First-Principles-Based Machine-Learning Molecular Dynamics for Crystalline Polymers with van der Waals Interactions

Abstract

Machine-learning (ML) techniques have drawn an ever-increasing focus as they enable high-throughput screening and multiscale prediction of material properties. Especially, ML force fields (FFs) of quantum mechanical accuracy are expected to play a central role for the purpose. The construction of ML-FFs for polymers is, however, still in its infancy due to the formidable configurational space of its composing atoms. Here, we demonstrate the effective development of ML-FFs using kernel functions and a Gaussian process for an organic polymer, polytetrafluoroethylene (PTFE), with a data set acquired by first-principles calculations and ab initio molecular dynamics (AIMD) simulations. Even though the training data set is sampled only with short PTFE chains, structures of longer chains optimized by our ML-FF show an excellent consistency with density functional theory calculations. Furthermore, when integrated with molecular dynamics simulations, the ML-FF successfully describes various physical properties of a PTFE bundle, such as a density, melting temperature, coefficient of thermal expansion, and Young's modulus.

From Organic Fragments to Photoswitchable Catalysts: The OFF–ON Structural Repository for Transferable Kernel-Based Potentials

“…Machine learning (ML)-based potentials have served as efficient ab initio surrogates for molecular dynamics (MD) simulations, enabling statistically converged free energy surfaces (FESs) to be computed at DFT-level accuracy with significantly reduced computational cost. − Commonly used architectures include artificial neural networks (ANN), − Gaussian approximation models (GAP), − and, more generally, kernel-based methods. − The success and robustness of these surrogate potentials rely on many aspects, with the curation of chemically and structurally diverse databases to train the underlying ML models being of utmost importance for achieving size-extensive extrapolations and transferable predictions. Generating such training sets requires not only a thorough mapping of chemical space but also the inclusion of out-of-equilibrium structures that cover the necessary energies and forces for MD simulations.…”

Transferring Chemical and Energetic Knowledge Between Molecular Systems with Machine Learning

“…In the last decade, first attempts to employ ML in molecular simulations have appeared. In particular, ML has been used to predict atomistic properties in molecular systems [9,21,28,36,39], also using first principle calculations (i.e., quantum mechanics) [8,19,30], and, more recently, in identification of free-energy states and slow degrees of freedom in molecular systems [35,40]. However, the wealth of data represents a major limitation for ML applications and despite the increasing computing power, the sampling capability of a system's phase space still represents a hindering factor in all ML applications to molecular simulations.…”

Symbolic Transformer Accelerating Machine Learning Screening of Hydrogen and Deuterium Evolution Reaction Catalysts in MA2Z4 Materials

“…Density functional theory (DFT) calculation combined with a machine learning (ML) model is a powerful tool for developing the advanced force fields of molecular dynamics , and predicting the property of targeted materials in the catalysis field, , which can provide a deep insight into the internal laws and accelerate the screening of candidate catalysts . It not only ensures the accuracy of DFT-calculated catalyst performance but also shortens the calculation time during the high-throughput screening of catalysts in various reactions, such as oxygen reduction reaction, nitrogen reduction reaction and hydrogen evolution reaction.…”