A Neural Network Approach to Predict Gibbs Free Energy of Ternary Solid Solutions

TL;DR

This paper introduces a neural network-based deep learning method to predict Gibbs free energy in ternary solid solutions, leveraging a large binary dataset to enhance thermodynamic predictions across compositions and temperatures.

Contribution

It demonstrates the feasibility of using neural networks trained on CALPHAD data to predict thermodynamics of ternary systems, with insights into data sampling and potential improvements.

Findings

Neural networks can accurately predict Gibbs free energy for ternary solid solutions.

Binary datasets of 102,000 points effectively train the neural network models.

Data sampling sensitivity impacts prediction accuracy significantly.

Abstract

We present a data-centric deep learning (DL) approach using neural networks (NNs) to predict the thermodynamics of ternary solid solutions. We explore how NNs can be trained with a dataset of Gibbs free energies computed from a CALPHAD database to predict ternary systems as a function of composition and temperature. We have chosen the energetics of the FCC solid solution phase in 226 binaries consisting of 23 elements at 11 different temperatures to demonstrate the feasibility. The number of binary data points included in the present study is 102,000. We select six ternaries to augment the binary dataset to investigate their influence on the NN prediction accuracy. We examine the sensitivity of data sampling on the prediction accuracy of NNs over selected ternary systems. It is anticipated that the current DL workflow can be further elevated by integrating advanced descriptors beyond the elemental composition and more curated training datasets to improve prediction accuracy and applicability.