Work Summary: Accurate forecasting of weather and climate depends on physics-based numerical simulations that are computationally expensive and require some of the fastest supercomputers in the world. In an effort to reduce computational cost, researchers are exploring deep learning to work in combination with the current numerical models. However, comparisons between deep learning models are difficult because they are trained and optimized on different datasets. In this project, we develop a benchmark to facilitate comparisons among deep learning architectures We use data from the Multi-RADAR/Multi-sensor System (MRMS) which contains precipitation rates in 5-minute intervals, for a 1 Mi sq. km region within Southwestern U.S. from 2001 to 2011. With our dataset, we design and test a convolutional long short-term memory (convLSTM) deep- learning architecture and defined this as our baseline model. Ultimately, the benchmark dataset can be used to compare the performances of other deep learning models with our baseline convLSTM implementation.

Tools Used: Python, Git, LaTeX, Slurm, Harovod, Tensorflow

Github URL: github.com/gkroiz/weatherDNN