Abstract: Refractory medium entropy alloys (RMEAs) have attracted significant attention in recent years due to their exceptional mechanical properties and high-temperature stability, making them suitable for a number of advanced applications. While computational modelling such as density functional theory (DFT) and molecular dynamics (MD) are powerful for investigating RMEAs, these traditional methods are often constrained by high computational cost and limited accuracy. In this work, a deep neural network potential (DNNP) was developed to address the complex compositional nature of FeCr₂V-based RMEAs with varying levels of tungsten doping. The DNNP demonstrated high accuracy, comparable to that of DFT calculations. Utilizing the DNNP, high-accuracy MD simulations were conducted to examine large-scale effects, including short-range ordering (SRO), twining, and dislocation behaviour, on the mechanical properties of the RMEAs. The results indicate that in the SRO structure, the covalency of V-V, Cr-Cr, and V-W ordered atomic pairs enhances local bonding strength and increases the elastic modulus of the RMEA. As the simulation temperature increases, dislocation mobility improves while dislocation density decreases, thereby enhancing the ductility of the material. Above 823 K, the SRO structure demonstrates superior mechanical performance, which is attributed to the increased length of 1/2 < 111>, dislocations facilitated by the formation of Cr-Fe and Cr-Cr ordered twins. This work underscores the potential of DNNP and MD simulations in predicting and analyzing the mechanical properties of RMEAs, advancing their development for various applications.

Key words: High entropy alloys (HEA), Medium entropy alloy (MEA), Density functional theory (DFT), Deep neural network potential (DNNP), Molecular dynamic (MD), Nuclear materials