Microstructure homogenization: human vs machine

Abstract Two approaches are presented to improve the capabilities of machine learning models in multiscale modeling for microstructure homogenization (graphical abstract in Fig. 1). The first approach features a Bayesian data mining scheme with a human in the loop, halving the prediction error compared to [1] using four novel and efficient to evaluate feature descriptors. The second purely machine learning-driven approach utilizes convolutional neural networks, where we introduce a novel module (the deep inception module) designed to capture characteristics of different length scales within the image. The new module features a new normalization block, which aids in calibrating the differently obtained feature characteristics. Further improvements, universally applicable to artificial neural networks, are found with a novel hyperparameter insensitive learning rate schedule, which adapts to the training progress of the model. A further improvement is given by a pre-trained feature bypass which utilizes global low-level features to serve as baseline prediction such that the model is able to dedicate its attention to high-level features. The proposed schemes have been applied to different literature models, yielding significant improvements in any of the investigated convolutional neural networks. The improvements found by the two overarching contributions, i.e., derived through feature development with a human in the loop, and via convolutional neural networks, are critically assessed in a thermal and mechanical setting. It is further expanded to variable material parameters while allowing for variable microstructural elements, yielding drastically reduced prediction errors across the board.