Warm East Asian winters become more predictable in a warmer climate; cold ones do not

Why it matters

Seasonal forecasts of unusually warm or cold winters matter for agriculture, power supply and hazard planning. The work suggests that the sources of seasonal predictability may shift as the climate warms. Models trained on past data may therefore not carry over unchanged.

What they did

The authors trained a convolutional neural network on CESM2 large ensemble simulations, historical and SSP3-7.0. It used 14 global November fields to predict winter (December to February) temperature over East Asia. A loss function called amplitude focal loss pushed the network to focus on extreme winters. They also fine-tuned the network on ERA5 reanalysis and compared it with the SMYLE forecast system. They then used Integrated Gradients to see which inputs the network relied on in each climate.

Key findings

• Against observations, the fine-tuned network reached an ensemble-mean correlation of about 0.5, versus 0.4 for the SMYLE physics-based system.
• In the model world, median correlation rose from about 0.38 in the historical period to 0.44 in the future, and median RMSE fell from about 1.16 °C to 1.06 °C.
• Skill for the warmest 5% of winters reached nearly 0.3 (Heidke skill score) in the future, while cold-winter skill stayed modest and broadly unchanged.
• Without amplitude focal loss, skill for extremes was close to zero.
• Attribution pointed to tropical Pacific sea surface temperature. In the future, warm El Niño-like anomalies drive a stronger circulation response toward East Asia than cold anomalies do.

Limitations

• Only one climate model (CESM2-LE) was used, and the future results come from a perfect-model setup, so model biases in ENSO and teleconnections may carry into the findings.
• Integrated Gradients show what the network relies on, not proven physical causes, and other drivers beyond the tropical Pacific were not examined.
• The observational skill may be slightly inflated, because leave-one-out testing assumes independent winters but adjacent winters are modestly correlated.

Glossary

• Integrated Gradients: An explainable-AI method that scores how much each input value contributes to a model’s prediction.
• Amplitude focal loss: A training loss that gives more weight to large-amplitude targets, so the model pays more attention to extremes.
• Heidke skill score: A score for yes/no event forecasts where 0 means no better than chance and 1 means perfect.
• Perfect-model framework: Training and testing on the same climate model’s output, so the model is treated as if it were reality.