Fast crack propagation prediction in SiCf/SiC via deep learning with phase-field-generated data

The high-accuracy numerical analysis for the crack propagation in solids with multiple initial pores by physics-based models can be greatly computational costing. A conditional generative adversarial network (cGAN)-based framework that is trained on the dataset generated by phase field method (PFM) is proposed for the crack propagation in silicon carbide fiber-reinforced silicon carbide matrix composites. The structural similarity calculations show that the cGAN achieves a high prediction accuracy by designing its object function and discriminator. By fixing the discriminator's receptive field to optimize the cGAN, the optimal discriminator's receptive field is 16 × 16, with their respective accuracy at 98%. Moreover, the trained cGAN is capable of simulating crack propagation with a wide range of geometric size, pore configurations, and pore shape, without any additional modification. Interestingly with a speed up by 28‒75 times in comparison to PFM, the computational time for the cGAN remains consistent at 10‒13 s, irrespective of the geometric size and node degrees of freedom. These characteristics make the cGAN framework a promising approach for quickly simulating the crack propagation.