Improving the accuracy of the neuroevolution machine learning potential for multi-component systems

Zheyong Fan

DOI: https://doi.org/10.1088/1361-648X/ac462b

2022-01-24

Abstract:In a previous paper [Fan Z \textit{et al}. 2021 Phys. Rev. B, \textbf{104}, 104309], we developed the neuroevolution potential (NEP), a framework of training neural network based machine-learning potentials using a natural evolution strategy and performing molecular dynamics (MD) simulations using the trained potentials. The atom-environment descriptor in NEP was constructed based on a set of radial and angular functions. For multi-component systems, all the radial functions between two atoms are multiplied by some fixed factors that depend on the types of the two atoms only. In this paper, we introduce an improved descriptor for multi-component systems, in which different radial functions are multiplied by different factors that are also optimized during the training process, and show that it can significantly improve the regression accuracy without increasing the computational cost in MD simulations.

Computational Physics

What problem does this paper attempt to address?

The main objective of this paper is to improve the accuracy of Neural Evolutionary Potentials (NEP) in multi-component systems. The authors propose an improved method for atomic environment descriptors, significantly enhancing regression accuracy by optimizing different radial functions multiplied by different coefficients, without increasing the computational cost in molecular dynamics (MD) simulations. Specifically, for multi-component systems, the traditional NEP method (referred to as NEP1) uses fixed factors between all radial functions that depend only on the types of the two atoms. In the improved method (referred to as NEP2), these factors are also optimized during the training process, and different radial functions are multiplied by different factors. In this way, NEP2 can better distinguish between different types of atoms, thereby improving the description capability for multi-component systems. The paper validates the effectiveness of this improvement through two case studies: PbTe crystals and Al-Cu-Mg alloys. The results show that NEP2 reduces the prediction error in energy and forces by approximately 30% compared to NEP1, achieving higher regression accuracy while maintaining high computational speed. Additionally, by performing non-equilibrium molecular dynamics simulations to calculate the thermal conductivity of PbTe, it is demonstrated that the high accuracy of NEP2 is not a result of overfitting.

Improving the accuracy of the neuroevolution machine learning potential for multi-component systems

Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport

GPUMD: A package for constructing accurate machine-learned potentials and performing highly efficient atomistic simulations

Combining the D3 dispersion correction with the neuroevolution machine-learned potential

Incorporating Neural Networks into the AMOEBA Polarizable Force Field

Neuroevolution of Physics-Informed Neural Nets: Benchmark Problems and Comparative Results

Highly efficient path-integral molecular dynamics simulations with GPUMD using neuroevolution potentials: Case studies on thermal properties of materials

Towards Physical Plausibility in Neuroevolution Systems

Accurate and efficient molecular dynamics based on machine learning and non von Neumann architecture

Molecular dynamics simulations of heat transport using machine-learned potentials: A mini review and tutorial on GPUMD with neuroevolution potentials

Accurate nuclear quantum statistics on machine-learned classical effective potentials

Deep Potential Molecular Dynamics: a scalable model with the accuracy of quantum mechanics

Global Neural Network Potential with Explicit Many-Body Functions for Improved Descriptions of Complex Potential Energy Surface.

Committee neural network potentials control generalization errors and enable active learning

NNP/MM: Accelerating molecular dynamics simulations with machine learning potentials and molecular mechanic

Tensorial properties via the neuroevolution potential framework: Fast simulation of infrared and Raman spectra

Reweighting from Molecular Mechanics Force Fields to the ANI-2x Neural Network Potential

Neural Network Potential with Multi-Resolution Approach Enables Accurate Prediction of Reaction Free Energies in Solution

Molecular Dynamics with Neural-Network Potentials

A2I Transformer: Permutation-equivariant attention network for pairwise and many-body interactions with minimal featurization