Abstract
The problem of forecasting weather has been scientifically studied forcenturies due to its high impact on human lives, transportation, foodproduction and energy management, among others. Current operational forecastingmodels are based on physics and use supercomputers to simulate the atmosphereto make forecasts hours and days in advance. Better physics-based forecastsrequire improvements in the models themselves, which can be a substantialscientific challenge, as well as improvements in the underlying resolution,which can be computationally prohibitive. An emerging class of weather modelsbased on neural networks represents a paradigm shift in weather forecasting:the models learn the required transformations from data instead of relying onhand-coded physics and are computationally efficient. For neural models,however, each additional hour of lead time poses a substantial challenge as itrequires capturing ever larger spatial contexts and increases the uncertaintyof the prediction. In this work, we present a neural network that is capable oflarge-scale precipitation forecasting up to twelve hours ahead and, startingfrom the same atmospheric state, the model achieves greater skill than thestate-of-the-art physics-based models HRRR and HREF that currently operate inthe Continental United States. Interpretability analyses reinforce theobservation that the model learns to emulate advanced physics principles. Theseresults represent a substantial step towards establishing a new paradigm ofefficient forecasting with neural networks.