A new survey takes stock of how large language models are being pointed at partial differential equations, the math that describes everything from heat flow to airflow over a wing.
The paper reviews LLM-assisted PDE research across three stages: discovery, where models help formulate equations from physical principles; solving, where they configure and run numerical solvers; and optimization, where they translate simulation results into design and control decisions. The authors also survey existing benchmarks for judging whether an LLM's output is scientifically valid, computationally efficient, and reliable as part of a larger workflow. Rather than presenting a new tool, this is a stock-take of a fast-moving research area, cataloging what has been tried and what has not.
PDE work has traditionally required years of specialized training in both the math and the numerical methods used to approximate solutions on a computer. If LLMs can reliably automate even parts of that pipeline, engineers and scientists without that specific background could iterate on simulations faster. That is a real shift for fields like aerospace, climate modeling, and materials science, where the bottleneck is often expert time rather than compute.
The authors admit the approach still lacks PDE-specific reasoning, transferable domain expertise, and a validated way to run model-driven experiments, which makes this more a research roadmap than a working assistant.