Machine Learning\na Simple Interpretable Short-Range\nPotential for Silica

A wide array of models, spanning from computationally\nexpensive\nab initio methods to a spectrum of force-field approaches, have been\ndeveloped and employed to probe silica polymorphs and understand growth\nprocesses and atomic-level dynamical transitions in silica. However,\nthe quest for a model capable of making accurate predictions with\nhigh computational efficiency for various silica polymorphs is still\nongoing. Recent developments in short-range machine-learned models,\nsuch as GAP and NNPScan, have shown promise in providing reasonable\ndescriptions of silica, but their computational cost remains high\ncompared to force fields such as BKS which are based on simple interpretable\nfunctional forms. Here, we build on the recent success of our reinforcement\nlearning (RL) workflow to derive a new set of optimal parameters for\na promising short-range BKS-based model proposed by Soules. We use\nRL to navigate the eight-dimensional parameter space of the Soules\npotential using an experimental training data set that includes both\nlocal and global structural features from approximately 21 experimentally\nrealized silica polymorphs, including high density phases and porous\nzeolites. We compare the performance of our machine-learned ML-Soules\nmodel with other high quality models including our recent machine-learned\nparametrization of BKS (ML-BKS), a machine-learned potential (GAP),\nas well as predictions of ab initio calculations with the highly fidelity\nSCAN functional. The ML-Soules accurately captures the relative energetic\nordering of various polymorphs as well as their structural features\nat a significantly reduced computational expense. The ML-Soules model\nalso reasonably captures the structure, density, and elastic constants\nof quartz, as well as metastable silica polymorphs. We further discuss\nthe limitations of the Soules functional form and propose potential\nenhancements, including the incorporation of additional three-body\nterms and/or the utilization of different short-ranged functional\nforms to achieve greater accuracy for both global and local features\nin the modeling of silica while retaining low computational cost.