Research map / Probabilistic and scientific modeling / Differential equations and PDEs
PINNeval: A Comprehensive Evaluation Standard for Physics-Informed Neural Networks
Kevin von Bargen, Hamid Ebrahimy, Tim Römer
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PINNeval, an evaluation standard that decomposes PINN training into four components (domain-importance adaptation, loss balancing, architecture, optimization) and evaluates them jointly rather than in isolation.
Problem
Existing PINN benchmarks study the components of the training pipeline in isolation, and the studies set the supervision/setup in inconsistent ways, making method comparisons hard.
Approach
Systematically divides the PINN training pipeline into four components and provides a unified evaluation standard that assesses them together rather than separately.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use two-stage MLP pipeline (SOAP then SSBroyden2, float64) instead of single-stage PirateNet-SOAP | PINN training pipeline · demanding PDE benchmarks | median relative L2 error and runtime | Improved |
| add quasi-Newton SSBroyden2 stage to MLP at same precision (ablation) | MLP under SOAP · Heat PDE | relative L2 error: error reduced by up to a factor of 1.7 x 10^4 | Improved |
| use float64 precision alone (ablation) | MLP under SOAP · PINN benchmarks | relative L2 error | No clear effect |
| replace PirateNet with MLP under SOAP alone (ablation) | PINN with SOAP · PINN benchmarks | relative L2 error | Hurt |
| use PINNpipe efficient implementation with LAS instead of jaxpi with Random-R | Ginzburg-Landau benchmark · Ginzburg-Landau | runtime (minutes): 746.0 -> 519.2 | Improved |
| adaptive domain-importance methods (RAD, R3, vRBA) instead of random resampling under SOAP (ablation) | PINN pipeline with SOAP · Wave, Heat, KdV, Rayleigh-Taylor | relative L2 error | Hurt |
| use LAS adaptive sampling instead of Random-R (ablation) | PINN pipeline with SOAP · Wave and Heat | relative L2 error | No clear effect |
| random resampling versus adaptive sampling (ablation) | PINN pipeline with SOAP · KdV and Rayleigh-Taylor | relative L2 error | Improved |
| use GradNorm-based loss balancing (ablation) | PINN pipeline with Random-R · nine PDE benchmarks | median relative L2 error: best on 5 of 9 benchmarks | Improved |
| fixed-weight loss baseline (ablation) | PINN pipeline with Random-R · KS, AC, KdV | median relative L2 error | Improved |
| use SoftAdapt loss balancing (ablation) | PINN pipeline · Wave and KS | relative L2 error | Hurt |
| use GradNorm-MTL loss balancing (ablation) | PINN pipeline · KdV | relative L2 error | Hurt |
Closest papers by meaning
- Per-Loss Adapters for Gradient Conflict in Physics-Informed Neural Networks (NeurIPS 2026 · similarity 0.91)
- Efficient Error Certification for Physics-Informed Neural Networks (ICML 2024 · similarity 0.90)
- Challenges in Training PINNs: A Loss Landscape Perspective (ICML 2024 · similarity 0.90)
- CoPINN: Cognitive Physics-Informed Neural Networks (ICML 2025 · similarity 0.89)
- PINNsFormer: A Transformer-Based Framework For Physics-Informed Neural Networks (ICLR 2024 · similarity 0.89)
- ConFIG: Towards Conflict-free Training of Physics Informed Neural Networks (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
