BayesianKAN: A reaction condition optimization framework integrating Kolmogorov‐Arnold network and Bayesian optimization

Abstract

Efficient optimization of chemical reaction conditions is crucial for enhancing reaction yield and selectivity, yet traditional methods face inherent limitations including experimental inefficiency, low predictive accuracy, and poor interpretability. This study proposes a novel framework for reaction condition optimization by integrating the Kolmogorov‐Arnold network (KAN) model and Bayesian optimization (BO) algorithm. The KAN model establishes accurate and explicit mappings between reaction conditions and outcomes like yield, while BO iteratively optimizes reaction outcomes to identify optimal conditions based on the KAN model, demonstrating efficacy even with sparse data. This framework is implemented as the BayesianKAN software and validated in two reaction systems: hydrogen peroxide (H ₂ O ₂ ) synthesis and photocatalytic acceptorless dehydrogenation to flavones. KAN exhibits superior fitting accuracy and generalization ability compared to the traditional response surface methodology and other seven common machine learning approaches. Experiments also verify the feasibility and effectiveness of BayesianKAN, achieving 60.7% and 5.2% increases in H ₂ O ₂ production and flavone yield, respectively.