Abstract: Undergraduate experimental design typically requires extensive trial-and-error experimentation to identify optimal reaction conditions, a process that demands considerable time and resources. To simplify this complex experimental design process and to enhance students’ interest in advanced technologies and interdisciplinary fields, an intelligent reaction optimization model framework was introduced. This model is designed to predict the optimal combination of reaction conditions for achieving the highest yield, and this framework has been integrated into the undergraduate organic chemistry curriculum, specifically for the synthesis of acetylsalicylic acid. Herein, 1054 reaction data points were collected conforming to this reaction mechanism as training data for the model, mainly including product, reactant, catalyst, solvent, the main reaction reagents, reaction temperature and yield. Initially, a yield prediction model was pre-trained based on a chemical multi-modal transformer, which served as the objective function for subsequent reaction optimization. Then the Bayesian optimization algorithm was utilized to ascertain the optimal combination of reaction conditions, with the aim of minimizing the negative value of the yield (the maximizing yield). Finally, the model-predicted yields were experimentally validated against the actual yields, and in 100 model tests, the predicted yields consistently fell within the range of the actual yields, demonstrating the model’s excellent performance. The model also identified a reaction reagent combination that yielded up to 90.1%, providing a variety of experimental options for this undergraduate experiment.

Key words: Optimization algorithm, Multi-modal transformer, Machine learning, Acetylsalicylic acid, Optimization