Topic: Building a reaction database using high-throughput technology: Enabling AI chemistry

Abstract: Chemists often fail to predict the outcome of an unseen reaction, recently chemists have shown great interest in developing data-driven AI models for reaction prediction. However, bias towards positive data, data inconsistency and lack of annotation are common problems that exist in both public and proprietary databases. High-Throughput Experimentation (HTE), a technology that uses robotics to rapidly perform many reactions in parallel, has shown great potential in building standardized databases.

Here we present our progress in developing an ultra-HTE platform, constructing a proprietary reaction database and building AI-based reaction prediction models. Our ultra-HTE platform could set up and analyze over 8,000 reactions per day, and almost all common reactions are suitable. To build a standardized reaction database, we are currently using the platform to run reactions and collect data on a daily basis. With the data set in hand, we have developed AI-based models to successfully predict optimal conditions and reaction yields for several commonly used reactions as well as new reactions. Data collection and AI-based model development for more reactions is ongoing in our team.