Emulators, correctors, and crashes: Improving FV3GFS weather and climate prediction using machine learning
Parameterizations of the unresolved physical processes such as clouds, turbulence, and convection are a large source of uncertainty for prediction and complexity for atmospheric models. To improve predictions, we might add more complexity to the parameterization, or increase the model resolution and remove the need for a parameterization altogether. In either case, the increase in accuracy has a tradeoff in cost, which is sometimes quite severe. Recent breakthroughs in prediction for vision or language-related tasks have shown the power of machine learning. With enough data and compute power upfront, one can create compact and skillful models capable of extraordinary things. Could modeling of physical systems see the same type of breakthrough?
In this talk, I will discuss work from our team at the Allen Institute for Artificial Intelligence (Ai2) in partnership with GFDL where we are exploring how to use machine learning to improve weather and climate prediction. I will discuss our experience with developing ML tools for use with FV3GFS and cloud computing, our successes, and of course, challenges encountered. In the first part, I will cover our search for stable and skillful replacements for some or all of the coarse-grid physics using neural-net emulators. Then, I will cover how we’ve improved forecast skill and time-mean precipitation by making corrective-adjustments to the physics with information from observations or fine-resolution simulations. In each case, the hurdles teach us a lot about how to implement ML into existing physics-based predictive models.