An efficient and accurate deep learning approach to weather prediction

Machine learning is being increasingly applied as a surrogate modeling technique for weather prediction, providing fast forecasts with similar accuracy to numerical weather prediction models. However, developing accurate state-of-the-art machine learning models requires a significant allocation of high-performance computing resources for processing datasets and training. In this work, we investigate the essential components of a deep learning model architecture for accurate weather prediction and formulate strategies that reduce the number of parameters needed in such a model based on physical assumptions to lower training time. Specifically, we investigate autoencoder architectures with convolutional and attention-based neural network layers for capturing the necessary information provided by weather data for prediction. These architectures are incorporated within the neural ordinary differential equations framework and then trained based on reanalysis data constructed from simulation and observation data to provide forecasts. The results and conclusions based on these experiments are discussed, and recommendations for future work are provided.