Machine learning potentials for refractory high-entropy alloys applied to atomistic modeling of dislocation slip behavior

In high-entropy alloys (HEAs)—metallic materials composed of several principal elements—mechanisms underlying their exceptional mechanical properties remain largely unknown. Atomistic modeling of HEAs is complicated by the chemical inhomogeneity of atomic environments, which requires the development of new interaction potentials for simulating the mechanical response of large systems within classical approximation. Machine learning interatomic potentials have proven to be an indispensable approach to answering this problem. The present study focuses on building machine learning potentials based on artificial neural networks for two body-centered cubic alloys: VNbMoTaW and TiZrNbHfTa. For the newly developed potentials, we demonstrate accurate representation of basic material properties such as the bulk modulus, elastic constants, and stacking fault energies. Both VNbMoTaW and TiZrNbHfTa demonstrate high strength over a wide range of temperatures, however the origin of the difference in their ductility requires further investigation. Our simulations of screw dislocation slip behavior in these two alloys suggest that a strong pinning effect contributes to the high brittleness of VNbMoTaW. In contrast, in TiZrNbHfTa, pinning is suppressed, and the (112) slip plane is easily activated, which agrees with previous observations of frequent cross-slip at high temperatures in this alloy. • Artificial neural networks are used to build two robust interatomic potentials for VNbMoTaW and TiZrNbHfTa high-entropy alloys. • The high brittleness in VNbMoTaW is attributed to the pinning effect of screw dislocation motion. • The lowered Peierls barrier of screw dislocation motion is evident from the dynamics of dislocations in TiZrNbHfTa. • The easy activated (112) slip plane in TiZrNbHfTa provides a plausible explanation of slip persistent on (112) plane observed in TiZrNbHfTa.