Reaction optimization has historically been done from a “cold-start”, using a trial-and-error approach uninformed by historical data. For reaction optimisation, we leveraged a method, known as multi-task Bayesian Optimization (MTBO), which learns from both historical reaction data as well as from new experiments. MTBO identifies optimal reaction conditions far faster than standard (single-task) Bayesian Optimization (STBO), which is not data-informed.
Bayesian Optimization (BO) offers a powerful and data-efficient way to navigate complex chemical reaction spaces. BO works by constructing a statistical model which is continually updated and improved with additional data. This model captures both the outcome (such as reaction yield) and the uncertainty across the design space.
At the start. when no data is available, uncertainty is high across the entire design space, so BO must balance exploration (learning the shape of the space) with exploitation (testing the most promising conditions).
A key insight is that this initial “map” of the reaction landscape doesn’t need to be constructed solely from the target reaction. BO can leverage information from related reactions to reduce early uncertainty and accelerate optimization. This principle underpins multi-task Bayesian Optimization (MTBO), allowing the model to learn from analogous chemical data to achieve faster, more efficient optimization.
• Upload prior data of similar/related experiments.
• Map the prior data to your target task.
• Receive intelligent predictions from the start.
Sample Processing
Filter Substrate Scope
Execute
