climate science. She already found that the method they developed could be used in another project involving aerosol modeling with deep learning, where she replaced part of a climate model with a neural network, aiming to conserve the aerosol mass within each time step. The same constraint layer helped solve that problem, and she anticipates that many other problems could use it. Will Paula be tackling some of these problems herself? “It’s going to be a mix,” she responds. “I’m still involved in some projects that use that, but I also know some people who took the idea and are working on it themselves, and I’m very happy about that. I just hope that it can help the community get closer to the goal of having good climate forecasts for the future.” Paula is keen to continue advancing deep learning models for applications in climate and weather modeling and super-resolution downscaling. In particular, how models can be improved to perform better when applied to new locations. “Sometimes, we have great data to train on in one location of the world, and then we apply the deep learning model in a different location, and it completely fails,” she tells us. “How can we make the models more robust under distribution shifts? I don’t have a solution yet, but it’s something I’m really curious to explore!” 47 Computer Vision News Computer Vision News Climate Downscaling
RkJQdWJsaXNoZXIy NTc3NzU=