Too Rare to Learn: Prescribed Cyclone Tracks Degrade a Bay of Bengal Ocean Emulator

The research team found during tests in the Bay of Bengal that using hurricanes as preset inputs for neural network ocean simulators impaired forecasting accuracy. The researchers removed 15 annual hurricanes with wind speeds between 65 and 150 knots from the GLORYS12 reanalysis data and compared two identical U-Net models, which differed only in four preset hurricane trajectory channels. The results showed that in all three seed runs, ocean models without storms performed better than those using continuous prediction methods. In contrast, models with preset storms lagged behind in each run, and their skill ranges did not overlap (p = 3.1e-5). The reason for this issue was exposure frequency rather than signal content: these channels were non-zero only in 7.9% of training days and fell outside the distribution once activated. Additional errors were concentrated within the footprints of preset storms; during the inference phase, using storm-free maps to replace real hurricane maps improved storm forecasting accuracy by 7.5% to 16.4% in each seed run. This indicates that the conditional network has learned to respond to rare and erroneous signals.