Machine learning potential-driven prediction of high-entropy ceramics with ultra-high melting points

Developing high-entropy ceramics (HECs) with ultra-high melting points (Tm) is crucial for their applications in ultra-high-temperature environments. However, related research has seldom been reported. Here, taking high-entropy diborides (HEBs) as an example, we develop a data-driven method to efficiently explore HEBs with ultra-high Tm via transferable machine-learning-potential-based molecular dynamics (MD). Specifically, a moment tensor potential (MTP) for HEBs with nine transition metal elements of group IVB, VB, and VIB is first constructed based on unary and binary diborides. Further studies on the performance of our constructed MTP have confirmed its remarkable accuracy, transferability, and reliability across both equimolar and non-equimolar HEB systems. Tm of HEBs are then accurately simulated through MD simulations based on the constructed MTP, and 24 features are simultaneously collected to enable reliable machine learning training. Five descriptors with the gradient boosting regression model are derived as the optimal combination for accurate Tm predictions in HEBs with genetic algorithms. Based on our established model, Tm of 32563 HEBs are eventually determined, achieving the maximum Tm of 3688 K in (Ti0.1Zr0.1Hf0.6Ta0.2)B2. The work presents a feasible approach to develop HECs with ultra-high Tm.