Seismic stability of cracked rock slopes based on physics-informed neural networks

Cracks have been proven to significantly impact soil slope stability, whereas their mechanisms influencing rock slope stability have not been effectively investigated. Within the upper-bound theorem framework, the non-linear Hoek-Brown yield envelope is divided into multiple segments using a series of tangent lines to establish a three-dimensional (3D) multi-segment failure mechanism and a discontinuity surface is introduced into the first segment to account for the presence of a pre-existing crack at the slope crest. A novel physics-informed neural network (PINN) is then developed to calculate the seismic acceleration induced by seismic waves. The PINN-based framework has the advantage of describing the spatiotemporal characteristics of seismic waves while adhering to the geometrical constraints of slopes. To acquire the space-dependent seismic force-generated work rates, a slice integration strategy is introduced, followed by the derivation of the stability number with seismic action. Determining the critical stability number of cracked rock slopes involves identifying a critical failure mechanism characterized by cracks that most adversely affect slope stability in terms of depth and location. This process is a multivariable optimization scheme supported by the innovative Marine Predators Algorithm (MPA). Results indicate that seismic excitation and the existence of cracks considerably narrow the stress distribution range on the Hoek-Brown strength envelope, leading to a reduction in rock slope stability. The PINN model can provide a more realistic distribution characteristic of seismic loadings within slopes, taking into account the slope geometry, rock properties, and seismic wave characteristics. Considering the powerful transformation capabilities of Fourier analysis, the proposed PINN-based seismic analysis framework demonstrates the potential for incorporating more realistic seismic data into an analytical framework of slope stability analysis.