Attention-Enhanced Variational Learning for Physically Informed Design of Ultra-hard Multicomponent Metallic Glasses

The discovery of ultra-hard multicomponent metallic glasses (MMGs) remains challenging due to the vast compositional complexity and the absence of physically grounded inverse design frameworks. Here, we present VIBANN—a novel AI-driven approach that synergistically combines Variational Information Bottleneck with Attention based Neural Networks, enabling interpretable and uncertainty-aware design of MMGs with record-level hardness. VIBANN compresses composition–mechanical load inputs into a disentangled latent space that captures chemically meaningful structure–property relationships, while attention layers identify critical elemental contributors to hardness. We integrate Monte Carlo dropout for calibrated uncertainty estimation and use Gaussian Mixture Models with latent-space gradient optimization to generate new alloy compositions autonomously. Experimental validation of five inverse-designed B–Co–Fe–Re–W–(Ni/Cr/V) glasses predicted by VIBANN yields bulk amorphous alloys with Vickers hardness exceeding 2200 HV, surpassing most known MMGs. SHAP analysis and latent traversals further expose smooth, interpretable hardness design directions, revealing compositional pathways with equivalent performance. In essence, VIBANN establishes a fully autonomous, interpretable, and experimentally validated inverse design loop, bridging deep learning with metallurgical insight and setting a new benchmark in data-driven alloy design.