Stable Autoregressive Deep Learning Climate Models Aligned with Unbiased Global Mean Evolution
Yue Sun
Author Information:
Yue Sun, National Computational Infrastructure, Canberra
Edison Guo, National Computational Infrastructure, Canberra
Mark Collier, CSIRO Environment, Aspendale, Melbourne
Terry O’Kane, CSIRO Environment, Hobart
Abstract Body:
Deep learning models have achieved impressive accuracy in forecasts and stability in long-term autoregressive predictions, but their training objectives—typically focused on minimising regression error metrics between predictions and ground truth—are often disconnected from the climate indices that matter most for scientific understanding and policy. Additionally, once a deep learning model is trained on data at a specific spatial resolution, researchers are typically constrained to use input data at that same resolution for inference, limiting flexibility and broader applicability.
To address these limitations, we replicated the LUCIE model—whose training is guided not only by RMSE but also by climate bias—to explore how incorporating different objectives can lead to better climate models. We also removed the positional embedding in its backbone model, SFNO, to allow inference on inputs with different spatial resolutions.
Our experiments show that the updated model maintains the original LUCIE’s accuracy and long-term stability over 40 years of autoregressive prediction at six-hour intervals, even at higher resolution and with more variables. These modifications improve the model’s flexibility and scientific utility without sacrificing performance.
climate researchers, machine learning researchers, practitioners seeking adaptable tools for climate analysis and prediction.
deep learning, climate modelling, LUCIE, SFNO, Gadi
Please use this thread for discussion about this talk.