Data-driven discovery of a formation prediction rule on high-entropy ceramics

The interest in high entropy ceramics (HECs) has increased steadily due to their superior properties. However, the prediction of their formation still poses challenges for the discovery of new systems. Here, we discover a rational rule for designing single-phase high entropy transition metal diborides (HEBs) using data-driven approach. The machine learning (ML) model is trained on data collected via high-throughput experiments (HTEs). K nearest neighbors (KNN) model shows an experimental validation accuracy of 93.75%. By imple-menting interpretable ML method, we demonstrate that a mismatch of the bonds between boron and transition metals (delta B_TM) dominates the formation of HEBs. We propose an empirical rule that HEBs favor forming a single phase when delta B_TM < 3.66; otherwise, multiphase. The rule has a high accuracy of 93.33% for new HEBs pre-dictions. In addition, we contribute 165 high quality HEBs data in total, which can promote the development of materials informatics in HEBs. Moreover, this data-driven strategy can be expanded to accelerate the search for new HECs, paving a pathway to design novel HECs with superior properties rapidly.