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