EvoPINN: Agentic Discovery of Executable Algorithms for Physics-Informed Neural Networks
Quick Answer
This paper shows that EvoPINN revolutionizes the development of Physics-Informed Neural Networks (PINNs) by automating algorithm discovery through a modular search space and LLM-driven modifications, achieving significant reductions in relative L2 error across various PDE regimes.
Quick Take
Notably, it autonomously invented the SLRC-PINN architecture, demonstrating its capability to create new scientific computing mechanisms.
Key Points
- EvoPINN transforms manual PINN design into an automated algorithm discovery process.
- Utilizes agents for iterative, memory-conditioned program modifications.
- Achieves significant L2 error reduction compared to traditional baseline methods.
- Introduces SLRC-PINN, a novel architecture with proven performance gains.
- Validates candidates through strict structural verification and PDE evaluations.
DeepSignal Analysis
What happened
EvoPINN is a new framework that automates the discovery of algorithms for Physics-Informed Neural Networks (PINNs). It utilizes a modular search space and Large Language Models (LLMs) to propose modifications, leading to significant reductions in relative L2 error across various partial differential equation (PDE) regimes. The framework has autonomously developed a new architecture called SLRC-PINN.
Key evidence
- EvoPINN reformulates the development of PINNs from manual design to an execution-grounded algorithm discovery problem.
- The framework utilizes an LLM agent to propose modifications, ensuring scientific validity through structural verification and PDE evaluation.
- EvoPINN has demonstrated significant reductions in relative L2 error across diverse PDE regimes, including oscillatory and nonlinear transport.
Why it matters
The automation of algorithm discovery in PINNs could streamline the development process, reducing reliance on manual engineering. This could lead to more efficient solutions for complex scientific problems, as EvoPINN's ability to autonomously create new architectures like SLRC-PINN suggests a potential shift in how scientific computing mechanisms are developed. The implications for research and industry could be substantial, particularly in fields reliant on solving PDEs.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Physics-informed neural networks (PINNs) have emerged as a powerful paradigm for solving partial differential equations (PDEs), yet their performance heavily relies on the manual, trial-and-error engineering of neural representations, loss formulations, and optimization dynamics. While Large Language Models (LLMs) offer a promising avenue for automated design, unconstrained code generation often yields mathematically invalid or numerically unstable solutions under strict scientific computing constraints. To bridge this gap, we propose \textbf{EvoPINN}, an agentic framework that reformulates PINN development from labor-intensive manual design into a rigorous, execution-grounded algorithm discovery problem. EvoPINN navigates a modular search space by decoupling neural representations from training programs, utilizing an LLM agent to iteratively propose memory-conditioned programmatic modifications. To ensure scientific validity, all candidates undergo strict structural verification and budget-matched PDE evaluation. Extensive experiments across diverse PDE regimes (oscillatory, elliptic, dissipative, and nonlinear transport) demonstrate that EvoPINN discovers PDE-specialized learning algorithms that significantly reduce relative $L_{2}$ error compared to baselines. Crucially, EvoPINN autonomously invented SLRC-PINN, a novel architecture whose performance gains persist under rigorous parameter-matched comparisons, establishing the viability of execution-grounded agents for discovering genuinely new scientific computing mechanisms.
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Neural and Evolutionary Computing (cs.NE) |
| Cite as: | arXiv:2607.26490 [cs.AI] |
| (or arXiv:2607.26490v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2607.26490 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Peng Yin [view email]
[v1]
Wed, 29 Jul 2026 05:36:28 UTC (445 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →HOBA: Hierarchical On-Policy Bidding Agents for Adaptive Online Advertising
HOBA (Hierarchical On-policy Bidding Agents) is a novel hierarchical reinforcement learning framework that enhances online advertising bidding systems by improving adaptability and reducing hyperparameter tuning costs. It utilizes a for hyperparameter inference, a SARSA agent for expert model selection, and a dynamic expert pool for bid execution, achieving a +3.6% increase in target cost during large-scale deployment and outperforming state-of-the-art baselines on AuctionNet.