Jerry Lin Dissertation Defense

Title: Scaling Across Multiple Dimensions to Accelerate Subgrid Machine Learning Parameterization Development

Abstract: Subgrid processes in climate models are a dominant source of uncertainty in long-term climate projections. Due to the prohibitive computational expense of direct simulation, these processes are instead parameterized using hand-tuned approximations and assumptions that lead to systematic biases and uncertainty. In theory, Machine-Learning (ML) parameterizations offer a few attractive advantages, namely being able to emulate higher resolution simulations at a fraction of the cost and scaling in performance with more data. However, they also introduce substantial challenges of their own, ranging from the highly empirical nature of their design and evaluation to the emergent behavior they can exhibit when coupled to the dynamical core of a climate model (i.e. tested online). Left unsolved, these challenges threaten the prospect of breaking "deadlock" in convective parameterization development on a timescale in which humanity has not already experienced the worst effects of climate change. This talk encompasses work that seeks to seed a fundamental step change in the progress of ML subgrid parameterization development by embracing scale on two seemingly orthogonal dimensions: online sampling and people. While these may seem like two unrelated directions, they confront the same central, existential problem: the solution space of possible ML parameterizations is too vast, too chaotic, and too poorly understood for any one graduate student or research lab to efficiently explore alone. However, with large scale online sampling, signal can be detected from emergent noise with statistical confidence. And by opening up the ML parameterization problem to seasoned machine learning researchers and data scientists around the world with a standardized dataset, benchmark, software suite, and competition, identifying the neural network architectures and paradigms with the best chance of solving this problem suddenly becomes plausible.