Data-driven void growth prediction of aluminum under monotonic tension using deep learning

Abstract

Void growth plays a significant role in ductile fracture prediction of aluminum. This study proposes 2 deep learning models to address this issue. For voids that retain their ellipsoidal characteristics during growth, the ellipsoidal void Semiaxes Long Short-Term Memory (SLSTM) method is proposed, using the 3 principal features of the ellipsoid to represent the voids. For voids that undergo arbitrary shape changes during growth, an innovative deep learning method called Voronoi tessellation-assisted LSTM (VLSTM) is proposed. This method uses the Voronoi algorithm to standardize data features and employs Principal Component Analysis (PCA) to perform data compression before neural network training. This new method combines the Voronoi algorithm, LSTM neural networks, and PCA algorithms, and is termed as VLSTM-PCA. In this study the deep learning-based SLSTM surrogate models and VLSTM-PCA surrogate models run approximately 514 and 537 times faster than ABAQUS finite element simulations, significantly enhancing efficiency while maintaining high prediction accuracy. Finally, growth patterns of ellipsoidal voids under different stress triaxialities are analyzed.