Foretelling microstructural interface with multi-generational convolutional-LSTM framework

期刊论文
发表日期：

2025年2月28日

查看全文

Upadesh Subedi,
Nele Moelans,
Tomasz Tánski,
Anil Kunwar

摘要

Predicting multi-generational microstructural evolution in the manifest space using convolutional-LSTM neural network models is a promising approach towards computational reducibility. Two predictive models were trained, analyzed, and compared using data generated from varying the driving force in phase-field simulations of phase decomposition in a binary alloy system. The four consecutive predicted generations exhibited microstructural similarity indices ( $\mu _{ SIM}$ $μ_{SIM}$ ) of 0.994, 0.987, 0.980, and 0.967, respectively, compared to the ground truth computational results. By utilizing blobs, $\mu _{ SIM}$ $μ_{SIM}$ , and shape index features, we validated that the predicted images retain microstructural information down to the interfacial level. The time-aware data-driven model was applied to rapidly forecast structural changes in Pd–Rh and Cu–Mn engineering alloys.