Accelerated offline setup of homogenized microscopic model for multi‐scale analyses using neural network with knowledge transfer

The FE 2 $$ {}^2 computational homogenization method is a predictive multi-scale without the need for constitutive assumptions and/or potential function postulates at macro engineering scale. Instead, effective micro-structural responses are extracted directly from representative volume element (RVE) underlying each point. However, still computationally too expensive most practical uses, since micro-macro coupling done loading step/iteration entire domain. To this end, machine learning has been utilized in literature offline training of surrogate model to predict RVE homogenized response general conditions. In contribution, neural network (NN) incorporated into finite framework non-intrusive manner. This termed as FE-NN framework, analogy method. general, online simulations very efficient, with predictions matching closely those obtained reference direct numerical (DNS). A bottleneck however, high cost associated data generation NN setup. paper, focusing on non-linear elastic deformation heterogeneous materials, sequential strategy knowledge transfer proposed, enable an efficient microscopic For given target RVE, we first consider simplified source where can be generated rapidly, pre-training model. pre-trained parameters next downloaded initialize model, followed by fine-tuning process, using only small dataset high-fidelity RVE. efficiency proposed over conventional training, well as, its excellent capability analyses, demonstrated multi-phase composite material. FE-NN-KT approach implemented easily complicated pre-processing procedures, involves standard extraction RVEs, together architecture.