Improving accuracy of artificial neural networks used to reproduce potential energy surface in BCC multi-component alloys

Lobzenko, I.   ; Tsuru, Tomohito   

In the atomistic modeling of mechanical properties of materials, machine-learning potentials (MLPs) based on ANNs, trained on large datasets obtained from first-principles calculations, found their place in reproducing the potential energy surface of new complicated materials. For materials such as multi-component alloys (MCA), reliable atomistic modeling is only possible by utilizing the fitting power of ANNs. However to achieve high accuracy and robustness of interatomic potentials built with ANNs, not only a large dataset is needed, but also the architecture of the network and the training process should be curated for each material. In the present study we model the mechanical properties by using newly built MLPs of a set of base-centered cubic (BCC) materials, such as MCAs based on MoNbTaVW and ZrNbTaTiHf alloys. As one of the of results, the modeling of dislocation movement with MLPs in MoNbTa is discussed. The unusual (112) slip plane, which is different from the usual 110 slip plane in single BCC metals, was found.