A data-driven approach to approximate the correlation functions in cluster variation method

Abstract

The cluster variation method is one of the thermodynamic models used to calculate phase diagrams considering short range order (SRO). This method predicts the SRO values through internal variables referred to as correlation functions (CFs), accurately up to the cluster chosen in modeling the system. Determination of these CFs at each thermodynamic state of the system requires solving a set of nonlinear equations using numerical methods. In this communication, a neural network model is proposed to predict the values of the CFs. This network is trained for the bcc phase under tetrahedron approximation for both ordering and phase separating systems. The results show that the network can predict the values of the CFs accurately and thereby Helmholtz energy and the phase diagram with significantly less computational burden than that of conventional methods used. © 2021 IOP Publishing Ltd.