Towards global reaction feasibility and robustness prediction with high throughput data and bayesian deep learning.

Area of Science:

• Organic Chemistry
• Chemical Engineering
• Machine Learning

Background:

• Predicting organic reaction feasibility and robustness is a significant challenge in chemical development.
• Existing high throughput experimentation (HTE) studies often focus on limited chemical spaces.

Purpose of the Study:

• To develop a practical framework for predicting organic reaction feasibility and robustness using integrated HTE and Bayesian deep learning.
• To create the most extensive single HTE dataset for acid amine coupling reactions at a volumetric scale.

Main Methods:

• Conducted 11,669 distinct acid amine coupling reactions using an in-house HTE platform over 156 hours.
• Developed and applied a Bayesian neural network model for reaction feasibility prediction.
• Implemented fine-grained uncertainty disentanglement for active learning and out-of-domain detection.

Main Results:

• Achieved a benchmark prediction accuracy of 89.48% for reaction feasibility.
• Demonstrated an 80% reduction in data requirements through efficient active learning.
• Successfully identified out-of-domain reactions and evaluated process robustness against environmental factors.

Conclusions:

• The integrated HTE and Bayesian deep learning approach provides a powerful tool for navigating complex chemical spaces.
• This framework enables the design of highly robust industrial chemical processes with reduced experimental effort.
• The study offers a practical solution for predicting reaction outcomes and assessing scalability.