We present a machine learning based emulator of a microphysics scheme for condensation and precipitation processes (Zhao-Carr) used operationally in a global atmospheric forecast model (FV3GFS). Our tailored emulator architecture achieves high skill (≥94%) in predicting condensate and precipitation amounts and maintains low global-average bias (≤4%) for 1 year of continuous simulation when replacing the Fortran scheme. The stability and success of this emulator stems from key design decisions. By separating the emulation of condensation and precipitation processes, we can better enforce physical priors such as mass conservation and locality of condensation, and the vertical dependence of precipitation falling downward, using specific network architectures. An activity classifier for condensation imitates the discrete-continuous nature of the Fortran microphysics outputs (i.e., tendencies are identically zero where the scheme is inactive, and condensate is zero where clouds are fully evaporated). A temperature-scaled conditional loss function ensures accurate condensate adjustments for a high dynamic range of cloud types (e.g., cold, low-condensate cirrus clouds or warm, condensate-rich clouds). Despite excellent overall performance, the emulator exhibits some deficiencies in the uppermost model levels, leading to biases in the stratosphere. The emulator also has short episodic skill dropouts in isolated grid columns and is computationally slower than the original Fortran scheme. Nonetheless, our challenges and strategies should be applicable to the emulation of other microphysical schemes. More broadly, our work demonstrates that with suitable physically motivated architectural choices, ML techniques can accurately emulate complex human-designed parameterizations of fast physical processes central to weather and climate models.
We use online data assimilation to combine information from a linear inverse model of coupled atmosphere-ocean dynamics with proxy records to create a new annual-resolution reconstruction of atmosphere and ocean fields over the last millennium. Instrumental validation of reconstructed sea-surface temperature and 0-700 m ocean heat content shows broad regions of positive spatial correlations, and high correlations (~0.6-0.9) for global averages and indices of large-scale modes of atmospheric variability. Compared to previous reconstructions, the online reconstructions show global and hemispheric averages with little-to-no millennial-scale trend and global-mean temperatures ~0.25-0.5 K cooler during early periods (1000-1400 C.E.). The spatial anomaly differences of average temperature between an early (1000-1250 C.E.) and later (1400-1700 C.E.) period show warm anomalies over high-latitude Europe and cool tropical conditions in partial agreement with previous assessments. The addition of online data assimilation, which provides dynamical memory to climate proxy information, is shown to be crucial for adequately characterizing decadal-to-centennial-scale variability of 0–700 m ocean heat content. Furthermore, the climate forecasts provide model-based physical constraints for atmosphere-ocean interaction, which become increasingly important during early periods when less proxy information is available for assimilation.
Due to limited resolution and inaccurate physical parameterizations, weather and climate models consistently develop biases compared to the observed atmosphere. These biases are problematic for forecasting on timescales from medium-range weather to centennial-scale climate. Using the FV3GFS model at coarse resolution, we propose a method of machine learning corrective tendencies from a hindcast simulation nudged towards an observational analysis. We show that a random forest can predict the nudging tendencies from this hindcast simulation using only the model state as input. This random forest is then coupled to FV3GFS, adding corrective tendencies of temperature, specific humidity and horizontal winds at each timestep. The coupled model shows no signs of instability in year-long simulations and has significant reductions in short-term forecast error for 500hPa height, surface pressure and near-surface temperature. Furthermore, the root mean square error of the annual-mean precipitation is reduced by about 20%.