Abstract
Pd-catalyzed C-N couplings are commonplace in academia and industry. Despite their significance, finding suitable reaction conditions leading to a high yield, for instance, remains a challenging and time-consuming task which usually requires screening over many sets of conditions. To help select promising reaction conditions in the vast space of reagent combinations, machine learning is an emerging technique with a lot of promise. In this work, we assess whether the reaction yield of C-N couplings can be predicted from databases of chemical reactions. We test the generalizability of models both on challenging data splits and on a dedicated experimental test set. We find that, provided the chemical space represented by the training set is not left, the models perform well. However, the applicability domain is quickly left even for simple reactions of the same type, as, for instance, present in our plate test set. The results show that yield prediction for new reactions is possible from the algorithmic side but in practice is hindered by the available data. Most importantly, more data that cover the diversity in reagents are needed for a general-purpose prediction of reaction yields. Our findings also expose a challenge to this field in that it appears to be extremely deceiving to judge models based on literature data with test sets which are split off the same literature data, even when challenging splits are considered.