| Summary: | Machine learning has grown in use for constructing potential energy surfaces for their ability to theoretically recreate any function given enough training as well as their fast predictive powers after being trained. When trained on ab initio data, this enables simulation of a large number of ab-initio-quality trajectories. Here, rigorous benchmarking of these machine-learned potential energy surfaces—both in terms of their static errors and dynamics errors—is carried out for the HBr+ + HCl system. In a novel comparison, both neural networks and a kernel regression method are compared for a global potential energy surface, covering multiple dissociation channels. Further, comparison with ab initio molecular dynamics simulations enables one of the first direct comparisons of dynamic, ensemble-average properties of the system. Finally, comparison with experimental results reveals remarkable agreement for the sGDML method for training sets of thousands to tens of thousands of molecular configurations.
|
|---|