Data-driven ocean, atmosphere, and land parameterizations calibrated from indirect data

The multiscale nature of climate necessarily requires approximations within all subcomponents, and may be based on machine learning, physics or both. Our approach to make accurate and trustworthy predictions of future, never-observed climate regimes combines physics-based models with subcomponents learned from accessible data, which are often only indirectly informative about the modeled processes. We have developed a suite of model-agnostic machine learning tools to learn about subcomponent models from such data. These tools rely on ensembles of model simulations, effectively carried out on GPUs in distributed systems; and our framework for assessing uncertainty of calibrations (CalibrateEmulateSample) requires 1,000 times fewer model evaluations than traditional approaches. Our approach has produced several scientific successes, such as a unified turbulence and cloud parameterization calibrated with a library of large-eddy simulations, a neural network snow model trained from station data, and calibration-directed development of a parameterization for upper-ocean turbulence (CATKE).