Evaluation of Tropical Cyclone Track and Intensity Forecasts from Purely ML-based Weather Prediction Models, Illustrated with FourCastNet

In the past few years, advanced Machine Learning (ML) techniques have provided an alternative to physically-based numerical weather prediction (NWP) models. FourCastNet, developed by the NVIDIA Corporation, is an ML-based model that predicts meteorological fields on a global 0.25 deg latitude/longitude grid out to seven days. The model was trained on ECMWF ERA5 reanalysis data for the period 1979-2015. Each forecast requires a global model analysis field for initialization, such as from the ECMWF or Global Forecast System (GFS) models, but does not require any additional input for the prediction. Evaluation by the development team showed that this model has comparable accuracy to state-of-the-art NWP models such as the ECMWF for large scale atmospheric variables. However, FourCastNet requires about four orders of magnitude less computational resources than current operational physically-based NWP models. Previous research has also demonstrated that the FourCastNet fields contain tropical cyclones, although the inner core is not fully resolved.

In the presentation, the implementation of the open-access version of FourCastNet at CIRA will be described, including time tests. The model is initialized with GFS analysis fields and takes about 30 seconds to make a seven-day forecast on a modest Linux workstation with one GPU. A simplified tropical cyclone (TC) tracker was adapted to the FourCastNet forecast fields. The tracker only requires horizontal wind fields at 850 hPa and 10 m and sea-level pressure fields, all of which are available from the FourCastNet predictions. The tracker is being run for global TCs during 2023 hurricane and typhoon seasons. The TC track and intensity error statistics will be compared with those from several operational global models, regional hurricane models, and statistical models. Vortex structure is evaluated by comparing the pressure-wind relationships from the model with those from observations. Potential applications of the FourCastNet TC prediction system will also be described, including very fast ensemble predictions. Plans are also underway to compare the TC forecast results from FourCastNet with those from other open source ML-based weather prediction models. The comparison with operational hurricane models is an important first step towards possible uses of ML-based models for real time forecasting.