Multi-year to decadal predictions using a model-analog framework and a learned mask of weights trained on observations

Multi-year to decadal climate predictions are key tools in understanding the range of potential regional and global climate futures. A framework that uses a learned weighted mask to select model-analogs for climate predictions on these timescales is presented. Analog forecasting is the method of finding the most similar climate states, called analogs, to the state being predicted. The mean and spread of the known future of these top analogs act as the prediction and prediction uncertainty. The neural network contributes to this method by learning the regions that are most important in determining if two climate states will evolve similarly, ensuring that the analog selection encodes information specific to the prediction task.

Results for temperature and precipitation will be shown for several regions, including the Western United States, for lead times ranging from 1 to 15 years. In addition to a discrete prediction, the ensemble of top analogs allows us to make statistical predictions, such as the probability of reaching temperature or precipitation thresholds. We find that training the weighted mask on a combination of Earth system models (selected from a set of 30 CMIP6 ESMs) and observations (ERA5, BEST) leads to better predictions. We compare our predictions to initialized ESMs (e.g., CESM DPLE) and traditional analog methods, which select the top analogs based on the target region or the entire globe, using common accuracy methods as well as the spatially-aware Earth mover's distance metric.