Abstract:We develop a framework for on-the-fly machine learned force field molecular dynamics simulations based on the multipole featurization scheme that overcomes the bottleneck with the number of chemical elements. Considering bulk systems with up to 6 elements, we demonstrate that the number of density functional theory calls remains approximately independent of the number of chemical elements, in contrast to the increase in the smooth overlap of atomic positions scheme.

What problem does this paper attempt to address?

This paper aims to address the bottleneck problem of computational complexity caused by the increase in the number of chemical elements in molecular dynamics simulations. Current methods, such as machine learning force fields based on Smooth Overlap of Atomic Positions (SOAP), are inefficient when dealing with multi-element systems, and the cost increases sharply with the number of elements. The paper proposes a framework based on multipole featureization for real-time machine learning force field simulations, which overcomes the bottleneck related to the number of chemical elements. Experimental results demonstrate that in systems containing up to 6 elements, this method makes the number of density functional theory (DFT) calls roughly independent of the number of elements, and shows improvement compared to the SOAP method. The newly developed method in the paper utilizes Normalized Gaussian Multipole (GMP) descriptors, whose feature vector size is independent of the number of chemical elements, thus reducing the cost of training and inference. This enhances the robustness of real-time training for systems with complex chemistry. By implementing GMP-based real-time potentials in the SPARC quantum chemistry software package, the researchers demonstrate its application in a range of metal and alloy systems, showing that the GMP model can achieve comparable accuracy and stability with lower computational costs compared to the SOAP model. Furthermore, the paper discusses the challenges of hyperparameter optimization, pointing out that inappropriate hyperparameter choices can lead to unstable models and even produce unphysical results. Despite these challenges, the results indicate the potential of the GMP featureization scheme for real-time force fields in multi-element systems, and its scalability to larger-scale simulations, particularly in computing complex properties such as the free energy of high-entropy alloys. Future work will focus on improving the efficiency of machine learning operations, automating hyperparameter selection, and developing more reliable uncertainty quantification and active learning methods.

Overcoming the chemical complexity bottleneck in on-the-fly machine learned molecular dynamics simulations

Efficient Machine Learning Force Field for Large-Scale Molecular Simulations of Organic Systems

Accurate interatomic force field for molecular dynamics simulation by hybridizing classical and machine learning potentials

Utilizing Machine Learning for Efficient Parameterization of Coarse Grained Molecular Force Fields

Accelerating Atomistic Simulations with Piecewise Machine Learned Ab Initio Potentials at Classical Force Field-like Cost

On-the-fly machine learning force field generation: Application to melting points

Simulate Time-integrated Coarse-grained Molecular Dynamics with Multi-Scale Graph Networks

Molecular Dynamics with On-the-Fly Machine Learning of Quantum-Mechanical Forces

Machine Learning Directed Optimization of Classical Molecular Modeling Force Fields

Overcoming the Size Limit of First Principles Molecular Dynamics Simulations with an In-Distribution Substructure Embedding Active Learner

Ab-Initio Molecular Dynamics Acceleration Scheme with an Adaptive Machine Learning Framework

Accelerating atomistic simulations with piecewise machine-learned ab Initio potentials at a classical force field-like cost

Towards Exact Molecular Dynamics Simulations with Machine-Learned Force Fields

Quantum-accurate machine learning potentials for metal-organic frameworks using temperature driven active learning

Parametrization of Non-Bonded Force Field Terms for Metal-Organic Frameworks Using Machine Learning Approach

Towards Electronic Structure-Based Ab-Initio Molecular Dynamics Simulations with Hundreds of Millions of Atoms

Molecular Dynamics with Neural-Network Potentials

Force field optimization by end-to-end differentiable atomistic simulation

Towards Linearly Scaling and Chemically Accurate Global Machine Learning Force Fields for Large Molecules

Machine Learning for Parameter Auto-tuning in Molecular Dynamics Simulations: Efficient Dynamics of Ions near Polarizable Nanoparticles