機械学習を用いた短繊維強化複合材料のマルチスケール解析

抄録

A new data-driven multi-scale material modeling method called Deep Material Network (DMN), based on the Representative Volume Element (RVE) method and machine learning techniques, has been developed to predict physical properties with high accuracy and speed. DMN represents the response of RVE by forming a network using a mechanical building block that avoids the loss of intrinsic physics that occurs in common neural networks. The network can be constructed based on the analysis of linear elastic RVEs and can predict the nonlinear characteristics of various RVEs. In this paper, the basic concept of DMN is explained and its application to short-fiber reinforced composites is verified. The results confirm that DMN can reproduce the nonlinear response of the RVE with the same accuracy and orders of magnitude faster than the results of direct numerical simulations.