EXTENDED ABSTRACT: Predicting organic reaction feasibility and robustness against environmental factors is a longstanding challenging. We address this by integrating high throughput experimentation and Bayesian deep learning. Unlike prior studies focused on niche chemical spaces, our platform conducted 11,669 distinct acid-amine condensation reactions in 156 working hours, yielding the largest single reaction type dataset at a volumetric scale for industrial delivery. Our Bayesian neural network model achieved a new benchmark with 89.48\% accuracy for reaction feasibility prediction. Furthermore, our fine-grained uncertainty disentanglement enables efffcient active learning, reducing 80% of data requirements. Our uncertainty analysis effectively identifies out-of-domain reactions and evaluates reaction robustness against environmental factors, offering a practical framework for navigating chemical spaces and designing highly robust reactions for industrial processes. This work has been integrated into the industrial development pipeline of ChemLex.

Keywords: High-Throughput Experimentation; Bayesian Learning; Organic Chemistry;

REFERENCES: [1] H. Zhong et al, ChemRxiv, doi:10.26434/chemrxiv-2024-gt72l, 2024 (preprint)