Optimizing Chemical Reactions with Deep Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Chemical Engineering
• Computational Chemistry

Background:

• Optimizing chemical reactions is crucial for efficient synthesis and discovery.
• Traditional methods often require extensive experimentation and can be time-consuming.
• Black-box optimization algorithms exist but can be inefficient in complex chemical spaces.

Purpose of the Study:

• To develop and evaluate a deep reinforcement learning model for optimizing chemical reactions.
• To improve the efficiency and reduce the number of experimental steps required for reaction optimization.
• To enhance the understanding of factors influencing microdroplet reactions.

Main Methods:

• Utilized deep reinforcement learning to iteratively guide experimental condition selection.
• Implemented an efficient exploration strategy using probability distributions for sampling reaction conditions.
• Integrated the DRL model with accelerated microdroplet reaction platforms.

Main Results:

• The DRL model outperformed a state-of-the-art black-box optimization algorithm, requiring 71% fewer steps.
• An efficient exploration strategy improved the optimization performance, reducing regret from 0.062 to 0.039.
• Optimal reaction conditions were identified within 30 minutes for four distinct reactions.
• The model demonstrated adaptability and improved performance on reactions with similar and dissimilar mechanisms.

Conclusions:

• Deep reinforcement learning offers a powerful and efficient approach to chemical reaction optimization.
• The developed exploration strategy enhances the performance and learning ability of the DRL model.
• Accelerated microdroplet reactions combined with DRL enable rapid determination of optimal conditions and provide mechanistic insights.