Data-Guided Feature Identification for Predicting Specific Heat of Multicomponent Alloys

The ever-increasing interest in multicomponent alloys over the last decade is attributed to their intriguing structural properties. Nonetheless, these materials do hold potential where novel functional properties, such as high specific heat at extremely low temperatures, are required for applications, such as the cryogenic regenerator bed, that have been predominantly dependent on rare earth metals and lead. Given the enormous design space of alloy compositions, spanning over half a trillion possible compositions, and consisting of three to six principal elements, machine-learning and data-guided methods are the emergent techniques to address such materials discovery challenges. Here, we implement a gradient boost regressor to estimate the specific heat capacity of these alloys, but more importantly to understand the key material descriptors driving the predictions. The mean lattice constant and the valence electron concentration are the most impactful descriptors, since they directly influence vibrational frequencies and electron-phonon coupling, both of which are crucial material features for specific heat of multicomponent alloys. Additionally, our workflow also establishes the need for an extensive and homogeneous dataset to enable accurate ML-based predictions for material properties.

Automated summary

The increasing interest in multicomponent alloys is due to their interesting structural properties. Machine-learning and data-guided methods have emerged as techniques to address the challenges of discovering novel materials. By using a gradient boost regressor to estimate the specific heat capacity of alloys and understanding the key material descriptors driving the predictions, this study provides guidance for the functional properties of alloys.