a conference paper at ICLR 2027

PINNMorph: Evolving Online Adaptation Policies for Physics-Informed Neural Networks

Abstract

Physics-informed neural networks (PINNs) provide a learning-based framework for solving partial differential equations (PDEs), yet their training behavior can change substantially throughout optimization. Residual distributions, gradient interactions, regional learning difficulty, and model-capacity requirements may evolve over time, while the network architecture and major training mechanisms are typically determined before training. We propose PINNMorph, an online PINN adaptation framework based on LLM-guided policy evolution. PINNMorph maintains a population of state-conditioned adaptation policies that map execution diagnostics to controlled interventions over topology modification, additive representation augmentation, objective balancing, gradient handling, adaptive sampling, and optimizer-phase control. At each intervention opportunity, candidate programs are instantiated from the current policy population, selected according to the observed training state, and applied directly to the PINN under training. The resulting model inherits its existing parameters and training state and continues optimization along the same trajectory. Execution outcomes are subsequently used to evaluate interventions and evolve the policy population. Unlike pre-training architecture search or fixed adaptation rules, PINNMorph jointly adapts the current PINN and the policies governing its interventions using feedback from actual training. Experiments on 13 PDE benchmarks show that PINNMorph achieves lower solution errors than SA-PINN, ConFIG, RoPINN, HARMONIC, and PINNsAgent across all evaluated problems. Ablation studies further examine the effects of online adaptation, state-conditioned intervention selection, and execution-feedback-driven policy evolution. Anonymous Code Repository: https://anonymous.4open.science/r/PINNMorph/

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Neural Networks, Online Structural Adaptation, Policy Evolution, Large Language Models, Execution Feedback, Partial Differential Equations, Scientific Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Qxt3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PINNs are difficult to train because of imbalanced objective functions and the the authors argue that the dominant bottleneck shifts during training. They propose a framework that periodically diagnoses the training state and use an LLM to generate policies that adapt topology changes in the PDE, add representation branches, reweigh the loss terms or make changes to the optimizer and apply to the model without restarting training. Each policy change is evaluated against an estimated no-intervention trajectory, and this feedback is used to evolve the policies. On the benchmarks the method reports the lowest MSE against five baselines, with ablations supporting each component and a conservative increase in wall-clock overhead.

Critical strengths
  • The authors brought to attention an important bottleneck in PINN training as the training shifts during the runs and existing adaptive methods fix each problem in advance as opposed to during runtime.

  • The paper looks at developing adaptive decisions across several different problem settings and shows improvements without restarting methods, which is an advantage over pre-trained search methods

  • The LLM generation was surprisingly low overhead in terms of runtime, making it useful for online training

Critical weaknesses

None

Questions for authors

Several reported improvements fall within one standard deviation of the best baseline (e.g., KS 1D, Burgers 1D, Poisson 5D). Could the authors report paired significance tests per PDE (e.g., Wilcoxon or t-test across seeds), and consider more than five seeds, to support the claim of best performance on all 13 benchmarks?

All benchmarks are relatively small. How do the costs of diagnosis, auditing (residuals on held-out sets), per-loss gradient statistics, and in-place structural changes scale with problem dimension, collocation set size, and model size? A cost breakdown, or results on a larger 3D time-dependent problem, would help. Do the LLM-generated policies transfer across problems, or must the search restart for every new PDE?

The No Morph base PINN performs far worse than the baselines on several tasks. Do all methods share the same architecture and tuning budget? Since interventions add parameters, please report final parameter counts compared with the baselines.

If the LLM-generated policy changes the objective weighting or sampling, the measured quantity changes too. How do the authors keep the utility comparable across interventions? Can they validate the no-intervention estimate by continuing training without intervention from the same checkpoint, and show that evolved policies beat the initial ones from the same training state?

1 comment

We thank the reviewer for the positive feedback and helpful questions. We address each point below.

1. Paired significance across seeds

With 10 matched seeds, PINNMorph had the lowest mean MSE on all 13 PDEs. Two-sided paired t-tests (9 df) were significant at 0.05 in 59/65 comparisons and 8/13 against the lowest-mean baseline. After Holm correction across all 65, the counts were 52/65 and 4/13 (NS 2D CG, Heat 2D MS, Poisson 2D CG, Gray–Scott 2D). The three cases you flagged were not significant after correction (KS 1D and Burgers 1D, p = 0.61447 each, and Poisson 5D, p = 0.07447).

Table R1. Lowest-mean baseline, first four PDEs (mean ± sample SD). All 13 in Appendix E.2, Table E.3.

PDEBest baseline MSEPINNMorph MSEReductiont(9)pHolm p
KS 1DPINNsAgent: 1.179 ± 0.1001.049 ± 0.23310.99%1.5530.154870.61447
Burgers 1DPINNsAgent: 7.14E-05 ± 1.71E-056.49E-05 ± 1.13E-059.09%1.0480.322070.61447
NS 2D CGPINNsAgent: 4.34E-03 ± 9.79E-041.56E-03 ± 5.50E-0464.16%8.3441.58E-054.58E-04
Heat 2D CGSA-PINN: 3.42E-03 ± 5.28E-052.54E-03 ± 2.21E-0325.88%1.2760.233880.61447

Table R2. Paired t / Holm p, first four PDEs. Positive t favors PINNMorph. All 65 comparisons in Appendix E.2, Table E.4.

PDESA-PINNConFIGRoPINNHARMONICPINNsAgent
KS 1D66.72 / 1.19E-1197.28 / 4.16E-1326.56 / 3.53E-0831.52 / 8.32E-091.55 / 0.61447
Burgers 1D36.12 / 2.59E-094.82 / 0.0170629.17 / 1.60E-087.25 / 0.001211.05 / 0.61447
NS 2D CG29.23 / 1.60E-089.65 / 1.53E-045.33 / 0.0104926.82 / 3.30E-088.34 / 4.58E-04
Heat 2D CG1.28 / 0.6144710.00 / 1.21E-043.33 / 0.0960436.14 / 2.59E-091.77 / 0.55688

2. Scaling and policy transfer

On two 10D PDEs we varied width (128/256) and interior points (32,768/65,536). All six configurations improved mean MSE over No Morph, with 2.7–32.1% extra runtime. Decision and Audit sets are capped at 4,096 and 32,768 points (B.3). Wider models require more work per derivative. Timings cover training, diagnostics, interventions and LLM calls.

Table R3. Five-run means, matched starting networks, point budgets and 20,000 steps per configuration.

PDEWidthInterior pointsMSE (No Morph / PINNMorph)Time (s, same order)Increase
Heat 10D12832,7686.18E-5 / 3.24E-71,662 / 2,19632.1%
Heat 10D25632,7686.14E-7 / 1.90E-72,892 / 3,19210.4%
Heat 10D25665,5361.54E-4 / 2.35E-75,544 / 5,7062.9%
Poisson 10D12832,7684.87E-4 / 6.97E-71,320 / 1,48812.7%
Poisson 10D25632,7681.00E-6 / 2.86E-72,520 / 2,92216.0%
Poisson 10D25665,5362.03E-5 / 2.51E-75,148 / 5,2862.7%

3. No Morph, tuning budget, and model size

No Morph uses PINNMorph’s initial network and data with adaptation disabled. The fixed baselines share its 4×128 Tanh backbone and add weighting, gradient or regional updates. PINNsAgent searches its architecture and optimizer. Each run has 20,000 steps. Total tuning costs can differ. The five baselines select by MSE on the reporting points (oracle). PINNMorph selects by its reference-free audit score (B.6).

Table R4. Final parameter counts (five-run means), first four PDEs. All 13 in Appendix E.1, Table E.2. No Morph matches “Fixed baselines” in Table E.2.

PDESA-PINNConFIGRoPINNHARMONICPINNsAgentNo MorphPINNMorph
KS 1D50,04950,04950,04950,04981,45150,04953,132
Burgers 1D50,04950,04950,04950,04956,09350,04954,467
NS 2D CG50,30750,30750,30750,30751,88150,30751,615
Heat 2D CG50,17750,17750,17750,17759,58450,17750,692

4. Utility and matched checkpoints

Changing training weights or samples leaves the Decision and Audit points fixed. Utility compares physical residuals with frozen pre-action forecasts. Active groups have equal weight, independent of training-loss weights. Mismatched components or regions withhold credit.

From each checkpoint we ran three branches for 500 steps. N continued without intervention. I and E took one action from initial and evolved policies, respectively. Forecast errors compare predictions with observed N-branch residuals.

Table R5. Solution MSE and relative forecast errors at 500 steps.

PDEStepN MSEI MSEE MSEDecision errorAudit error
Burgers 1D10,0002.51E-042.13E-041.08E-040.0860.097
Heat 2D MS7,6001.23E-039.31E-045.32E-040.1130.135
NS 2D CG12,0003.51E-033.06E-032.07E-030.0910.103
Poisson 5D9,0002.14E-061.75E-061.12E-060.0710.084

Mean forecast errors were 9.03% (Decision) and 10.48% (Audit). At all four checkpoints, E improved on I, with a mean MSE reduction of 40.1%.

Reviewer_zsMPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PINNMorph, an LLM-based framework that adjusts a PINN during training. Based on the current training state, it selects changes to the network, loss terms, sampling, or optimization settings while keeping previously learned parameters where possible. The effects of these changes are then used to improve later decisions. Experiments compare the method with existing PINN approaches on 13 PDE benchmarks.

Critical strengths

The method achieves the lowest mean MSE across all 13 tested problems, with clear improvements on several tasks. The ablations also suggest that selecting changes based on the training state and using execution feedback are useful. These results support the potential value of adapting the training process rather than keeping all settings fixed.

Critical weaknesses
  1. Limited novelty and unclear need for LLMs. PINNsAgent already uses LLMs to automate PINN configuration and architecture search. PINNMorph extends this idea to changes within a single training run, which is useful, but the contribution currently appears incremental to me. More importantly, the experiments do not clearly show why an LLM is needed for these decisions. A small neural controller or a simple rule-based method might provide similar benefits using the same training statistics and available actions. Comparisons with such alternatives are needed to separate the value of online adaptation from the value of using an LLM.

  2. The role of repeated LLM intervention during training needs stronger justification. The LLM selects training changes rather than directly replacing the numerical optimizer. Nevertheless, it is unclear whether repeatedly involving it during training is more useful than using it to choose an initial configuration or a training schedule. A comparison between these settings under a matched total compute budget would help justify the proposed design. The analysis should also consider whether repeated changes make training less stable or introduce costs that simpler tuning methods could avoid.

  3. Limited evidence of performance at larger scales. The evaluation focuses on numerical PDE benchmarks. Although it includes Poisson 5D, the results do not establish how well the method scales to substantially higher dimensions or larger application problems. Tests with larger networks and more training points, together with accuracy and total runtime, would help show whether the benefits remain as the computational cost grows.

Questions for authors

none

4 comments

We thank the reviewer for the helpful comments on online adaptation and the role of the LLM.

1. What the LLM contributes

PINNsAgent evaluates configurations across separate runs. PINNMorph uses an LLM to generate and revise policies across six intervention families during one run. A second LLM scores candidate actions using the current training diagnostics. For example, a localized residual can prompt a local expert, while uneven residuals across the domain can prompt resampling (main paper, Table 4). Deterministic checks decide whether an action can be executed, and the numerical optimizer continues from the current weights.

2. Live decisions versus a predetermined schedule

Fixed Population freezes the initial policies but keeps live diagnosis and action selection. Recorded Replay executes a schedule from an independent donor seed without adapting it to the target run.

Table R1. Solution MSE, mean ± SD over five runs.

PDEFixed PopulationRecorded ReplayPINNMorph
Burgers 1D1.93E-4 ± 2.03E-44.42E-4 ± 7.65E-46.48E-5 ± 3.29E-5
Heat 2D MS1.12E-3 ± 3.41E-52.42E-3 ± 3.13E-32.34E-4 ± 2.21E-5
NS 2D CG2.76E-3 ± 4.31E-45.23E-3 ± 4.31E-31.47E-3 ± 4.13E-4
Poisson 5D1.39E-6 ± 1.31E-73.34E-6 ± 5.75E-66.26E-7 ± 8.32E-8

PINNMorph had lower mean MSE than both controls on all four PDEs. The median ratios of control MSE to PINNMorph MSE were 2.60× for Fixed Population and 6.08× for Recorded Replay. Replay directly tests whether a schedule carries over to a new seed.

3. Stability, cost, and scale

We tested Heat 10D and Poisson 10D at two starting widths and two point budgets. Within each configuration, both methods used the same four-layer starting network, collocation budget, and 20,000 steps. The largest cases used 98,304 training points for Heat and 81,920 for Poisson. Each PDE used 32,768 fixed evaluation points sampled separately from training.

Table R2. Five-run mean test MSE and end-to-end runtime, including diagnosis, interventions, and LLM calls.

PDEInitial widthInterior pointsMSE: No Morph / PINNMorphRuntime (s): No Morph / PINNMorphRuntime increase
Heat 10D12832,7686.18E-5 / 3.24E-71,662 / 2,19632.1%
Heat 10D25632,7686.14E-7 / 1.90E-72,892 / 3,19210.4%
Heat 10D25665,5361.54E-4 / 2.35E-75,544 / 5,7062.9%
Poisson 10D12832,7684.87E-4 / 6.97E-71,320 / 1,48812.7%
Poisson 10D25632,7681.00E-6 / 2.86E-72,520 / 2,92216.0%
Poisson 10D25665,5362.03E-5 / 2.51E-75,148 / 5,2862.7%

PINNMorph had lower mean MSE in all six configurations. At the largest point budgets, the additional time was 162 s for Heat and 138 s for Poisson. These runs test higher dimension, network width, and point count.

Thank you for the response and additional experiments. While these strengthen the empirical evidence, I still find the theoretical contribution and methodological novelty limited. The approach appears to be an incremental extension of existing LLM-based PINN automation, and its advantages over simpler adaptive controllers remain unclear. I therefore maintain my original rating of weak rejection.

Thank you for raising this important question. In PINNMorph, the outcome of one intervention can change the policy used for the next decision while the same PINN continues training. To test what this adds, we compared PINNMorph with a rule-based adaptive controller (R) that reads its training diagnostics, uses the same execution checks, and makes no LLM calls.

1. How the adaptive controller chooses an intervention

We fixed eight rules covering all six intervention families and used the same rules across PDEs and seeds.

Base scoreFamilyTriggerAction
1.00Objective balancingBoundary or initial conditions dominate the residualsIncrease the weight on these constraints
0.95Gradient handlingGradient conflict persists and physics residuals plateauProject conflicting gradients
0.90Objective balancingComponent gradient norms are imbalancedRebalance loss weights
0.85SamplingResiduals concentrate in a small regionResample there at a fixed point budget
0.80Additive representationResiduals concentrate in a small regionAdd a localized expert
0.75Additive representationMultiscale or oscillatory evidence coincides with a plateauAdd a Fourier branch
0.70TopologyDiagnostics indicate a capacity or representation limitWiden the middle hidden layers
0.65Optimizer controlIn late Adam, physics residuals plateau, gradients are healthy, and L-BFGS is availableSwitch to L-BFGS

When several rules apply, R adds to each base score the mean physical-feedback score from up to three recent uses of the same operation under the same trigger. With no qualifying history, this adjustment is zero. R selects the highest-scoring eligible action if its score is positive. Otherwise, training continues unchanged.

2. Complete training runs

Table 1 reports the solution MSE (mean ± std over five runs), with all methods evaluated under the same 20,000-step training budget.

PDENo MorphAdaptive controller (R)PINNMorph
Burgers 1D6.82E-4 ± 5.05E-43.67E-3 ± 3.51E-36.48E-5 ± 3.29E-5
Heat 2D MS4.26E-3 ± 3.91E-44.64E-3 ± 5.36E-42.34E-4 ± 2.21E-5
NS 2D CG8.72E-3 ± 3.38E-36.07E-3 ± 4.59E-41.47E-3 ± 4.13E-4
Poisson 5D9.00E-6 ± 3.72E-61.45E-5 ± 9.34E-66.26E-7 ± 8.32E-8

On NS 2D CG, the adaptive controller reduced mean MSE by 30.4% relative to No Morph. PINNMorph reduced it by a further 75.8% relative to R. Across all four problems, PINNMorph had the lowest mean MSE.

3. Decisions from the same checkpoint

Table 2 compares one decision from each of four saved training states. N continued without a new intervention, R used the fixed rules, and E used the evolved policies. All three branches then trained for 500 additional optimizer steps without further intervention, and solution MSE was evaluated at the end.

PDECheckpoint stepNRE
Burgers 1D10,0002.51E-42.49E-41.08E-4
Heat 2D MS7,6001.23E-32.31E-35.32E-4
NS 2D CG12,0003.51E-33.19E-32.07E-3
Poisson 5D9,0002.14E-62.75E-61.12E-6

E had lower MSE than R at all four checkpoints shown. For Heat 2D MS, the checkpoint comparison in Table 2 yielded MSEs of 1.23E-3 for N, 2.31E-3 for R, and 5.32E-4 for E. E therefore achieved a 77.0% lower MSE than R from the same starting state. This result supports the effectiveness of the evolved policy relative to the fixed-rule controller at this training state.

These results highlight the advantage of evolving intervention policies rather than relying on predefined rules. While the adaptive controller can adjust the ranking of fixed operations using feedback, its actions and trigger conditions must be specified in advance. In contrast, PINNMorph can generate, refine, and combine interventions based on the current training state and accumulated execution feedback. This allows it to address interacting training difficulties without requiring a separate hand-written rule for every situation. The results suggest that this flexibility contributes to PINNMorph's stronger performance over the adaptive controller.

Thank you for the additional experiments. However, it remains unclear whether repeated LLM interventions outperform configuration search across multiple runs under the same total computational budget. The checkpoint experiments show short-term gains but provide limited evidence for the long-term value of evolving policies. Moreover, extending existing LLM automation to online PINN interventions appears narrow in scope and incremental. I therefore find the methodological novelty insufficient and maintain my weak rejection rating.

Reviewer_TPu7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PINNMorph adapts a PINN while it trains instead of fixing its configuration beforehand. At regular intervals the training state is diagnosed on a held-out set of collocation points. Five LLM-generated policies each propose an intervention (topology, additive branches, loss weights, gradients, sampling, optimizer phase); a second LLM scores them, and at most one is applied in place, keeping the unchanged parameters and the compatible optimizer state. Each intervention is judged against an extrapolated no-intervention trend of the physics residuals, and that feedback evolves the policy population.

Critical strengths

Interventions are applied to the network as it trains, with unchanged parameters and compatible optimizer state kept, a real difference from between-run search such as PINNsAgent. The evaluation (13 PDEs, five seeds, five baselines, three ablations, a cost table) is broad enough to be informative, and the numbers agree across Tables 1, 3 and 5.

Critical weaknesses
  1. The ablations only partly isolate what real-time decision-making adds, and the baselines' training protocol is not clearly specified. The three ablations in Table 2 show that state-conditioned selection and feedback-guided evolution beat their randomized versions, and that the base network alone does worse. They do not show what live decisions add over a predefined intervention schedule with the same mechanisms, say a recorded schedule replayed on new seeds. The paper's framing "Unlike pre-training architecture search or fixed adaptation rules" rests on exactly that difference. For the baselines, Appendix B gives the step budget and PINNsAgent's LLM, but not which B.1 settings apply to each of them, nor their collocation budget, constraint handling or final-model rule. So I cannot read "the same budget" in Table 1 as the same protocol.

  2. Where the results stand against PINNacle's published numbers is never addressed, and on two problems those numbers are two to three orders of magnitude lower. The paper takes its problems and reference solutions from PINNacle but compares against none of PINNacle's methods and cites none of its numbers. Under PINNacle's standard protocol (Hao et al., 2024), NTK weighting reports an MSE of 5.00E-5 on Poisson 2D C and 6.99E-5 on Poisson 2D CG; Table 1 has 1.01E-1 and 2.62E-2. Evaluation points, the MSE convention for multi-output problems and collocation budgets may differ, so this is a gap between reported numbers, not a verdict, and the result against the five compared methods stands. But the comparison leaves out the benchmark's own effective methods, and until the protocols are aligned I cannot place the reported gains against what is already known on this benchmark.

Questions for authors
  1. Do the Appendix B.1 settings (Adam followed by L-BFGS) apply to the five baselines too, and do they share the collocation budget, constraint handling and final-model rule? RoPINN, for instance, reports Adam-only training on PINNacle. What is the dashed line in Fig. 2, and what are the final parameter counts per PDE?
  2. Random Evolution replaces feedback-guided evolution with randomized evolution (Sec. 4.3), so Table 3 shows that feedback beats random mutation, not that evolving the population beats keeping it fixed. Is there a fixed-population variant? The most direct additional control would be a recorded intervention schedule replayed on new seeds without live diagnosis.
1 comment

We thank the reviewer for the helpful comments on online decision-making and the baseline training protocol.

1. Baseline protocol

All methods used the same PDEs, prescribed boundary/initial conditions, and a nominal budget of 20,000 optimizer steps per run. Each started with 5,000 interior and 256 boundary points, plus 256 initial-condition points for time-dependent PDEs. All except PINNsAgent started from four hidden layers of 128 Tanh units. During training, sampling, loss weighting, gradient updates, and total search cost can differ. Appendix B.1 and B.6 give the details.

Table R1. Optimizers and adaptation mechanisms (Appendix E.1, Table E.1). For the specified optimizers, the learning rates were 1e-3 for Adam and 1.0 for L-BFGS.

MethodOptimizerAdaptation mechanism
SA-PINNAdam → L-BFGSPoint-wise adaptive weights
ConFIGAdamConflict-free gradients
RoPINNAdamRegional optimization
HARMONICAdamGradient harmonization
PINNsAgentSearchedConfiguration search
PINNMorphAdam → L-BFGSOnline multi-mechanism adaptation

The five baselines selected the checkpoint with the lowest solution MSE on the same reference points used to report MSE (oracle selection). PINNMorph selected its checkpoint by the physics-audit score without reference-solution labels.

2. Figure 2 and final parameter counts

PINNMorph employs two independently sampled, fixed held-out collocation sets: a Decision set for online diagnosis and action selection, and an Audit set for evaluating intervention outcomes. Neither set participates in gradient updates or uses reference solutions.In Figure 2, the Decision physics score (orange dashed line) is computed on the Decision set and measures physical constraint satisfaction to support training-state diagnosis and action selection. The Audit physics score (green dash-dotted line) is computed on the separate Audit set and evaluates physical constraint satisfaction to assess intervention outcomes.For both scores, lower values indicate better satisfaction of the governing equations and physical constraints.

Table R2. Mean final parameter counts over five runs, first four PDEs. Appendix E.1, Table E.2 gives all 13. “Fixed baselines” gives the count for each of SA-PINN, ConFIG, RoPINN, and HARMONIC.

PDEFixed baselinesPINNsAgentPINNMorph
KS 1D50,04981,45153,132
Burgers 1D50,04956,09354,467
NS 2D CG50,30751,88151,615
Heat 2D CG50,17759,58450,692

3. Comparison with PINNacle

We did not include a direct comparison with PINNacle’s published results in the paper because the two studies use different training configurations for the same PDEs. We use PINNacle’s problem definitions and reference solutions. On Poisson 2D C and Poisson 2D CG, however, PINNacle uses 8,192 interior and 2,048 boundary points, whereas our runs start with 5,000 and 256, respectively. On both Heat 2D MS and Heat 2D CG, PINNacle uses 32,768 interior points, 8,192 boundary points, and 8,192 initial-condition points. Our runs start with 5,000, 256, and 256 points, respectively. Similarly, PINNacle adopts substantially larger budgets for boundary and initial conditions on the other PDEs.

4. Fixed population and schedule replay

Random Evolution tests feedback against random updates. We added two controls to answer the distinct questions you raised. Fixed Population freezes the initial policies while keeping live diagnosis and action selection. Recorded Replay applies schedules recorded from independent donor seeds, so the target run cannot change its schedule in response to its own state.

Table R3. Solution MSE, mean ± std over five runs.

PDEFixed PopulationRecorded ReplayPINNMorph
Burgers 1D1.93E-4 ± 2.03E-44.42E-4 ± 7.65E-46.48E-5 ± 3.29E-5
Heat 2D MS1.12E-3 ± 3.41E-52.42E-3 ± 3.13E-32.34E-4 ± 2.21E-5
NS 2D CG2.76E-3 ± 4.31E-45.23E-3 ± 4.31E-31.47E-3 ± 4.13E-4
Poisson 5D1.39E-6 ± 1.31E-73.34E-6 ± 5.75E-66.26E-7 ± 8.32E-8

PINNMorph had lower mean MSE than both controls on each of the four PDEs. Across PDEs, the median ratio of control MSE to PINNMorph MSE was 2.60× for Fixed Population and 6.08× for Recorded Replay. These results support updating the policies and choosing actions from the current trajectory on the tested problems.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.