The School of Earth and Atmospheric Sciences Presents, Dr. Maria J. Molina, NCAR

How Robust is Deep Learning in a Changing Climate?

Computational advances have enabled substantial progress in large-data analysis and deep learning for the atmospheric, climate, and related sciences. These tools can be leveraged to identify and classify features in the Earth system with skill. However, as the climate continues to change, deep learning models may not perform with skill when faced with extreme events that lie outside of the distribution of the training data.

This seminar will highlight a benchmark study that evaluated the ability of a convolutional neural network to skillfully classify severe convective storms in high-resolution model output representative of a future climate (late 21st century). Explainable artificial intelligence methods, such as saliency maps, were also applied to the trained model, which revealed that the convolutional neural network learned physical information that was not prescribed during the training process