AutoPDE: Reliable Agentic PDE Solving via Explicitly Represented Solver Strategies

AutoPDE is a new code agent designed for solving partial differential equations (PDEs) by explicitly representing solver strategies throughout the solving process. It operates in three stages: PDE analysis, numerical method selection, and adaptive tuning, leveraging a library of reusable PDE-solving skills. In evaluation against the PDE Agent Bench, AutoPDE achieved a pass rate of 54.5%, outperforming the strongest baseline by 14.2 percentage points, which enhances reliability and adaptability in numerical solver development for practitioners.