A machine learning potential with low error in the potential energies does not guarantee good performance for the simulations. One of the reasons is that it is hard to train machine learning potentials with balanced descriptions of different PES regions, especially for global PES data with many strongly distorted molecular geometries that have high deformation energies.

We discuss this problem and show how to solve it by training machine learning potentials to improve performance in simulations rather than on the validation or test set. For this, we have implemented energy-weighting training which can be tuned to get better simulation results. The obtained potentials can be used in heavy diffusion Monte Carlo simulations requiring billions of calculations for accurate anharmonic zero-point vibrational energies.

See our paper in JPCL for more details as well as the tutorials on how to use MLatom for such simulations.

• Fuchun Ge, Ran Wang, Chen Qu, Peikun Zheng, Apurba Nandi, Riccardo Conte, Paul L. Houston, Joel M. Bowman, Pavlo O. Dral. Tell Machine Learning Potentials What They Are Needed For: Simulation-Oriented Training Exemplified for Glycine. J. Phys. Chem. Lett. 2024, 15, 4451–4460. DOI: 10.1021/acs.jpclett.4c00746.