A machine learning-based atomistic-continuum multiscale technique for modeling the mechanical behavior of Ni3Al

In this paper, a machine learning-based atomistic-continuum multiscale method is developed to model the mechanical behavior of Ni-based superalloys. The kinematic and energetic consistency principles are employed to link the atomistic and continuum scales. The Ni-based superalloys with different void defects are proposed, and the effect of various parameters is investigated to distinguish the proper atomistic RVE in the multiscale analysis. The dataset of stress-strain samples is generated using the molecular dynamics analysis under various loading conditions that is divided into several groups according to strain values by the K-means algorithm to enhance the regression accuracy, in which a feedforward neural network is trained for each group. In order to optimize the unknown parameters in neural networks, the Bayesian regularization approach is employed through the optimization process. The material properties of the coarse-scale domain required for the multiscale analysis, i.e., the stress tensor and tangent modulus, are then derived from the trained neural networks for each Gauss integration point. Finally, several numerical examples are solved to illustrate the capability of the proposed machine learning-based multiscale technique. The results of the multiscale method are compared with those of molecular dynamics analysis that shows the capability of the proposed computational algorithm in capturing the phenomena, such as the rupture and stress concentration in Ni-based superalloys