Exploring physics-informed neural networks for glacier flow.

Deep learning-based surrogate models have emerged as computationally inexpensive tools for simulating glacier dynamic systems defined by complex, nonlinear partial differential equations. Despite the potential, the application of physics-informed neural networks  (PINNS) in glacier modelling is sparse. Thus, this study explores the potential of  NVIDIA Modulus-Sym, a PyTorch-based framework, in simulating glacier velocities. The framework NVIDIA Modulus-Sym provides ground for building, training and fine-tuning physics-based surrogate models targeting computational fluid problems. This study presents the pipeline to generate glacier velocities using the physics-constraint approach that incorporates the physics regularisation term within the loss function to enhance generalisation performance. The study further emphasises the challenges and limitations of tools in glaciological research. 

Keywords: Deep learning, Surrogate Models, Glacier Dynamics,  Glaciology