DMC-TPE: tree-structured Parzen estimator-based efficient data assimilation method for phase-field simulation of solid-state sintering

Abstract

In phase-field simulations, accurate material parameters are required to quantitatively predict microstructural evolutions. Non-sequential data assimilations enable the estimation of unknown material parameters by minimizing a cost function that represents the misfit between numerical simulation results and time-series observation data. In this study, a new non-sequential data assimilation method Minimizing the Cost function using tree-structured Parzen estimator (TPE), namely DMC-TPE, with higher estimation accuracy than that of a conventional method (DMC-Bayesian optimization (BO)) is developed. The estimation accuracy of DMC-TPE is compared with that of DMC-BO via numerical experiments, where these methods are applied to a phase-field simulation of solid-state sintering. The comparison results demonstrate that the estimation accuracy of DMC-TPE is higher than that of DMC-BO, specifically in cases where numerous parameters have to be estimated, because TPE can continuously minimize the cost function by increasing the number of iterative minimization calculations. Furthermore, DMC-TPE provides less scatter in the estimation results than that in the case of DMC-BO. The DMC-TPE developed herein leads to highly accurate PF simulations of microstructural evolution by simultaneously estimating the states and many unknown material parameters with high accuracies based on experimental data.

Keywords

• data assimilation
• bayesian optimization
• phase-field simulation
• sintering