Abstract
The design of multi-principal element alloys (MPEAs) is challenged by data scarcity and the complexity of optimizing multiple mechanical properties simultaneously. To address these challenges, we investigate an inverse material design approach that integrates experimental data with machine learning using generative adversarial networks. A conditional Wasserstein generative adversarial network (cWGAN) is developed and compared to a baseline conditional GAN (cGAN). Both models are designed to generate alloy compositions conditioned on hardness and elastic modulus. The proposed cWGAN framework demonstrates superior performance, achieving more stable training dynamics, closer alignment with experimental distributions, and more accurate reproduction of elemental correlations compared with the cGAN model. Furthermore, compositions generated by cWGAN improve the performance of downstream predictive models, confirming their quality and utility for high-throughput evaluation and discovery of novel MEPAs with targeted properties. Collectively, these results demonstrate the advantages of the cWGAN architecture under data-limited conditions and establish it as a practically useful framework for accelerating the design of MPEAs with enhanced mechanical performance.