Taimoor Sohail: Towards a Machine Learning Ocean Emulation Framework for Decadal Climate Prediction
Long-term (decadal to centennial) changes to the global ocean, including trends in ocean warming, sea level rise and carbon absorption, are vital to understanding the impacts of climate change. Long-term ocean projections from climate and ocean models remain highly uncertain. This is in part due to a lack of historical ocean observations to validate existing models, and due to the computational cost of resolving key processes over the timescales needed to capture the long arc of climate change. In light of this, machine learning methods have been proposed as a means to learn and predict future ocean changes at a fraction of the cost. This is achieved predominantly by training deep learning methods on ocean properties, learning the statistical relationships between these properties, and emulating property change over time. Notable examples include Samudra, an ocean machine learning emulator that replicates the mean ocean state 150 times faster than traditional, physics-based models.
Given the advent of machine learning emulation of the ocean state, there is a need to collaboratively develop capacity towards an ocean emulation framework for decadal predictions here in Australia. Amongst other efforts, this project proposes one pathway towards this. Together with researchers at University of Tasmania, the Bureau of Meteorology and ACCESS-NRI, we explore the use of Autoencoders from the PyEarthTools package, alongside probabilistic neural networks, physics-informed neural networks and reinforcement learning, to develop a robust ocean emulation framework. This aims of this project are manifold: (a) develop an ocean emulator which reproduces realistic future ocean projections given a set of input parameters at a fraction of the cost of current modelling efforts, and (b) provide another platform for capacity building in earth system machine learning emulation. This project is envisioned as much as a training opportunity as a research project, and hence we welcome additional collaborators, development expertise and machine learning pros to build an open, flexible platform that can be used by climate researchers into the future.