An active representation learning method for reaction yield prediction with small-scale data.

Area of Science:

• Chemistry
• Machine Learning
• Chemical Engineering

Background:

• Reaction optimization is crucial for chemical research and industrial production.
• Exploring large reaction systems requires reducing extensive experimental efforts to identify high-yield conditions.
• Current methods often involve a heavy experimental load, limiting efficiency.

Purpose of the Study:

• To develop an efficient machine learning tool, RS-Coreset, for predicting chemical reaction yields.
• To significantly reduce the experimental data required for reaction space exploration.
• To guide an interactive procedure for representing the full reaction space using deep representation learning.

Main Methods:

• Utilizing deep representation learning techniques within an interactive procedure.
• Developing the RS-Coreset tool to represent the full reaction space.
• Employing small-scale datasets (2.5% to 5% of instances) for yield prediction.

Main Results:

• Achieved state-of-the-art performance on three public datasets.
• Demonstrated the ability to predict reaction yields accurately using minimal data.
• Successfully applied RS-Coreset to assist in exploring Lewis base-boryl radicals enabled dechlorinative coupling reactions.

Conclusions:

• RS-Coreset offers an efficient approach to reaction optimization, drastically reducing experimental workload.
• The tool can predict reaction yields and identify previously overlooked feasible reaction combinations.
• This machine learning approach advances chemical research and industrial production by enabling faster and more effective exploration of reaction spaces.