Physics-informed neural networks applied to catastrophic creeping landslides

In this study, a new paradigm compared to traditional numerical approaches to solve the partial differential equation (PDE) that governs the thermo-poro-mechanical behavior of the shear band of deep-seated landslides is presented. In particular, this paper shows projections of the temperature inside the shear band as a proxy to estimate the catastrophic failure of deep-seated landslides. A deep neural network is trained to find the temperature, by using a loss function defined by the underlying PDE and field data of three landslides. To validate the network, we have applied this network to the following cases: Vaiont, Shuping, and Mud Creek landslides. The results show that, by creating and training the network with synthetic data, the behavior of the landslide can be reproduced and allows to forecast the basal temperature of the three case studies. Hence, providing a real-time estimation of the stability of the landslide, compared to other solutions whose stability study has to be calculated individually for each case scenario. Moreover, this study offers a novel procedure to design a neural network architecture, considering stability, accuracy, and over-fitting. This approach could be useful also to other applications beyond landslides.