Last weekend AI researchers and enthusiasts from around the world gathered in Prague for this year’s edition of ML Prague – the biggest European conference about ML, AI, and Deep Learning applications. At Meteopress, we are pleased that we can be part of Prague becoming one of the significant European AI hubs alongside other companies, universities, and initiatives, and we couldn’t miss this prestigious event.
On Friday, our senior AI researcher Matej Choma held a workshop titled “Predicting Weather With Deep Learning” in cooperation with researchers Petr Šimánek, Jiří Pihrt, and Miroslav Čepek from DataLab of the Faculty of Information Technology at CTU Prague. We focused the workshop on radar echo nowcasting – short-time prediction of precipitation in high-resolution from weather radar observations. Together, with more than 50 participants, we learned the basics of working with sequential radar reflectivity data and implemented a recurrent ConvLSTM for next-step prediction in PyTorch.
Precipitation is a process in the atmosphere and is thus guided by laws of physics. We have presented various ways of combining physical models with neural networks and demonstrated the usefulness of these approaches in a simple example. We ended the workshop by implementing the advection-diffusion equation into the PhyDNet architecture, utilizing the gained knowledge in the precipitation nowcasting scenario.
On Sunday, during the third and final day of the conference, our AI researcher Matej Murín gave a talk on the main stage of the event. The topic was about PhyDGAN, our recent research about improving upon PhyDNet even more by utilizing the GAN framework. This type of neural network is able to produce realistically looking, deep physics-constrained predictions that were shown to be on par or superior to previous popular methods of probabilistic nowcasts.
It was an excellent opportunity to meet with not only Prague’s AI community. Thank you to the organizers; we are already looking forward to the following year.