Topic: Machine learning (ML) to break deadlock in achieving climate simulations with high-resolution physics.

Short Description: Climate model predictions are full of uncertainty. Some of it is because processes like cloud formation are crudely represented simply because they are too computationally intensive to model explicitly. Multi-scale climate models that embed small realizations of these explicit physics provide an opportunity to sidestep Moore’s law, by learning these physics with machine learning models. Once trained, the machine learning models can be coupled to the climate predictions, with fast inference allowing high-resolution physics in climate simulations ahead of schedule. An open research challenge is finding reproducible, reliable ways to achieve performant ML workflows in situations of real-world operational complexity. In this context, LEAP has innovated a ML training data set harvested from a state-of-the-art climate simulator, which exposes REU students to the relevant pipeline issues in a real-world research setting at the frontier of climate science and data science.

Learning Outcomes: Familiarity with datasets used in modern climate simulations, baseline “ML Operations” such as quality controlling and data engineering, building training pipelines to perform machine learning, assessing goodness of fit and exploring new algorithms. For those who already come with this basic foundation, advanced learning outcomes could include experimenting with stochastic and generative deep learning algorithms that attempt to fit the non-deterministic component of chaotic cloud dynamics.

Potential Continuation of Research After REU: Successful or standout fits could be considered for follow-on work that goes to the next step of coupling well-performing architectures to global climate simulators. The resulting “hybrid” (ML + physics) climate simulators may exhibit interesting pathologies or physically desirable behaviors, either of which provide  opportunities for collaborative analysis of climate prediction output.