acceptodds
Under review as a conference paper at ICLR 2027

An Energy-Aware Sampling Method for Physics-Informed Neural Networks

Abstract

Adaptive sampling is a critical technique in Physics-Informed Neural Networks (PINNs), particularly for solving stiff and multiscale partial differential equations (PDEs). Existing adaptive sampling methods predominantly rely on the PDE residual. However, residuals measure local equation violations rather than true approximation errors, frequently neglecting regions with large actual errors, which leads to inefficient resource allocation and slow convergence. To overcome this limitation, we explore an energy-aware sampling paradigm that leverages the intrinsic structure of energy-driven PDEs. Specifically, focusing on phase-field equations as a representative class, our theoretical analysis reveals that the double-well potential not only identifies physically important regions, but also dictates the necessity of a higher collocation point density in high-potential regions to achieve uniform generalization accuracy. Motivated by these insights, we propose Potential-based Adaptive Sampling (PAS), which exploits the derivative-free nature of the potential to integrate global exploration with local refinement. Extensive comparisons demonstrate that PAS significantly outperforms residual-based baselines in accuracy, convergence speed, and training efficiency.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PINNs, PDEs, Potential-based Adaptive Sampling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3GYnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the author mainly focuses on AC and CH types of equations, which could be derived from the energy dissipation law. The author first pointed out the drawback of the traditional PINN method, which uses the residual as an indicator for the sampling. Then, they propose potential-based adaptive sampling and calibrate the idea with some experiments on AC, CH, and RD equations from 1d to high dimensions

Critical strengths
  1. The authors proposed a simple but efficient way for AH and CH types of equations using the potential.
  2. The results presented in the paper confirm the efficiency for these types of equations.
Critical weaknesses
  1. It seems that the author does not acknowledge the existing work on the PINN residual and error analysis. In 2024, Liu et al., in the paper: Preconditioning for Physics-informed neural networks, and Tim De Ryck et al on ICLR, AN OPERATOR PRECONDITIONING PERSPECTIVE ON TRAINING IN PHYSICS-INFORMED MACHINE LEARNING, have pointed out the fundamental problems of PINNs using the residual as an indicator. They have done rigorous analysis and related the gap with the condition number of the differential operator.
  2. The authors should be very careful about the Modica-type gradient estimations. It is only valid for the stationary AH equation. For non-stationary AH, a small correction term is needed. For the Cahn-Hilliard equation, it is worse since there is not maximum principle for CH equation.
  3. I have some doubts about the numerical results in Table 1 and Table 7. why the Random-R, R3, Rsmote, and PAS always have the same memory in Table 1 . And why you increase Npool, does the memory not change in Table 7. 4.The authors overclaimed the applicable scope. It is not applicable for general energy-based govening equation for example the PNP, NS equation, where there is no potential energy inside or is not the dominate one. So the auther should narrow down there scope.
Questions for authors
  1. Can the authors clarify whether F(uθ)F(u_\theta) is intended as an interface/transition indicator or as a surrogate for solution error? What do you mean by "suitable assumptions," especially for CH? For example, if my u_0= sin(kx), then the condition is not satisfied at the point x=0

  2. How is GPU memory measured for each method? In particular, are candidate pools processed in batches, on CPU or GPU, and is adaptive-sampling overhead included consistently for PAS and RAD/RAR?

  3. Can the authors test PAS in a case where the predicted interface is initially misplaced, to assess whether F(uθ)F(u_\theta)-based sampling can recover the correct interface rather than only refine an already well-located one?

  4. Please do a careful literature search. Your current theoretical results have been mentioned.

  5. Your current method is only valid for AH. The applicability to Cahn–Hilliard requires a more discriminating test. I suggest evaluating PAS during spinodal decomposition from a near-homogeneous, non-well-prepared initial state, and reporting performance separately during early interface formation and late coarsening. Also, the result in Proposition 2 is not correct. For CH, there is a conservation law. So the solution cannot be any constant. If it is a constant solution, it must be the avarge of the initail value.

  6. The proposed PAS indicator relies on a potential-energy term F(u)F(u) that is strongly correlated with transition/interface regions, which is a special structural feature of phase-field models such as Allen–Cahn. The authors should substantially narrow the claimed scope of applicability and clearly specify the class of PDEs for which a meaningful F(u)F(u)-based sampling indicator exists.

Reviewer_BEqoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper propose a new adaptive sampling method for training PINNs. The authors experimantally found and theoretically proved that in ill-conditions a small residual can coexist with a large solution error, so the commonly used residual-based sampling method may mislead PINN training. Motivated by the prior estimation Eq.(10) in the paper, they propose to use the potential F(u)F(u) as the sampling indicator. F(u)F(u) is physical meaningful and interface sensitive, so it can adaptively samping collocation points in dynamically important regions, without additional gradient computation cost. Experiments are conducted on 1D/2D Allen-Cahn, 1D Cahn-Hilliard, and 10D/20D reaction-diffusion equations, and cpmpared with baselines Random-R, RAR (Lu et al., 2021), RAD (Wu et al., 2023), R3 (Daw et al., 2023) and RSmote (Luo et al., 2025), to support the method.

Critical strengths

This is a simple, cheap, genuinely physics-informed indicator that works empirically on hard phase-field problems. The potential requires one forward pass and no derivatives, yet it is much better aligned with the pointwise error than the residual.

Critical weaknesses

1.The potential indicator is equation-specific, and the theory does not establish the potential to the denser-sampling link. All but one benchmark use the same double-well potential (Eq. 3), and the one exception (the 1D elliptic inverse problem) uses a hand-crafted one. Nothing in the paper identifies a general class of PDEs for which a usable potential is available.

2.The empirical evidence for robustness and efficiency is too narrow and not statistically or computationally controlled. The hyperparameter study is effectively on single-equation, not a general class of PDEs. The abstract's "significantly outperforms … in training efficiency" overstates the 1D CH result, where PAS is slower than four of five baselines (17,799 s vs. 15,480 s for Random-R) while improving the error by only ~1.3×.

Questions for authors

1.The potential FF is exactly the potential term of the Ginzburg–Landau energy, and for the elliptic inverse problem (Eqs. 118–121) the solution minimizes Eε,λE_{ε,λ}. Why is the potential used only as a sampling signal, and not compared against energy-based training (Deep Ritz–type/variational PINNs rather than the strong form PINN)? Similarly, could you compare against the concurrent energy/thermodynamics-based samplers (EDRAS[1]) on 1D Allen–Cahn, the one setting where their formulations are validated? If these are not meaningful comparisons, please state the technical reason; otherwise the contribution's positioning and novelty are unclear.

[1]Li, C., Yu, W., & Wang, Q. (2026). Energy Dissipation Rate Guided Adaptive Sampling for Physics-Informed Neural Networks: Resolving Surface-Bulk Dynamics in Allen-Cahn Systems. International Journal of Numerical Analysis and Modeling, 23(4), 576-604. https://doi.org/10.4208/ijnam2026-1025

  1. Were hyperparameters K=3K = 3 and P1=0.2P_1 = 0.2 selected using the 1D Allen–Cahn results (the best entry of Table 6)? What are the per-benchmark optima, and do the same settings transfer to CH/2D/10D/20D/inverse? Also, Eq. (11) normalizes by the pool mean of FkF^k and then adds c=1c = 1, which differs from the sum-normalized form of RAD (Wu et al., 2023); how sensitive is PAS to cc and to the normalization (e.g., c∈0,0.1,1c \in {0, 0.1, 1})?
Reviewer_hvasOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies adaptive collocation for physics-informed neural networks (PINNs), focusing on phase-field equations with narrow interfaces and stiff dynamics. The authors point out that the internal PDE residual may not align with the true solution error, and that computing residuals over a large pool of candidate points requires costly automatic differentiation.

To address this, the authors propose Potential-based Adaptive Sampling (PAS), which uses the potential energy of the currently predicted solution as the sampling criterion. The algorithm first performs potential-weighted global sampling, then locally refines by perturbing high-potential points, ultimately reconstructing a fixed-size training point set. The design is justified through a counterexample where residuals fail, a conditional regional approximation complexity analysis, and a local resolution analysis, and is validated on Allen–Cahn, Cahn–Hilliard, high-dimensional reaction–diffusion, and elliptic inverse problems.

Critical strengths

For phase-field equations, the transition regions between stable phases typically require high resolution. Using the double-well potential to highlight these regions is a targeted and easily understandable design. Compared to relying entirely on residuals, the method introduces a useful physical prior. The potential can be computed directly from the network's predicted values, without additional input derivatives. When the candidate pool is large and the equations involve high-order derivatives, this design reduces the computational burden of the sampling stage. The authors compare not only the full method, but also replace the residual with the potential within RAD and RAdSAM (RSmote), and provide a gradient-metric control under the same sampling pipeline. These experiments help clarify the role of the potential metric. The appendix also clearly delimits the scope of some theoretical conclusions, avoiding outright claims of optimal sampling or convergence guarantees.

Critical weaknesses

A key efficiency motivation of this paper is to avoid computing expensive residuals over the entire candidate pool, but existing methods have also addressed this problem through low-cost approximate weights. For example, DMIS (AAAI 23) uses a dynamic grid: it computes the residual loss on a small set of points and interpolates to estimate sampling weights for the remaining points. It differs from PAS in both its metric and its sampling targets, but it remains an important related method when evaluating sampling cost. The current manuscript neither discusses nor compares against it. PAS simultaneously changes the importance metric, the global collocation distribution, and the local refinement mechanism. The existing ablations provide partial evidence, but a control is still missing in which, within the full PAS pipeline, only the potential is replaced by the residual. Table 4 shows that replacing RAD's metric with the potential yields only a relatively limited average accuracy improvement, whereas the gains of the full PAS are considerably larger. This suggests that the final performance depends on the interplay between the metric and the sampling mechanism.

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.