LLMs as Workflow Interfaces for Partial Differential Equations

An arXiv paper (2608.03600) investigates how large language models (LLMs) can enhance workflows related to partial differential equations (PDEs). The authors emphasize that PDEs should not be viewed merely as standalone equations but as integrated workflows that encompass modeling assumptions, governing equations, numerical solvers, diagnostics, and decision-making processes. LLMs are starting to facilitate these workflows by bridging natural language, symbolic math, code, solver outputs, and feedback. The study outlines recent progress in LLM-aided PDE research through three phases: discovering and formulating governing models, creating and refining executable numerical solvers, and utilizing simulation feedback for control, design, and optimization. Despite advancements, challenges remain due to a lack of quality datasets and benchmarks, particularly for knowledge discovery and practical applications, necessitating improved evaluation frameworks and collaborative initiatives in scientific computing.