Hugo Jehoshua Bethancourt, Hokkaido University, Japan

WRF and MPAS are useful models for weather forecasting. In particular, the Voroni mesh and discretization structures for MPAS allow for highly parallel systems to elucidate future climate events. Conveniently, many parallelizable systems may take advantage from convolutional neural networks (CNNs). First, this presentation talks about neural networks and different categories of Artificial Intelligence. Then, we talk about how physics and differential equations are useful for solving real-world applications, starting with the usual heat equation as an example. We combine the themes of differential equations and their solvability from neural networks in detail, including the advantages of each neural network model to make more accurate MPAS. There will be a detailed discussion about the use of Spherical CNNs over unstructured grids and its future applicability to MPAS. Next, there will be a discussion on limited training data for neural networks and how physics-informed neural networks (PINNs) are able to further enhance accuracy of MPAS models by construction of an inherent loss function and its optimization. Lastly, we shall look at solving Burger’s equation with a PINN. Given initial conditions and boundary values we will notice the usefulness of PINNs for elucidating limited datasets which use MPAS. A short discussion will be given on future avenues of research for PINN-enhanced MPAS.