Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
167 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
42 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks (2207.10289v1)

Published 21 Jul 2022 in physics.comp-ph and cs.LG

Abstract: Physics-informed neural networks (PINNs) have shown to be an effective tool for solving forward and inverse problems of partial differential equations (PDEs). PINNs embed the PDEs into the loss of the neural network, and this PDE loss is evaluated at a set of scattered residual points. The distribution of these points are highly important to the performance of PINNs. However, in the existing studies on PINNs, only a few simple residual point sampling methods have mainly been used. Here, we present a comprehensive study of two categories of sampling: non-adaptive uniform sampling and adaptive nonuniform sampling. We consider six uniform sampling, including (1) equispaced uniform grid, (2) uniformly random sampling, (3) Latin hypercube sampling, (4) Halton sequence, (5) Hammersley sequence, and (6) Sobol sequence. We also consider a resampling strategy for uniform sampling. To improve the sampling efficiency and the accuracy of PINNs, we propose two new residual-based adaptive sampling methods: residual-based adaptive distribution (RAD) and residual-based adaptive refinement with distribution (RAR-D), which dynamically improve the distribution of residual points based on the PDE residuals during training. Hence, we have considered a total of 10 different sampling methods, including six non-adaptive uniform sampling, uniform sampling with resampling, two proposed adaptive sampling, and an existing adaptive sampling. We extensively tested the performance of these sampling methods for four forward problems and two inverse problems in many setups. Our numerical results presented in this study are summarized from more than 6000 simulations of PINNs. We show that the proposed adaptive sampling methods of RAD and RAR-D significantly improve the accuracy of PINNs with fewer residual points. The results obtained in this study can also be used as a practical guideline in choosing sampling methods.

Citations (283)

Summary

  • The paper introduces novel adaptive sampling methods, RAD and RAR-D, that significantly enhance PINN accuracy with optimized residual point allocation.
  • It compares six uniform sampling techniques with adaptive approaches through over 6000 simulations on various forward and inverse PDE problems.
  • The study demonstrates that residual-based adaptive strategies are essential for capturing complex PDE features, such as steep gradients and multi-scale behaviors.

A Comprehensive Study of Sampling Strategies for Physics-Informed Neural Networks

The paper presents an in-depth exploration of sampling strategies for Physics-Informed Neural Networks (PINNs), which have garnered attention for their ability to address forward and inverse problems involving partial differential equations (PDEs). Unlike conventional numerical solvers that rely on mesh-based approaches, PINNs employ neural networks that leverage automatic differentiation to incorporate PDE constraints directly into their loss functions. The selection and distribution of residual points, the points at which the PDE loss is evaluated, are crucial to the performance of PINNs. However, earlier research on PINNs has not extensively explored the implications of different residual point sampling methods.

The authors classify sampling strategies into two main categories: non-adaptive uniform sampling and adaptive nonuniform sampling. The paper evaluates a range of methods including six uniform sampling approaches—equispaced grids, uniformly random sampling, Latin hypercube sampling, and quasi-random low-discrepancy sequences such as Halton, Hammersley, and Sobol sequences. A resampling approach is tested to assess its efficacy in enhancing accuracy. Furthermore, the paper introduces two novel residual-based adaptive sampling methods—Residual-based Adaptive Distribution (RAD) and Residual-based Adaptive Refinement with Distribution (RAR-D)—which dynamically adjust the distribution of residual points based on the PDE residuals as training progresses.

The authors conducted over \num{6000} simulations encompassing six PDE problems, providing a thorough evaluation of both forward and inverse problems. Numerical results indicate that the RAD and RAR-D methods significantly enhance the accuracy of PINNs while using fewer residual points. The RAD method consistently outperformed other sampling approaches across all tested scenarios, delivering substantial advances in prediction accuracy, especially for PDEs with complex solutions.

For forward PDE problems such as the diffusion, Burgers', Allen-Cahn, and wave equations, the authors found that RAD and RAR-D offered superior performance over traditional uniform sampling methods. Notably, the RAR-D and RAD approaches demonstrated exceptional robustness in cases where the solutions exhibited steep gradients or multi-scale behaviors.

In the context of inverse problems, such as those involving diffusion-reaction systems and the Korteweg-de Vries equation, RAD excelled in inferring unknown parameters with remarkable precision. The paper reveals that while uniform sampling methods yield decent performance for problems with smoother solutions, adaptive sampling approaches like RAD are indispensable for capturing the dynamics of more challenging PDE systems with limited residual points.

The implications of this paper are significant for the advancement of PINNs in computational science and engineering. The adaptive methods proposed not only improve accuracy but also provide practical guidelines on optimal residual point distribution, enhancing the utility of PINNs in diverse applications. Moving forward, the adaptive sampling strategies introduced could serve as a benchmark for developing more sophisticated sampling schemes in high-dimensional settings, potentially involving advanced sampling technologies like GANs for generating residual points. There is also an opportunity to explore meta-learning techniques to dynamically adapt the sampling distributions further.

Overall, this comprehensive paper on sampling strategies enriches the repertoire of methods available for enhancing the efficacy of PINNs, thereby cementing their role as a versatile tool in solving complex PDE problems.