Wei Sun

Random fracture propagation is frequently employed in geotechnical and structural engineering systems. However, it is inherently nonlinear, strongly discontinuous, and highly stochastic, posing significant challenges to efficient prediction and rapid engineering assessment. To overcome the prohibitive computational cost and limited scalability of conventional high-fidelity numerical simulations, this study proposes a physics-constrained conditional generative adversarial network (PC-GAN) for the full-process simulation of random fracture propagation based on datasets generated by peridynamics (PD). As a prerequisite, convolutional long short-term memory (ConvLSTM) networks are first examined and shown to effectively capture the temporal dependencies of fracture propagation under given historical sequences, thereby serving as the generative backbone. However, since step-wise temporal prediction is insufficient for full-process fracture simulation, adversarial learning is introduced to reformulate the task as a conditional image-to-image generation problem, complemented by physics-consistency constraints to ensure physical admissibility. The numerical results demonstrate that the proposed PC-GAN can simulate complete fracture propagation process directly from the initial fracture state, material parameters and loading conditions. The propoesd method is capable of achieving high precision, while reducing 50% of the computational time in representative numerical examples, compared with conventional step-by-step PD simulations.

Earthquake-induced liquefaction poses a serious threat to the safety of embankment dams. Numerous failures of earth embankments have been reported during past earthquakes, particularly when the underlying saturated granular soils experiences liquefaction, leading to cracking, settlement, lateral spreading, and slumping. However, most previous numerical studies have employed conventional continuum-based methods, such as the finite element method (FEM), which struggle to capture crack initiation and propagation in embankments affected by soil liquefaction. To address this limitation, this study proposes a porodynamic coupled peridynamics–FEM (PD–FEM) framework that integrates the CycLiq plasticity model for liquefiable foundation sand and a Mohr–Coulomb model with a tensile bond failure criterion for clayey embankment. A hybrid computational scheme is also proposed that adaptively transitions from FEM to PD–FEM, ensuring both accuracy and efficiency. Validation against centrifuge shaking table tests on clayey embankments over liquefiable Nevada sand demonstrates that the model successfully reproduces observed deformation, excess pore pressure evolution, and embankment cracking. A five-stage cracking mechanism is identified: initial free-field liquefaction, lateral spreading of the near-field soil, subsequent horizontal stretching of the embankment, tensile crack initiation at the crest/slopes, and progressive propagation and eventual stabilization of cracks. The coupled framework captures the release of tensile stress after cracking and avoids the unrealistic stress accumulation inherent in pure FEM simulations. Reinforcement evaluations show that foundation densification effectively suppresses liquefaction and cracking by enhancing the liquefaction resistance of foundation soils, whereas sufficiently large-volume gravel berms primarily provide confinement to reduce near field excess pore pressure accumulation and restrain deformation and crest cracking.