Christian Stassen: Machine learning for regional climate downscaling in the Australian Climate Service
High-resolution climate and hazard projections are an essential tool to understand how risks can change through time. Physics-based regional climate models, such as the Bureau’s Atmospheric Regional Projection for Australia (BARPA) and CSIRO’s Conformal Cubic Atmospheric Model (CCAM), provide the gold standard for those projections. However, with increased resolution, physics-based models can become computationally very expensive to run. This limits their application to a few select socio-economic pathways, projection ranges, or storyline approaches. Machine learning (ML), trained on the physical model data, can be used to supplement the physical models by providing a cheaper and quicker alternative to create large ensembles of future projections.
This presentation provides an update on the progress of developing and evaluating two ML models. We outline the progress in choosing the model architecture and present preliminary results. Parallel to the development of model architecture, preliminary work has been done in planning and developing an evaluation suite which will be used to benchmark the ML model performance against dynamical models.