DASyR-LLM: DOMAIN-AWARE SYMBOLIC REGRESSION WITH LLMS FOR KINETIC MODEL DISCOVERY

Identifying accurate kinetic models from experimental data is a fundamental challenge in chemical engineering. Traditional symbolic regression (SR) methods often struggle with large search spaces and the tendency to produce mathematically accurate but physicochemically implausible models. This paper asks whether embedding a large language model (LLM) into the discovery loop can inject domain expertise to improve the efficiency and physical consistency of kinetic model discovery.

The authors introduce DASyR-LLM, an iterative framework that embeds an LLM module within an existing SR-based pipeline (ADoK-S). The process follows a nine-step loop: generating concentration data, identifying candidate models via SR, estimating reaction rates, and performing parameter estimation. The LLM module acts at each iteration to critique the best-performing SR candidates and propose new, scientifically grounded rate expressions. These LLM-generated candidates are then evaluated alongside SR-generated ones using statistical criteria (AIC) and model-based design of experiments (MBDoE) to guide future data collection.

DASyR-LLM significantly accelerates the discovery of ground-truth models. Across four case studies of increasing complexity—ranging from heterogeneous catalysis to bioprocess systems—the framework reduced the number of iterations required to identify the correct model by 41.7% to 79.3%. Notably, the LLM successfully proposed the correct model structure in over half of the guided runs. Despite this efficiency, the predictive performance of the models remained high, with R² values exceeding 0.98, matching the performance of the baseline SR-only approach. Ablation studies confirmed that both the SR component and the LLM's reasoning capabilities are essential, though even smaller LLMs were sufficient to maintain discovery efficiency.