Abstract:The introduction of modern Machine Learning Potentials (MLP) has led to a paradigm change in the development of potential energy surfaces for atomistic simulations. By providing efficient access to energies and forces, they allow to perform large-scale simulations of extended systems, which are not directly accessible by demanding first-principles methods. In these simulations, MLPs can reach the accuracy of electronic structure calculations provided that they have been properly trained and validated using a suitable set of reference data. Due to their highly flexible functional form the construction of MLPs has to be done with great care. In this tutorial, we describe the necessary key steps for training reliable MLPs, from data generation via training to final validation. The procedure, which is illustrated for the example of a high-dimensional neural network potential, is general and applicable to many types of MLPs.

What problem does this paper attempt to address?

This paper is a tutorial that primarily discusses how to train Neural Network Potentials (NNPs). The researchers point out that modern Machine Learning Potentials (MLPs) have revolutionized the development of Potential Energy Surfaces (PES), enabling atomistic simulations of large systems that are difficult to handle directly using traditional first-principles methods. Despite the flexibility of MLPs, careful construction is required to ensure proper training and validation. The tutorial provides a detailed overview of the key steps involved in generating data, training, and validating MLPs, using high-dimensional neural network potentials as an example. The success of MLPs lies in their ability to learn atomic interactions from reference data obtained from electronic structure calculations and provide energy and forces with accuracy approaching that of reference methods. However, their extrapolation capability is limited, and they typically require a large and diverse training dataset. The paper also mentions different generations of MLPs, including the first generation (limited to low-dimensional systems), the second generation (introducing atomic center symmetry functions applicable to high-dimensional systems), as well as the third and fourth generations, which incorporate long-range charge and dispersion interactions, respectively. Despite the excellent performance of MLPs in describing various chemical bonds and interactions, their computational costs are often higher than those of simple classical force fields. The generation of datasets, neural network settings, training processes, and validation are crucial aspects in constructing neural network potentials. The authors emphasize that building MLPs is not a simple task and requires users to understand the applicability and limitations of the underlying datasets. Finally, the tutorial provides a detailed flowchart outlining the entire process of training neural network potentials, illustrated using a simple model system of lithium-hydroxide hydrate. In summary, this paper aims to fill the gap in tutorials for MLP training and validation, helping researchers construct and apply neural network potentials effectively in large-scale atomistic simulations.

Tutorial: How to Train a Neural Network Potential

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

The Potential of Neural Network Potentials

Considerations in the use of ML interaction potentials for free energy calculations

Strategies for the Construction of Machine-Learning Potentials for Accurate and Efficient Atomic-Scale Simulations

Training Machine Learning Potentials for Reactive Systems: A Colab Tutorial on Basic Models.

Machine Learning Potentials with the Iterative Boltzmann Inversion: Training to Experiment

Scalable Training of Neural Network Potentials for Complex Interfaces Through Data Augmentation

Force Training Neural Network Potential Energy Surface Models

Machine learning potentials with Iterative Boltzmann Inversion: training to experiment

Introduction to machine learning potentials for atomistic simulations

chemtrain: Learning Deep Potential Models via Automatic Differentiation and Statistical Physics

De novo exploration and self-guided learning of potential-energy surfaces

A Generator of Neural Network Potential for Molecular Dynamics: Constructing Robust and Accurate Potentials with Active Learning for Nanosecond-scale Simulations

Machine Learning of Reactive Potentials

A Hessian-Based Assessment of Atomic Forces for Training Machine Learning Interatomic Potentials

Synthetic pre-training for neural-network interatomic potentials

Neural Network Potentials for Chemistry: Concepts, Applications and Prospects

Learning Interatomic Potentials at Multiple Scales

Peering inside the black box: Learning the relevance of many-body functions in Neural Network potentials

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