Interpretable AI predicts a 2026 summer dry anomaly in central China

Seasonal precipitation predictions in China remain challenging due to complex interactions among atmospheric, oceanic, and land-surface processes. This study investigates whether a deep learning framework can provide credible, physically interpretable real-time predictions of regional summer precipitation anomalies for 2026 aheadofthetargetseason.

The authors employ a circulation-to-precipitation deep learning bridging model that translates dynamical circulation predictions into regional precipitation estimates. Predictions initialized between March and May 2026 are evaluated using multiple lines of evidence. First, historical analogue analyses identify past years resembling the 2026 anomaly to assess model skill under similar climate conditions. Second, composite analyses examine the large-scale atmospheric circulation and sea surface temperature patterns associated with the predicted drying. Finally, layer-wise relevance propagation (LRP) and perturbation tests are used to verify which input features drive the model's prediction.

Predictions initialized from March to May consistently indicate a substantial dry anomaly over central China in summer 2026, with regional-mean reductions ranging from 29% to 38%. Historical evaluations reveal that the bridging model achieves higher predictive skill in analogue years featuring central equatorial Pacific warming. This warming favors an anomalous cyclonic circulation over the western North Pacific and South China Sea, inducing northerly winds and moisture divergence that suppress rainfall over central China. Feature attribution via LRP confirms that the model relies heavily on these northerly winds, and LRP-guided perturbation tests demonstrate that removing these features effectively eliminates the predicted dry anomaly.

By combining deep learning forecasts with explainable AI and physical climate diagnostics, this study offers a rigorous framework for assessing the credibility of regional climate projections before observational data become available. This approach helps decision-makers evaluate seasonal forecasts based on transparent physical mechanisms rather than relying solely on black-box model outputs.