A Generative Adversarial Networks (GAN) based efficient sampling method for inverse design of metallic glasses

Metallic glass has garnered significant attention due to its unique physical properties. However, the complex composition design space of alloy presents a challenge for traditional experimental methods in the development of metallic glass. In this paper, we propose a novel approach for rapidly generating hypothetical metallic glass compositions using a generative adversarial network (GAN) based sampling model. We evaluated GAN-generated samples in terms of validity, novelty, and uniqueness. Two different XGBoost models were employed to validate the validity of the generated samples, where the phase classifier evaluated that 85.6% of the GAN-generated samples were amorphous, and the critical casting diameter (Dmax) regressor evaluated that 89.2% of our generated samples had a Dmax greater than 1 mm. Moreover, we demonstrated the GAN-generated samples’ novelty and uniqueness by comparing their distribution with the real samples. Our GAN model is expected to improve the sampling efficiency of metallic glass and thus shorten its development cycle.