OpenPINN/PINO: An Extensible Open-Source Framework for Agent-Assisted Physics-Informed Neural Networks and Neural Operators

Substantial challenges in physics-informed learning have motivated many improvements, but fragmented implementations hinder systematic organization, integration, and comparison. We propose OpenPINN/PINO, an extensible open-source PyTorch framework for physics-informed neural networks and neural operators. Two shared specifications connect problem setup, local computation, and evaluation. Agents support problem formulation, solver planning, and diagnosis. ProblemSpec defines geometry, governing equations, constitutive relations, parameters, initial and boundary conditions, and optional data. MethodSpec configures the model, sampling, loss, optimization, and differentiation. Researchers can register new problem templates and method components, then combine or compare strategies such as Fourier features, residual-adaptive sampling, and self-adaptive loss weighting for a fixed problem. Agents translate natural-language descriptions into specifications, flag missing information, and propose solver plans. The framework validates these specifications and deterministically generates local training and testing code. A diagnostic agent reviews training logs and evaluation results to identify problems and suggest revisions. Three examples show how method choices and agent diagnostics affect accuracy. For viscous Burgers, Fourier features and residual-adaptive sampling each reduced relative $L^2$ error by about $75\%$ from a shared baseline. For periodic Poisson, spectral differentiation reduced mean-free relative $L^2$ error by about $68\%$ across 16 held-out forcings. For a plate with a hole, successive agent diagnoses guided nondimensionalization and optimization. After initial training failure, all three runs met the $10\%$ target, with relative $L^2$ displacement errors of $3.51\%$--$6.31\%$.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-17
DOI
https://doi.org/10.48550/arxiv.2611.23135
Primary Topic
Model Reduction and Neural Networks
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

OpenPINN/PINO: An Extensible Open-Source Framework for Agent-Assisted Physics-Informed Neural Networks and Neural Operators

Zhenyu Yin, Timon Rabczuk, Jidong Zhao, Peinan Li et al.
arXiv (Cornell University)
Model Reduction and Neural Networks
preprint

OpenPINN/PINO: An Extensible Open-Source Framework for Agent-Assisted Physics-Informed Neural Networks and Neural Operators

Zhenyu Yin, Timon Rabczuk, Jidong Zhao, Peinan Li, Hehua Zhu, Xiaoying Zhuang, Hao Sun, Xi Wang, Wei Wu, Cheng Zhang, Jun Lu
preprint en

Abstract

Substantial challenges in physics-informed learning have motivated many improvements, but fragmented implementations hinder systematic organization, integration, and comparison. We propose OpenPINN/PINO, an extensible open-source PyTorch framework for physics-informed neural networks and neural operators. Two shared specifications connect problem setup, local computation, and evaluation. Agents support problem formulation, solver planning, and diagnosis. ProblemSpec defines geometry, governing equations, constitutive relations, parameters, initial and boundary conditions, and optional data. MethodSpec configures the model, sampling, loss, optimization, and differentiation. Researchers can register new problem templates and method components, then combine or compare strategies such as Fourier features, residual-adaptive sampling, and self-adaptive loss weighting for a fixed problem. Agents translate natural-language descriptions into specifications, flag missing information, and propose solver plans. The framework validates these specifications and deterministically generates local training and testing code. A diagnostic agent reviews training logs and evaluation results to identify problems and suggest revisions. Three examples show how method choices and agent diagnostics affect accuracy. For viscous Burgers, Fourier features and residual-adaptive sampling each reduced relative $L^2$ error by about $75\%$ from a shared baseline. For periodic Poisson, spectral differentiation reduced mean-free relative $L^2$ error by about $68\%$ across 16 held-out forcings. For a plate with a hole, successive agent diagnoses guided nondimensionalization and optimization. After initial training failure, all three runs met the $10\%$ target, with relative $L^2$ displacement errors of $3.51\%$--$6.31\%$.

arXiv (Cornell University)
Model Reduction and Neural Networks
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon