Deseasonalized Attention for Scientific Discovery of Extreme and Compound Climate Events

The Earth system is characterized by complex, nonlinear interactions where the combination of multiple drivers can lead to extreme or compound events with significant impacts. While traditional statistical methods often struggle to capture these multivariate dependencies, deep learning models have emerged as powerful tools for forecasting hydro-climatic time series. However, their utility in Earth system science is currently limited by a lack of transparency. While  ML/DL is useful in predicting extremes, the explainability of the physical mechanisms or compound drivers is limited. Furthermore, standard interpretability techniques applied to geophysical data are often misleading, as they tend to highlight dominant seasonal cycles rather than the dynamic, event-specific interactions that are crucial for scientific discovery. This research proposes a diagnostic framework that repurposes the internal decision-making process of an attention-based encoder-decoder LSTM as a hypothesis generation tool, specifically targeting the latent drivers of extreme and compound events, exemplified here by drought. 

Using multivariate Terrestrial System Modeling Platform simulation data, we trained an attention-based encoder-decoder LSTM where 14 climatological variables serve as both input features and prediction targets in round robin training experiments, generating a comprehensive 14×14 matrix of target-specific attention maps. To transition from predictive modeling to physical interpretation, we apply a post-hoc analysis pipeline to deseasonalize the model's attention weights, which effectively filters out the model’s background behavior, isolating time periods of anomalies. We hypothesize that these anomalies, specifically the extreme 1st and 99th percentile attention, signal instances where standard linear relationships break down. This forces the model to rely on complex, transient feature interactions to maintain predictive accuracy. 

In order to understand the complex dynamics of these events and to disentangle the driving factors from the resultant effects, we employed stacked time series visualisations with multi-scale event windows (±15 and ±90 days). We compared the attention anomalies directly against the anomalies from the simulation results. This granular approach identified distinct attention signatures, revealing dynamic shifts in feature importance, such as an increased focus on surface sensible heat flux and pressure, which were specific to anomalous periods. While our analysis is mainly focused on drought evolution, these synchronized shifts suggest a capacity to reveal the multi-driver interactions of compound events. Consistent patterns across historical events demonstrate that the model’s reliance on specific inputs spikes significantly during these windows, effectively isolating potential compound drivers. By pinpointing exactly when and where system dynamics shift, this framework transforms the LSTM from a passive predictor into an active tool for scientific discovery. It provides domain scientists with targeted starting points for studying the physical precursors of compound climate extremes.