Machine-learned interatomic potentials by active learning: amorphous and liquid hafnium dioxide

Abstract

We propose an active learning scheme for automatically sampling a minimum number of uncorrelated configurations for fitting the Gaussian Approximation Potential (GAP). Our active learning scheme consists of an unsupervised machine learning (ML) scheme coupled with a Bayesian optimization technique that evaluates the GAP model. We apply this scheme to a Hafnium dioxide (HfO 2) dataset generated from a "melt-quench" ab initio molecular dynamics (AIMD) protocol. Our results show that the active learning scheme, with no prior knowledge of the dataset, is able to extract a configuration that reaches the required energy fit tolerance. Further, molecular dynamics (MD) simulations performed using this active learned GAP model on 6144 atom systems of amorphous and liquid state elucidate the structural properties of HfO 2 with near ab initio precision and quench rates (i.e., 1.0 K/ps) not accessible via AIMD. The melt and amorphous X-ray structural factors generated from our simulation are in good agreement with experiment. In addition, the calculated diffusion constants are in good agreement with previous ab initio studies.

Automated discovery of a robust interatomic potential for aluminum

“…The basic steps of active learning (AL) for atomic-scale modeling are to sample new atomic configurations, query the ML model for uncertainty in its predictions, and selectively collect new training data that would best improve the model 24 , 35 – 40 . Previous work employed AL to drive nonequilibrium sampling of large datasets through organic chemical space, yielding the highly general ANI-1x potential 41 .…”

Short range order and network connectivity in amorphous AsTe3: a first principles, machine learning, and XRD study

“…This technique has been successfully applied to model glasses, liquids, and crystals. [63][64][65][66] In this work, we apply ML-GAP technique as implemented in QUIP (http://www.libatoms.org). We combine two descriptors for the representation of the atomic structure where the Smooth Overlap of Atomic Positions (SOAP) 67 is complemented by a nonparametric two-body distance descriptor in order to prevent nonphysical clustering of atoms.…”

A Transferable Active-Learning Strategy for Reactive Molecular Force Fields

“…Within small chemical subspaces, models can be achieved using neural networks (NNs), 6,[16][17][18][19][20] kernel-based methods such as Gaussian processes (GP) 21,22 and gradient-domain machine learning (GDML), 23 and linear fitting with properly chosen basis functions, 24,25 each with different data requirements and transferability. 26 In the present work, we employ the Gaussian Approximation Potential (GAP) framework, 21 which has been used to generate force fields for a range of elemental, [27][28][29] multicomponent inorganic, 30,31 and recently gas-phase organic systems. 14,32 Initial studies of condensed phase molecular systems with GAPs include fluid methane 33 and phosphorus.…”

“…Active learning (AL) has started to emerge as one of the key strategies for generating reference databases and accelerating the fitting process. 30,[37][38][39][40][41] Notable examples in the modelling of materials include an early demonstration of a "query-by-committee" approach in fitting a high-dimensional NN potential for elemental copper, 38 the fitting of Moment Tensor Potential (MTP) 25 models based on the D-optimality criterion 42 which has been applied, e.g., to the prediction of elemental crystal structures 37 and multicomponent alloys, 39 and the deep potential generator (DP-GEN) 43,44 that provides an interface to deep neural network potential models. 45 More recently, AL approaches have also been combined with GP based force fields including GAP, 46 and included within a first-principles MD implementation such that it allows the "on the fly" fitting of force fields for a specific simulation system.…”