A forward look at the alteration of the weather forecast by artificial intelligence

Abstract

Most modern weather forecasts are accomplished by atmospheric simulations based on physics models. This method requires a lot of processing time due to the massive amount of computing work. Meanwhile, accuracy faces a limit due to the chaotic nature of the atmosphere. As machine learning gets better and becomes more involved in society, which has been an accelerated trend in recent years, some machine learning-based weather forecast models have been developed. This paper provides a forward-looking study of the potentials of machine learning in terms of theoretical abilities, shown advantages, and possible alterations in the field of weather forecasting based on existing literature and data. Explanations and discussions of two outstanding machine learning-based models, MetNet-2 and Pangu, are provided. Based on the results of tests of these two models, machine learning has good advantages that can be largely applied in the field of weather forecasting and other atmosphere-related work. Machine learning-based methods are likely going to replace physics models in most parts, while physics knowledge is still valuable in this field as it can help with the training of machine learning models and improve performance.