Assoc. Prof. Dr. Gregor Skok (FMF): Use of neural networks in atmospheric sciences

Neural networks (NN) are one of the more advanced methods of machine learning. Although the beginnings of NN development date back to the 1940s, their use in atmospheric sciences has been limited until now. The reason for this was that NNs often were not more successful than other simpler approaches, that the computational power available needed to be increased for solving large tasks, and that there needed to be more data available for training, a necessary precondition for successful use of such networks. In recent years, with the advent of more powerful graphics cards containing several thousand computational cores, the development of new architectures of neural networks (e.g., convolutional neural networks), which are more suitable for tasks where input or output data are available in the form of spatial fields, and the publication of extensive databases (e.g., reanalysis - reconstructions of past states of the atmosphere), the situation has changed considerably. Consequently, there has been a significant increase in the use of NN and other machine-learning approaches in atmospheric sciences lately. In the lecture, I will present some of the types of NN that are most used in meteorology. I will also present some interesting examples of NN usage from meteorological literature and our research group.
Cookies, tea