O3CastNet: Global Ozone Forecasting from Atmospheric Predictions Without Ozone Input

Global ozone forecasts commonly depend on an initialized chemical state. We present O3CastNet, a diagnostic neural network that estimates ozone from two consecutive atmospheric states and their timestamps, without ozone input. The diagnostic operates on weather model outputs and can in principle be adapted to different medium range forecast systems. Here, FuXi supplies atmospheric predictors from ECMWF operational analyses and remains fixed while a single diagnostic model is trained for all leads. The ozone output covers nine pressure levels from 250 to 1000 hPa. Training spans 2003 to 2022, validation uses 2023, and testing covers all initializations in the 2024 test set at leads from 6 to 120 h. Verification against EAC4 shows lower root mean square error than CAMS throughout the forecast range at 500 to 1000 hPa. At 250 and 300 hPa, CAMS retains an advantage at longer leads. Spatial diagnostics reveal broad regional improvements and positive temporal anomaly correlation, with substantial geographical variation. These results demonstrate useful ozone prediction from forecast meteorology and time information without chemical initialization. They establish the performance of the FuXi implementation, while transfer to other weather models remains to be evaluated.