(521du) Unlocking the Power of Machine Learning in Catalytic Reaction Optimization

Rough estimations suggest that a typical catalytic reaction optimization could have a total reaction space of up to 52 million different possible combinations. Traditional methods are not well-suited for the optimization of multiple reaction parameters simultaneously, and as a consequence, optimization campaigns are often lengthy, expensive, and wasteful, requiring multiple rounds of experimentation and forcing scientists to rely heavily on intuition and previous experience. Thus making it a very challenging task for beginners and almost impossible for experienced researchers to know how efficient their optimizations truly are. SuntheticsML is an online platform specifically designed to address these challenges and help R&D scientists optimize catalytic reactions with the help of proprietary machine-learning algorithms based on Bayesian optimization (BO).

Bayesian optimization is a probabilistic approach that uses Bayesian inference to model the underlying function being optimized. While BO has shown promising results for numerical variables, it is less effective for categorical variables. Categorical features significantly increase dimensionality in optimization approaches given that every variable (e.g., solvent) may have different possible values (e.g., ethanol, methanol, toluene), and typical optimization approaches (i.e., one hot encoding) rely on introducing a new dimension to the optimization for each categorical value. In other words, optimizing the variable “solvent” will require optimizing as many additional dimensions as the number of potential solvents considered.

In this talk we will present recent advances in the development of ML algorithms and parametrization techniques tailored to the optimization of multitarget categorical features. Highlighting the power of machine learning and chemical knowledge to accelerate reaction optimization using small data. Starting with as little as 5 data points, SuntheticsML algorithms are able to reach unprecedented efficiencies requiring up to 15x less time and experiments. We will also discuss a wide range of recent case studies featuring catalytical applications, including cross coupling, mechanochemistry and electrochemistry.