Skillful Seasonal Tropical Cyclone Prediction with AI Weather--Climate Models Forced by Dynamically Predicted SSTs

Hiroyuki Murakami , Baoqiang Xiang

AI weather-climate models are far cheaper than dynamical seasonal prediction models and have shown skillful tropical cyclone (TC) prediction under persisted sea surface temperature (SST) anomalies, but their skill under dynamically predicted ocean boundary conditions, benchmarked against an established dynamical seasonal prediction system, is untested. We evaluate eight seasonal hindcast configurations against GFDL-SPEAR, the coupled model behind GFDL's real-time forecasts, over 1993-2025 for the July-November season in the western North Pacific, eastern North Pacific, and North Atlantic. The hybrid NeuralGCM and data-driven ACE2 use predicted SST and sea ice from operational models; six NeuralGCM configurations differ only in their SST source. Their advantage over SPEAR is basin-, metric-, and lead-dependent. At zero-month lead, six of seven AI-based configurations achieve North Atlantic correlations of 0.60-0.76, versus SPEAR's 0.59. At the grid-point scale, nearly every AI-based configuration exceeds SPEAR's significant-skill area in the eastern North Pacific and North Atlantic by up to 20-25 percentage points, and one retains skill at four-month lead where SPEAR does not; SPEAR remains unmatched for accumulated cyclone energy in the western North Pacific. TC prediction skill reflects a complex interplay among biases in a configuration's TC climatology, large-scale environment, and genesis response to it, so higher skill can arise from compensating biases rather than a physically valid mechanism. Overall, the rapid advancement of AI weather--climate models, together with the enduring value of dynamical models, suggests considerable potential for hybrid dynamical-AI approaches to advance seasonal TC prediction.