Core Concepts
This work introduces a method to access the alchemical degrees of freedom inherent in machine learning interatomic potentials, enabling smooth interpolation between different compositional states of materials and efficient calculation of alchemical gradients.
Abstract
The content describes a method to introduce continuous and differentiable alchemical degrees of freedom in atomistic materials simulations using machine learning interatomic potentials (MLIPs). The key points are:
Alchemical Graph Augmentation:
The original atomic graph is augmented by introducing alchemical atoms with corresponding weights, representing different compositional states.
Edges are connected between alchemical atoms and non-alchemical atoms based on a defined scheme.
An asymmetric weighting scheme is used to scale the message contributions during message passing.
Alchemical Message Passing and Energy Readout:
The message passing mechanism is modified to incorporate the alchemical weights during the message aggregation step.
The energy readout is also modified to perform a weighted sum of the alchemical node contributions.
Applications:
Modeling of solid solutions: The method can capture nonlinear deviations in lattice parameters and enable efficient optimization of compositions to match target properties.
Free energy calculations: The end-to-end differentiability of the model with respect to alchemical weights allows efficient calculation of free energy differences between compositional states, demonstrated for the free energy of vacancy formation in BCC iron and the relative phase stabilities of CsPbI3 and CsSnI3.
The proposed approach offers an efficient way to extend the capabilities of universal MLIPs in modeling compositional disorder and characterizing the phase stabilities of complex materials systems.