Abstract:Long-range interactions are crucial in determining the behavior of chemical systems in various environments. Accurate predictions of physical and chemical phenomena at the atomic level hinge on accurate modeling of these interactions. Here, we present a framework that substantially enhances the predictive power of machine learning interatomic potentials by incorporating explicit polarizable long-range interactions with an equivariant graph neural network short-range potential. The pretrained universal model, applicable across the entire periodic table, can achieve first-principles accuracy. This versatile model has been further applied to diverse areas of research, including the study of mechanical properties, ionic diffusivity in solid-state electrolytes, ferroelectricity, and interfacial reactions, demonstrating its broad applicability and robustness.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the predictive ability of machine - learning inter - atomic potential functions in simulating long - range interactions. Specifically, the authors propose a new framework to enhance the predictive ability of general machine - learning potential functions by combining short - range equivariant graph neural network potential functions with explicitly polarized long - range interactions. This framework aims to address the shortcomings of existing machine - learning potential functions in handling long - range interactions, especially the influence of the dynamic changes of charge transfer and polarization effects on long - range interactions. Through this method, the pre - trained general model can achieve an accuracy close to that of ab - initio methods while maintaining a low computational cost, and is suitable for the simulation of all elements in the periodic table. ### Main Contributions 1. **Framework Innovation**: Propose a new framework that integrates short - range equivariant graph neural network potential functions and long - range polarized charge - balancing methods. 2. **Performance Improvement**: Significantly improve the accuracy of the model in predicting material properties (such as mechanical properties, ion diffusivity, ferroelectric phase transitions, etc.) by introducing long - range interactions. 3. **Wide Application**: Demonstrate the application potential of this model in multiple fields, including mechanical property prediction, lithium - ion diffusion in solid - state electrolytes, phase transitions and interface reactions of ferroelectric materials. ### Specific Problems - **Accurate Modeling of Long - Range Interactions**: Existing machine - learning potential functions have limitations in handling long - range interactions, especially when simulating the behavior of complex materials. This paper solves this problem by introducing the polarized long - range charge - balancing method. - **Balance between Computational Efficiency and Accuracy**: Traditional ab - initio molecular dynamics simulations are highly accurate but computationally expensive. The model proposed in this paper significantly reduces the computational cost while maintaining high accuracy. - **Multi - field Application**: This model not only performs well in a single field but also shows broad application prospects in multiple fields (such as mechanical properties, ion diffusion, ferroelectric phase transitions, etc.). ### Conclusion By combining short - range equivariant graph neural network potential functions with long - range polarized charge - balancing methods, the new framework proposed in this paper significantly improves the predictive ability of machine - learning potential functions in simulating long - range interactions. This model not only performs well in multiple fields but also lays a solid foundation for future high - throughput material screening and in - depth understanding of material behavior. This research provides valuable tools and methods for the development of a new generation of high - performance materials and transformative energy solutions.

Enhancing universal machine learning potentials with polarizable long-range interactions

A Universal Framework for General Prediction of Physicochemical Properties: the Natural Growth Model

Capturing long-range interaction with reciprocal space neural network

Electrostatic interactions in atomistic and machine-learned potentials for polar materials

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

Machine Learning Interatomic Potentials and Long-Range Physics

Heterogeneous Relational Message Passing Networks for Molecular Dynamics Simulations

Systematic assessment of various universal machine-learning interatomic potentials

Self-consistent Coulomb interactions for machine learning interatomic potentials

Extending the atomic decomposition and many-body representation, a chemistry-motivated monomer-centered approach for machine learning potentials

Learning QM/MM Potential using Equivariant Multiscale Model

Constructing accurate and efficient general-purpose atomistic machine learning model with transferable accuracy for quantum chemistry

Long-range dispersion-inclusive machine learning potentials for structure search and optimization of hybrid organic-inorganic interfaces

Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation

A deep potential model with long-range electrostatic interactions

Fast and flexible range-separated models for atomistic machine learning

Overcoming systematic softening in universal machine learning interatomic potentials by fine-tuning

CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling

CHGNet: Pretrained universal neural network potential for charge-informed atomistic modeling

Learning Atomic Multipoles: Prediction of the Electrostatic Potential with Equivariant Graph Neural Networks

Bridging Machine Learning and Thermodynamics for Accurate Pka Prediction