Generative model-based inverse design of Fe-based metallic glasses with high saturation magnetic flux density

Fe-based metallic glasses (MGs) have attracted extensive interest owing to the excellent soft magnetic properties. However, discovering Fe-based MGs with high saturation magnetic flux density (Bs) remains great challenge, because of the vast composition space and the limitations of trial-and-error methods. Supervised machine learning methods accelerate the composition search, but require manual input of alloy compositions, which becomes impractical as the number of possible compositions increases exponentially with increase of alloy complexity. To address these challenges, we developed an interpretable generative model-based inverse design approach by using the Wasserstein Autoencoder combined with a property predictor (joint-WAE model), which can tackle the high-dimensional optimization problem and efficiently generate unreported Fe-based MGs with target Bs values without human intervention. Our joint-WAE model exhibits high fidelity and accuracy in composition reconstruction and property prediction. Furthermore, by integrating the optimization algorithms, our joint-WAE model can further accelerate the search in complex composition space. Unreported Fe-based MGs with Bs > 1.5 T have been successfully predicted by our model and validated by experiments. The interpretability of our model also provides physical insights into how constituent elements and compositions influence magnetic properties. These findings demonstrate that our generative model is promising to achieve accurate and efficient discovery and optimization of novel MGs with desired properties.