Machine Learning of Potential-Energy Surfaces Within a Bond-Order Sampling Scheme

Predicting the values of the potential energy surface (PES) for a given chemical system is essential to running the associated dynamics and modeling its evolution in time. To the purpose of modeling chemical reactions involving few atoms, this task is usually accomplished by fitting or interpolating a set of energies computed at different nuclear geometries through accurate, though computationally demanding, quantum-chemical calculations. Among the several approaches for choosing an appropriate set of geometries and energies, a new scheme has been recently proposed (Rampino S, J Phys Chem A 120:4683-4692, 2016) which is based on a regular sampling in a space-reduced bond-order (SRBO) domain rather than in the more conventional bond-length (BL) domain. In this work we address the performances of four machine-learning (ML) models, as opposed to pure mathematical fitting or interpolation schemes, in predicting the PES of a three-atom system modeling an atom-diatom exchange reaction when coupled to the SRBO sampling scheme. The models (two ensemble-learning, an automated ML, and a deep-learning one), trained on both SRBO and BL datasets, are shown to perform better than popular fitting or interpolation schemes and to give the best results if coupled to SRBO data.