Our team has developed an interpretable AI approach for transfer learning to improve S2S prediction of temperature along the U.S. West Coast. The approach shows skill at S2S timescales (a timescale which is notoriously difficult to predict) where extensive observations are lacking for training on observations alone. We have also devised an interpretable AI approach to make skillful climate forecasts on subseasonal-to-decadal timescales via analog forecasting. We are exploring a different interpretable model approach for S2S prediction based on the method of “analogs” for NOAA.

We successfully deployed robust, interpretable and XAI methods to identify skillful “forecasts of opportunity” (i.e. states of the climate system that are more predictable than others). We demonstrated that by training on climate model output (CESM2) and then fine-tuning with as well as reanalysis/observations (ERA5) we can make skillful S2S forecasts of North American temperatures in the reanalysis. This approach led us to identify representative patterns (i.e. prototypical patterns) within the real-world that lead to strong tropical Madden-Julian oscillation teleconnections to the midlatitudes.

Our prototype-based interpretable neural network when location matters (i.e. that is applicable to geophysical applications) for use for S2S prediction can be found to the left (see figure).

Related Project: Explainable AI, Robust AI

We a new method to predict average regional water levels along the Northwest Gulf of Mexico for S2S to multi-year predictions. While deep Gulf of Mexico heat profiles were expected to be skillful predictors, the method is presently only based on past processed regional average water levels, the former likely still driving the system at the S2S scale. This work was done in collaboration with the University of Valencia, Spain.