Predicting void nucleation in microstructure with convolutional neural networks

Abstract

Void nucleation in ductile materials subjected to high strain-rate loading remains a critical yet elusive phenomenon to understand. Traditional methods to understand void nucleation typically rely on experiments and molecular dynamics and do not capture the underlying factors leading to void nucleation. In this study, a convolutional neural network, specifically a U-Net enhanced with attention gates is developed, to predict void nucleation probability in pristine tantalum microstructures. The approach leverages a multi-channel input, incorporating four channels of grain orientations and an additional channel of grain boundary energy calculated via the lattice matching method. Void nucleation probability fields are determined from post-mortem micrographs and serve as ground truth, distinguishing void from no-void regions at the pixel level. Pixel-level class imbalance, commen in such images, is addressed by using Focal loss to guide the network's training to predict void nucleation sites more effectively. The model not only predicts void nucleation sites consistent with ground-truth but also reveals additional potential void nucleation sites, capturing the stochastic nature of void nucleation. This study shows that CNN-based models can predict void nucleation sites while considering combined interplay of factors such as grain boundary energy and grain orientation. In this way, machine learning can serve as a means to understand the underlying factors leading to void nucleation thereby contributing to a fundamental understanding of failure due to spallation in ductile materials.