Collaborative Research: CAIG: Understanding Ocean Physics via Multiscale AI Emulators

This project will build the first multiscale AI ocean emulator, spanning from submesoscales to large-scale ocean flows, for long-term ocean variability. The emulator will be used to investigate the complex nonlinear dynamics of the ocean circulation and to bridge the gap between understanding and simulating ocean variability across many time and space scales. This project will also advance the development of physics-informed, autoregressive neural networks capable of learning from heterogeneous, multi-resolution datasets. It addresses core challenges in scientific machine learning, including stability and interpretability in complex systems, and introduces methodologies that can extend to other multiscale problems. Understanding the interactions between oceanic processes across scales is essential for advancing our knowledge of ocean circulation, heat and momentum transport, and their role in shaping long-term variability. The primary objective is to investigate the spatio-temporal interactions between submesoscales, mesoscales, and large-scale flows on regional and global scales. This will be achieved by building a three-dimensional multiscale AI emulator, using deep neural networks, trained on a suite of numerical and observational datasets at different spatio-temporal resolutions. The questions to be addressed include: How to construct a physically-based 3D AI ocean emulator from heterogeneous datasets? How to evaluate a multiscale emulator from sparse and imperfect datasets? What fraction of ocean submesoscale and mesoscale physics drives momentum, energy, and heat transport at large scales? What is the role of multiscale processes on local ocean variability, such as marine heat waves and sea level extremes? This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.