A machine learning framework augmented by generative models for accelerated design of refractory high-entropy alloys with superior strength-ductility synergy

Abstract

Achieving a balance between strength and ductility at room temperature remains a grand challenge in developing body-centered cubic (BCC) refractory high-entropy alloys (RHEAs). Facing the vast compositional design space, we propose an alloy design framework, GAMP, combining a Generative Adversarial Network (GAN) with machine learning (ML) prediction models to accelerate the screening of potential alloy compositions. Unlike common inverse design or image-based approaches, the GAN in this work functions as a generator within a forward search loop, specifically capturing the latent distribution of existing alloys to propose high-potential new candidates. Among various models, eXtreme Gradient Boosting (XGBoost) was identified as the most accurate for predicting ultimate tensile strength (UTS) and elongation (El) to failure. A designed alloy, Al_9.6Ti_38.0Nb_20.5Hf_16.5V_15.4, was experimentally validated to exhibit a UTS of 1058.1 MPa and an El of 18.5% at room temperature. Interpretability analysis further revealed that shear modulus local mismatch (G_LM) and atomic size difference (δ) are the dominant descriptors for UTS, while enthalpy of mixing (ΔH_mix) is the key indicator for El. Microstructural analysis revealed that the superior strength-ductility synergy resulted from the activation of multiple deformation mechanisms, in which the formation of kink bands is a key factor enabling the excellent room-temperature ductility. This study presents an efficient framework for designing advanced alloys with complex property requirements.