Date of this Version



Deep Learning, Weather Forecast, Chaotic Systems, High-Resolution Data, High-Performance Computing


Weather prediction is indispensable across various sectors, from agriculture to disaster forecasting, deeply influencing daily life and work. Recent advancement of AI foundation models for weather and climate predictions makes it possible to perform a large number of predictions in reasonable time to support timesensitive policy- and decision-making. However, the uncertainty quantification, validation, and attribution of these models have not been well explored, and the lack of knowledge can eventually hinder the improvement of their prediction accuracy and precision. Our project is embarking on a two-fold approach leveraging deep learning techniques (LSTM and Transformer) architectures. Firstly, we model the Lorenz 63 and 96 systems, crucial for grasping chaotic dynamics. By harnessing these neural networks on local computers and the RCAC GPU cluster (Gilbreth), we aim for accurate multi-step forecasts, emphasizing hyperparameter influence on model performance. This research sets a foundation for advanced, transformer-based weather predictions. Secondly, noting the dearth of high-resolution weather data in the US Midwest, including cities like Chicago, we're employing Nvidia's FourCastNet model. Integrated with vision transformers and Adaptive Fourier Neural Operators (AFNOs), it simulates severe Midwest weather events. Using the RCAC's Gilbreth cluster and tapping into the ECMWF Reanalysis (ERA5) dataset, FourCastNet forecasts up to a week ahead in under two seconds, outpacing existing systems. This efficient model promises enhanced weather predictions and extreme event risk assessments. Our goal: simulate the potent January 23, 2016, mid-Atlantic snowstorm and contrast results with traditional forecast models.