ESPEI for efficient thermodynamic database development, modification, and uncertainty quantification: application to Cu-Mg

TL;DR

ESPEI is a software tool that streamlines thermodynamic database development by combining linear fitting and Bayesian optimization, demonstrated on the Cu-Mg system to quantify and propagate uncertainties in Gibbs energy models.

Contribution

The paper introduces ESPEI, a novel software package that integrates linear fitting and Bayesian methods for efficient thermodynamic modeling and uncertainty quantification.

Findings

Successful application to Cu-Mg system down to 0 K

Quantification and propagation of model parameter uncertainties

Demonstration of Bayesian optimization in thermodynamic modeling

Abstract

The software package ESPEI has been developed for efficient evaluation of thermodynamic model parameters within the CALPHAD method. ESPEI uses a linear fitting strategy to parameterize Gibbs energy functions of single phases based on their thermochemical data and refine the model parameters using phase equilibrium data through Bayesian optimization within a Markov Chain Monte Carlo machine learning approach. In this paper, the methodologies employed in ESPEI are discussed in detail and demonstrated for the Cu-Mg system down to 0 K using unary descriptions based on segmented regression. The model parameter uncertainties are quantified and propagated to the Gibbs energy functions.