Enhanced Sampling Path Integral Methods Using Neural Network Potential Energy Surfaces with Application to Diffusion in Hydrogen Hydrates

Abstract

The high cost of ab initio molecular dynamics (AIMD) simulations to model complex physical and chemical systems limits its ability to address many key questions. However, new machine learning‐based representations of complex potential energy surfaces have been introduced in recent years to circumvent computationally demanding AIMD simulations while retaining the same level of accuracy. As these machine learning methods gain in popularity over the next decade, it is important to address the appropriate way to develop and integrate them with well‐established simulation methods. This paper details the parameterization and training, using accurate electronic structure calculations, of artificial neural network potentials (NNPs) to model the intermolecular interactions in a hydrogen clathrate hydrate, wherein hydrogen molecules are confined inside the cavities of the 3D crystalline framework of hydrogen‐bonded water molecules. This new NNP is used in conjunction with new path‐integral based enhanced sampling methods, for inclusion of nuclear quantum effects and promotion of free energy barrier crossing, in order to determine properties that are important for understanding the diffusion of hydrogen gas in the clathrate system. These simulations demonstrate the influence of cage occupancy on the free energy barriers that determine the diffusivity of hydrogen gas through the network of large cages.

Forecasting nonadiabatic dynamics using hybrid convolutional neural network/long short-term memory network

“…Another option for obtaining the dynamical map is via machine learning (ML). Deep neural networks (NN) have been applied in many fields successfully, such as image recognition, 30,31 natural language processing, 32 potential energy surface fitting, 33 quantitative structureproperty relationship, 34 quantum machine learning, 35 and many more. 36 For learning time series data, convolutional neural network (CNN) 30,31 and long short-term memory (LSTM) 37,38 have been serving as the two main frameworks.…”

Hydrogen and Deuterium Molecular Escape from Clathrate Hydrates: “Leaky” Microsecond-Molecular-Dynamics Predictions

“… 45 Therefore, more elevated concentrations need to overcome larger free-energy barriers to jump from one cage to another, with all large cages having high occupation. 9 , 41 , 42 , 46 , 47 …”

“…Certainly, in the case of quadruple large-cage occupation, the level of molecular “crowding” is quite elevated, given that there is a growing experimental and simulation body of consensus that large cages are typically no more than doubly occupied . Therefore, more elevated concentrations need to overcome larger free-energy barriers to jump from one cage to another, with all large cages having high occupation. ,,,, …”

Affordable Ab Initio Path Integral for Thermodynamic Properties via Molecular Dynamics Simulations Using Semiempirical Reference Potential

“…41 Affordable studies are limited to relatively simple systems, 42 low levels of theory, 43 or with the aid of artificial intelligence. 44 Therefore, it is desirable to develop a practical method for the study of thermodynamics properties based on PIMD that can dramatically reduce the computational expense while maintain the accuracy. The multistate thermodynamic perturbation (MsTP) is the method of choice, of which the efficiency and reliability has been examined by Li et al in an early study of some model reactions.…”