I am increasingly intrigued by the idea that the new tools of modern data science, especially deep neural networks, could have breakthrough potential for ushering in an era of explicit turbulence in climate modeling ahead of schedule towards more faithfully addressing questions like the above. The field of machine learning assisted climate simulation is in its infancy but my group is working on critical-path issues at the frontier like physical credibility, hybrid physics-data-driven modeling, and mapping out trade-offs when emulators are allowed to couple with explicit fluid dynamics.
Following our initial proof of concept (Rasp et al. 2018), we have now developed a new neural network architecture algorithm that allows emulators of explicit convection physics to further enforce important physical constraints like mass and energy conservations to machine precision (Beucler et al. 2019). We also created new software to ease the process of testing DNN emulators trained in python software environments as predictive kernels within fortran-based physical models (Ott et al. 2020) where their coupled trade-offs when interacting with fluid dynamics are revealed, a major current frontier in weather and climate prediction. This work is in increasing collaboration with UCI computer science Prof. Pierre Baldi, and now also through co-advisement of CS PhD and MS students. It is highly collaborative with Prof. Pierre Gentine at Columbia University.
Please see this blog post for the Department of Energy’s E3SM climate modeling project from August 2020 for a snapshot of our machine learning research.