Improving accuracy of interatomic potentials: more physics or more data? A case study of silica

Abstract

In this paper we test two strategies to improving the accuracy of machinelearning potentials, namely adding more fitting parameters thus making use of large volumes of available quantum-mechanical data, and adding a chargeequilibration model to account for ionic nature of the SiO 2 bonding. To that end, we compare Moment Tensor Potentials (MTPs) and MTPs combined with the charge-equilibration (QEq) model (MTP+QEq) fitted to a density functional theory dataset of α-quartz SiO 2 -based structures. In order to make a meaningful comparison, in addition to the accuracy, we assess the uncertainty of predictions of each potential. It is shown that adding the QEq model to MTP does not make any improvement over the MTP potential alone, while adding more parameters does improve the accuracy and uncertainty of its predictions.

Machine‐learning‐based interatomic potentials for advanced manufacturing

“…92 MTPs combined with the chargeequilibration (MTP+QEq) cannot improve MTP potentials alone for alpha-quartz SiO 2 -based structures. 93 The method combining ab initio MD simulations and Bayesian optimization agrees well with the ab initio simulations by considering the example of glassy silica. 94 A statistical learning model was trained using the least absolute shrinkage and selection operator with a gradient boost machine (GBM-LASSO), which only needs to be trained using a data set consisting of only binary and ternary glass samples, 95 as shown in Figure 10.…”

“…An ANN model for silicon‐carbon systems and ceramic matrix composites (CMCs) based on silicon carbide is proposed 92 . MTPs combined with the charge‐equilibration (MTP+QEq) cannot improve MTP potentials alone for alpha‐quartz SiO 2 ‐based structures 93 . The method combining ab initio MD simulations and Bayesian optimization agrees well with the ab initio simulations by considering the example of glassy silica 94 …”

Uncertainty quantification in molecular simulations with dropout neural network potentials

“…A quantity of interest (QOI) is computed separately using each member of the ensemble and the spread in these predictions provides an estimate of uncertainty. The methods differ in how they construct the ensemble: (1) a single IP form is fit to partial datasets drawn at random from the full training set (i.e., training set subsampling) [16][17][18] ; (2) different IP forms are constructed and fit to the same training set 19,20 ; and (3) the same IP form is fit to the same training set, but using different initial values of IP parameters 21,22 . Another class of uncertainty estimation methods relies on a measure of the distance between an evaluated configuration and the training set.…”

Machine-learning interatomic potentials for materials science

“…Rotationally invariant descriptors are obtained from contractions of these tensors to produce scalars. These descriptors are used as part of the moment tensor potentials (MTP) [90][91][92][93][94]. In the recent tests of several ML potentials [24], the MTP potentials have shown the optimal combination of accuracy and computational efficiency.…”