Inverse design of short-range order arrangement via neural network

Abstract

Short-range order (SRO) plays a critical role in the mechanical behavior of metallic alloys. The arrangement of atoms at short ranges significantly impacts how the material responds to external forces. Nevertheless, the mechanics of these phenomena remain poorly understood. The identification of SRO in experiments is constrained by the low resolution of intensity distribution and the limitations associated with the direct observation of atomic arrangements in the lattice within the SRO domains. On the other hand, modeling the mechanical properties of short-range-ordered alloys is challenged by computationally expensive density functional theory (DFT)-based Monte Carlo (MC) simulations required to achieve the SRO structure dictated by energy minimization. This study aims to replace these expensive simulations with a trained machine learning (ML) model that yields accurate energy values for a given atomic structure. Further, we train an inverse model to determine the atomic configuration corresponding to the given SRO parameter. We propose a data-driven approach to map out the SRO directly to the atomic arrangements, combining ab initio calculations and a Neural Network (NN) model. We perform DFT-based MC simulations for Ni-V alloys in a wide range of solute compositions. Then, forward and inverse NN models are trained to map the SRO parameters into atomic arrangements or vice-versa. We predict the critical resolved shear stress (CRSS) for slip for all the studied configurations encompassing random and SRO structures and discuss the effect of SRO on the flow stress. The proposed ML methodology provides atomic arrangements from target order parameters with high accuracy, thereby eliminating the need for expensive simulations, and it advances the understanding of SRO at the atomistic scale.