Reproduce “Towards physics-informed neural networks for landslide prediction” on OpenResearch

About this paper

Ashok Dahal, Luigi Lombardo

A Physics-Informed Neural Network (PINN) predicts regional coseismic landslide susceptibility by inferring latent geotechnical parameters from proxy data and integrating Newmark's permanent deformation method. This approach demonstrates superior spatial transferability and robustness compared to purely data-driven models, while also providing physically interpretable geotechnical property maps.

Key insights

• Integrating physical laws into neural networks significantly enhances model generalizability and robustness, particularly when predicting in spatially distinct, unseen regions.
• PINNs can infer physically meaningful, spatially distributed geotechnical parameters from readily available environmental proxy data, overcoming a major data scarcity challenge in regional modeling.
• The ability to provide physically interpretable intermediate outputs fosters trust in the model's predictions and advances explainable AI in geosciences.

Summary by alphaXiv.