Abstract:Machine learning interatomic potentials (MLIPs) have become a workhorse of modern atomistic simulations, and recently published universal MLIPs, pre-trained on large datasets, have demonstrated remarkable accuracy and generalizability. However, the computational cost of MLIPs limits their applicability to chemically disordered systems requiring large simulation cells or to sample-intensive statistical methods. Here, we report the use of continuous and differentiable alchemical degrees of freedom in atomistic materials simulations, exploiting the fact that graph neural network MLIPs represent discrete elements as real-valued tensors. The proposed method introduces alchemical atoms with corresponding weights into the input graph, alongside modifications to the message-passing and readout mechanisms of MLIPs, and allows smooth interpolation between the compositional states of materials. The end-to-end differentiability of MLIPs enables efficient calculation of the gradient of energy with respect to the compositional weights. Leveraging these gradients, we propose methodologies for optimizing the composition of solid solutions towards target macroscopic properties and conducting alchemical free energy simulations to quantify the free energy of vacancy formation and composition changes. The approach offers an avenue for extending the capabilities of universal MLIPs in the modeling of compositional disorder and characterizing the phase stabilities of complex materials systems.

What problem does this paper attempt to address?

This paper discusses the introduction of continuous differentiable chemical degrees of freedom into machine learning inter-atomic potentials (MLIPs) to address the computational cost issue in large-scale simulation and statistical methods of chemically disordered systems. The authors leverage the property of discrete elements represented as real-valued tensors in graph neural networks MLIPs and propose a method to add chemical atoms and their corresponding weights in the input graph structure, and adjust the message passing and energy reading mechanisms. This approach allows for smooth interpolation between material composition states and efficient computation of energy gradients with respect to composition weights through end-to-end differentiability. Using these gradients, the paper proposes a strategy to optimize the composition of solid solutions towards desired macroscopic properties and performs chemical free energy simulations to quantify the free energy of vacancy formation and composition changes. This method extends the capabilities of general MLIPs in modeling phase stability of disordered and complex material systems. The main contributions of the paper include: 1. Introducing chemical atoms and weights to enable smooth interpolation between different material compositions in MLIPs. 2. Efficiently computing energy gradients with respect to composition weights through end-to-end differentiability for optimizing solid solution composition and conducting chemical free energy simulations. 3. Applying to predict properties of solid solution solutions such as lattice parameters and optimizing these properties by optimizing the composition. 4. Evaluating the formation free energy of point defects (such as vacancies) through chemical free energy calculations, providing a new computational pathway with higher statistical efficiency compared to traditional methods. In summary, the paper aims to enhance the capabilities of MLIPs in handling chemically disordered systems by improving them, thereby achieving more efficient and accurate calculations in material science simulation and design.

Interpolation and differentiation of alchemical degrees of freedom in machine learning interatomic potentials

Deep machine learning potentials for multicomponent metallic melts: Development, predictability and compositional transferability

Transferability and Accuracy of Ionic Liquid Simulations with Equivariant Machine Learning Interatomic Potentials

Improving Molecular Dynamics Simulations for Solid‐Liquid Interface with Machine Learning Interatomic Potentials

Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations

Transferability of datasets between Machine-Learning Interaction Potentials

Systematic assessment of various universal machine-learning interatomic potentials

Evaluating and improving the predictive accuracy of mixing enthalpies and volumes in disordered alloys from universal pre-trained machine learning potentials

Transferable Water Potentials Using Equivariant Neural Networks

Machine Learning Interatomic Potentials and Long-Range Physics

Recent advances in machine learning interatomic potentials for cross-scale computational simulation of materials

Data-efficient multi-fidelity training for high-fidelity machine learning interatomic potentials

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

Towards High-Throughput Microstructure Simulation in Compositionally Complex Alloys Via Machine Learning

Robust training of machine learning interatomic potentials with dimensionality reduction and stratified sampling

Learning Together: Towards foundational models for machine learning interatomic potentials with meta-learning

Learning Interatomic Potentials at Multiple Scales

Automated atomistic simulations of dissociated dislocations with ab initio accuracy

Accuracy, Performance, and Transferability of Interparticle Potentials for Al–Cu Alloys: Comparison of Embedded Atom and Deep Machine Learning Models