Decadal Predictions with Diffusion Models: Combining Machine Learning and Earth System Modelling

In the past years, decadal prediction systems have started to fill the gap between seasonal forecasts and long-term climate projections. Despite huge progress in predictive skill and decadal predictions outperforming climate projections in almost all forecast tasks, decadal predictions still possess large rooms for improvement. Machine learning based forecast systems have already outperformed traditional weather forecast systems in recent years. Similarly, machine learning has successfully transformed or assisted in data assimilation or climate data reconstruction tasks. Despite its success in the climate sciences, machine learning methods have not yet been successfully integrated in decadal prediction systems.

Combining machine learning and numerical modeling, we attempt to produce decadal climate predictions utilizing Diffusion Models, essentially probabilistic neural networks. We use such a neural network to predict global 2m-air temperatures by training it on the historical MPI-ESM-LR Grand Ensemble and finetuning it on the MPI-ESM-LR decadal predictions and on ERA5 reanalyses. The resulting predictions are qualitatively comparable to the standard MPI-ESM-LR decadal prediction system, surpassing their predictive skill for leadyears 1 and 2. With diffusion models still new to climate predictions, we expect this result to stand only at the beginning of further machine learning integration into climate predictions in general and decadal predictions in particular.