Deep learning-based surrogate model for generation of wood microstructures

Abstract

A machine learning-based surrogate model for efficient wood microstructure generation compatible with a physics-based model is developed. The model is based on the U-Net neural network, a variant of convolutional neural network which, due to its architecture, is suitable for image-to-image transformation, and focuses on the crucial step of the microstructure generation, which is distortion of the wood slices according to the prescribed distortion map. For training the U-Net, a dataset consisting of a variety of wood microstructures slices before and after the distortion is generated, along with the corresponding distortion maps, using the previously developed parametric model. Transfer learning is shown to improve the performance of the U-Net, especially if the dataset of wood slices is small. The best results with transfer learning are obtained if either the whole U-Net is fine-tuned or only the bottleneck block is frozen during the fine-tuning. The surrogate model is a promising tool for generating a large dataset of wood microstructures within the structural parameter ranges it was trained on in a relatively short time compared to the original method. That could in turn be useful for a parametric study and optimization of the physical properties of wood-based materials.