Talk abstract:
Global atmospheric `storm-resolving' models with horizontal grid spacing of less than ~5km resolve deep cumulus convection and flow in complex terrain. While computationally expensive, they can be run for yearlong scales and can serve as reference models for improving more economical coarse-grid global weather and climate models. Machine learning (ML) offers an avenue for translating the patterns seen in storm-resolving models onto the coarser grid, with the ultimate goal of reducing uncertainties in regional precipitation and temperature trends in global climate models. AI2's Climate Modeling group is undertaking an effort to apply this approach in an operational weather and climate model, among other efforts to leverage ML in climate modeling. Here we describe an experiment to machine learn the corrections between a coarse and storm-resolving model as functions of coarse model state, in order to improve upon the physical parameterizations of temperature, humidity, and winds, in a real-geography coarse-grid model (FV3GFS with a 200 km grid). We evaluate whether the ML correction improves the coarse model's weather forecast metrics and spatial distributions of temperature and precipitation. The best configuration of ML uses learned nudging for temperature and humidity but not winds, with neural nets slightly outperforming random forests. Forecasts of 850 hPa temperature gain 18 hours of skill at 3-7 day leads and time-mean precipitation patterns are improved 30% by applying the ML correction. Ongoing work focuses on improving climate-scale metrics such as seasonal and multi-year temperature and precipitation biases.
All seminars this quarter will be in a hybrid format.