Vadim Rezvov: Rehearsal Talk Before PhD Defense at the Shirshov Institute of Oceanology

At the 34th seminar of the Center for Earth Sciences, Vadim Rezvov presented a rehearsal version of his upcoming PhD defense talk, prepared at the Shirshov Institute of Oceanology and the Laboratory of Machine Learning in Earth Sciences at MIPT. His research explores neural network–based approaches to statistical downscaling, aiming to generate high-resolution atmospheric data without running additional numerical models. Focusing on the Barents and Kara Seas, he showed that convolutional neural networks trained on ERA5 reanalysis and WRF model outputs can effectively reconstruct mesoscale atmospheric features such as polar mesocyclones and the Novaya Zemlya bora. The method achieves over fiftyfold computational efficiency compared to traditional high-resolution modeling while maintaining realistic dynamical structures. The seminar took place in an open discussion format, helping the speaker refine both scientific content and delivery ahead of his formal PhD defense at the Shirshov Institute of Oceanology.