DASyR-LLM: Domain-Aware Symbolic Regression with LLMs for Kinetic Model Discovery

The paper introduces DASyR-LLM, an iterative symbolic-regression framework in which an LLM critiques candidate kinetic models using qualitative physicochemical reasoning and proposes new rate expressions. Across four in-silico cases covering heterogeneous catalysis and bioprocess systems, the method reduced iterations needed to recover the ground-truth model by 41.7%–79.3% versus a state-of-the-art symbolic-regression framework. The LLM directly proposed the correct structure in more than half of guided runs, while independent validation achieved R² > 0.98 in every case. Ablations suggest that a smaller LLM retained much of the discovery efficiency.