Kavli Affiliate: Ran Wang
| First 5 Authors: Fuchun Ge, Ran Wang, Chen Qu, Peikun Zheng, Apurba Nandi
| Summary:
Machine learning potentials (MLPs) are widely applied as an efficient
alternative way to represent potential energy surface (PES) in many chemical
simulations, e.g., geometry optimizations, frequency calculations, molecular
dynamics, and Monte Carlo computations. However, there is a growing realization
that the evaluation of the reliability of MLPs cannot be reduced to the common
metrics such as the mean absolute (MAE) or root-mean-square (RMSE) errors on
the test set drawn from the same distribution as the training data. Here, we
systematically investigate the relationship between such test errors and the
accuracy of chemical simulations with MLPs on an example of a full-dimensional,
global PES for the glycine amino acid, which is of great significance in both
biology and astronomy. Our results show that, indeed, the errors in the test
set do not unambiguously reflect the MLP performance in different simulation
tasks such as relative conformer energies, barriers, vibrational levels, and
zero-point vibrational energies. Importantly, we also offer a solution and an
easily accessible tool to improve the fidelity of MLPs in a simulation-oriented
manner. Guided by the loss function based on real simulation-oriented metrics,
our solution ensures the quality of MLPs, ultimately yielding the most accurate
one that can even obtain near-zero simulation errors in some tasks such as
predicting relative conformer energies and barriers. Furthermore, this solution
also passed the stringent test in the diffusion Monte Carlo simulations that
confirmed the comprehensive description of the global PES by the resulting MLP.
| Search Query: ArXiv Query: search_query=au:”Ran Wang”&id_list=&start=0&max_results=3