HOSSNet: An efficient physics-guided neural network for simulating micro-crack propagation

Abstract

The Hybrid Optimization Software Suite (HOSS), which combines the finite-discrete element method (FDEM), is an advanced approach for simulating high-fidelity fracture and fragmentation processes. However, the application of pure HOSS simulation is computationally expensive. Meanwhile, machine learning methods, which have shown tremendous success in various scientific problems, are increasingly being considered as promising alternatives to physics-based models in scientific domains. Thus, the objective of this work is to develop a new data-driven methodology to accurately reconstruct micro-crack fractures in both spatial and temporal fields. We leverage physical constraints to regularize the fracture propagation in long-term reconstructions. Additionally, we introduce perceptual loss and several machine learning optimization approaches to further enhance the reconstruction performance of fracture data. We demonstrate the effectiveness of the proposed method through both extrapolation and interpolation experiments. The results confirm that the proposed method can reconstruct high-fidelity fracture data across space and time in terms of pixel-wise reconstruction error and structural similarity. Visual comparisons also yield promising results in long-term reconstruction.