Scientific Code Generation

The field studies how to generate and adapt code for scientific tasks such as solving partial differential equations, configuring simulators, reproducing papers, and analyzing experimental data. Common approaches combine language models with documentation retrieval, task decomposition, specialized agents, and execution feedback; search, evolutionary methods, and training on scientific code are also used. Evaluation increasingly checks numerical accuracy, physical behavior, and fidelity to a paper’s methods—not just whether code runs. A central challenge is that runnable programs can still encode the wrong science, while tacit knowledge, large-scale workflows, and expert judgment remain difficult to capture and verify.