Subseasonal-to-Decadal Analog Forecasting Using Spatially-Weighted Masks Learned By a Neural Network

Analog forecasting—forecasting the evolution of an initial climate state based on the evolution of “analogous” climate states—is an intuitive approach to climate prediction. Recent work used an interpretable neural network to identify a spatially-weighted mask that quantifies how important each grid point is for determining whether two climate states will evolve similarly. Analog forecasts using these weighted masks have been shown to improve sea surface temperature predictions in the tropical Pacific and North Atlantic relative to traditional analog approaches. In addition, the weighted mask reveals robust precursor patterns in the climate system. This presentation will describe how the weighted-mask analog approach is being extended to various forecasting problems across subseasonal to decadal timescales.