AI/ML-Assisted Inverse Alloy Design of Multi-Principal Element Alloys for High-Temperature Energy Systems: Comparative Tree Ensembles and Deep Neural-Network Surrogates

Abstract

High-temperature energy systems require structural materials that retain strength while preserving sufficient deformation capacity under thermal exposure. The compositional design space of multi-principal element alloys (MPEAs) is, however, too large for exhaustive trial-and-error exploration. This work develops a reproducible artificial-intelligence/machine-learning (AI/ML) workflow for forward property prediction and inverse alloy design of MPEAs using the expanded experimental database reported by Borg et al. The original database contains 1,545 literature-derived records; the present analysis isolates tensile data and represents alloy chemistry through normalised elemental atomic fractions together with test temperature, grain size, processing route, microstructure, phase class, and composition-derived descriptors. Four supervised regressors were compared: Random Forest (RF), Extremely Randomized Trees (Extra Trees), Extreme Gradient Boosting (XGBoost), and a feed-forward deep neural network (DNN). Leakage was reduced by splitting data by normalised alloy composition rather than by random row. Independent surrogates were trained for yield strength and elongation. Under the verified grouped holdout configuration, the best yield-strength result was obtained by RF (R2 = 0.164, MAE = 173 MPa, RMSE = 251 MPa), whereas XGBoost gave the highest elongation R2 (0.292, MAE = 15.21%, RMSE = 20.63%) and the DNN gave the lowest elongation MAE (14.58%). The modest R2 values demonstrate that heterogeneous literature data contain substantial variance not captured by nominal chemistry and coarse processing descriptors. Nevertheless, surrogate-assisted screening at T ≥ 500 ◦C using target constraints of predicted yield strength ≥ 500 MPa and predicted elongation ≥ 15% identified a literature-supported Al-Co-Cr-Fe- Ni-Ti composition at 700 ◦C as the highest-ranked condition, with predicted yield strength of 521.7 MPa and predicted elongation of 15.72%. The study therefore positions AI/ML not as a replacement for experiment, but as a transparent decision-support layer for narrowing high-dimensional alloy spaces and prioritising candidates for experimental validation in high-temperature energy applications.