Data-driven discovery of ultrahigh specific hardness alloys

Abstract

Data-driven discovery of advanced alloys using experimental data is challenging due to the difficulty of obtaining large and reliable experimental datasets, the vastness of the compositional design space, and the risk of overfitting machine learning (ML) models. This study employed an iterative process of ML prediction paired with combinatorial experimental verification to discover new ternary alloys with ultrahigh-specific hardness. Combinatorial experimental datasets and an ensemble of six ML algorithms (elastic net, support vector machine, Gaussian process regressor, random forest, artificial neural network, and convolutional neural network) were used to explore a compositional space blended by 28 metallic elements. Through this approach, we discovered 22 new compositions in the Al-Ti-Cr system within a high specific hardness regime (>3254 kN m/kg) and 86 new compositions in Al- and Mg-based alloys within the high specific hardness/density regime (>0.61 kN m4/kg2) that were not accessed by previously reported alloys. Notably, the Al-Ti-Cr system reached a specific hardness of 3254 kN m/kg, surpassing the highest reported value by 12%. Explainable artificial intelligence was applied to elucidate the mechanisms behind these superior mechanical properties.

Keywords

• Data-driven materials sciences
• Complex concentrated alloys
• Combinatorial experiments
• Nanoindentation
• Specific hardness