Flexible Molecules Need More Flexible Machine Learning Force Fields

Robust machine learning (ML) models should reliably predict molecular potential-energy surfaces (PES), including equilibrium and transition regions, using limited amounts of reference data. We assess the performance of state-of-the-art ML models, namely sGDML, SchNet, GAP/SOAP, and BPNN, in solving this task on an example of cis to trans thermal relaxation in molecular switches exemplified by the azobenzene molecule. Our results demonstrate difficulties all the employed methods face in this task. The local ML models, GAP/SOAP, and BPNN show large errors caused by the limitations of descriptors in learning long-range interactions. The global model, sGDML, demonstrates the best accuracy, but it strongly depends upon the training set and the choice of descriptor. Moreover, the optimal descriptor is different for different transition mechanisms. Finally, SchNet is the most overall reliable model, but its prediction errors vary for different parts of the PES. Our findings reveal that constructing accurate and data-efficient ML force fields is still an open challenge, requiring further developments. To resolve this, we propose moving from learning the entire PES within a single ML model to the employment of local models that are combined into a global force field.