Design and Experimental Validation of a Photocatalyst Recommender Based on a Large Language Model.

Area of Science:

• Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Photocatalysis is crucial for chemical transformations.
• Identifying optimal photocatalysts is often challenging and time-consuming.
• Data-driven approaches can accelerate catalyst discovery.

Purpose of the Study:

• To develop a machine learning (ML) model for recommending suitable photocatalysts.
• To leverage a large literature dataset for training the ML model.
• To provide a freely accessible tool for researchers.

Main Methods:

• Trained an ML model on over 36,000 literature examples of photocatalytic reactions.
• Employed a Bidirectional Encoder Representations from Transformer (BERT)-inspired architecture.
• Validated the model's accuracy using cross-validation and experimental testing.

Main Results:

• The ML model achieved approximately 90% accuracy in suggesting appropriate photocatalysts.
• Experimental validation showed the model's suggested catalysts yielded results competitive with human experts.
• The model identified alternative, potentially superior photocatalysts for tested reactions.

Conclusions:

• The developed ML platform is a valuable tool for photocatalyst recommendation and reaction optimization.
• This AI-driven approach can accelerate the discovery and application of photocatalysts.
• The accessible online tool empowers researchers in synthetic chemistry.