Fig. 5: Machine learning (ML) optimisation of the nickel-catalysed Suzuki reaction.
From: Highly parallel optimisation of chemical reactions through automation and machine intelligence
a The first scatter plot shows the area percent (AP) yield and AP selectivity of experiments selected by our ML Bayesian optimisation (BO) workflow at each iteration. The second plot shows the Pareto points, representing the best trade-offs between AP yield and selectivity, identified at each iteration. The Pareto points illustrate the optimal yield-selectivity combinations achieved. b Using all collected experimental data from the campaign (576 reactions), we trained an ML model to perform Shapley additive explanations (SHAP)39 feature importance analysis. The SHAP values quantify the magnitude of each feature’s contribution to the model’s predicted AP yield, considering both positive and negative impacts to identify the most influential factors (see Methods for more details). The left panel shows the mean absolute SHAP values, while the right panel displays the individual data points that contribute to these means.
