Exploring Optimal Reaction Conditions Guided by Graph Neural Networks and Bayesian Optimization

Area of Science:

• Artificial Intelligence in Chemistry
• Computational Chemistry
• Organic Synthesis

Background:

• Optimizing organic reaction conditions is vital for efficient chemical synthesis, but traditional experimental methods are time-consuming and costly.
• Artificial intelligence (AI) offers data-driven solutions for predicting reaction conditions, yet novel syntheses present challenges.
• Bayesian optimization (BO) is effective but suffers from inefficiencies, particularly the cold-start problem requiring numerous initial trials.

Purpose of the Study:

• To develop an efficient AI-driven algorithm for optimizing organic reaction conditions.
• To overcome the limitations of traditional Bayesian optimization in chemical synthesis.
• To accelerate the identification of high-yield reaction pathways.

Main Methods:

• Integration of Bayesian optimization (BO) with a graph neural network (GNN).
• The GNN was trained on a large dataset of one million organic synthesis experiments.
• The GNN guides the BO process to efficiently explore and identify optimal reaction parameters.

Main Results:

• The proposed method achieved high-yield reaction conditions 8.0% faster than state-of-the-art algorithms.
• The algorithm outperformed 50 human experts by finding conditions 8.7% faster.
• In 22 additional tests, the method required an average of only 4.7 trials to surpass expert-recommended yields.

Conclusions:

• The AI-driven approach significantly enhances the efficiency of optimizing organic synthesis.
• This method offers a powerful tool for accelerating chemical discovery and development.
• The algorithm is particularly effective when a substantial reaction dataset is available for GNN training.