Do Climate Models Need Microphysical and Convective Parameterizations to Generate Accurate Precipitation Fields?

Abstract

Accurately representing surface precipitation is crucial for the operational use of weather and climate models. Presently, global numerical weather prediction (NWP) models struggle to accurately generate precipitation due to their parametrization of unresolved deep convective clouds and, in regions of grid-resolved ascent, inadequate parameterizations of cloud microphysics. Here we bypass these parameterizations with a machine learning model that diagnoses precipitation from 13 ERA5 fields that are easily observed and assimilated, as opposed for example, to fields like rain or cloud liquid water. We train a pair of models; ML_ERA5 using ERA5 precipitation as the target, and ML_IMERG using a satellite based precipitation product. ML_ERA5 closely reproduces the ERA5 precipitation at all intensities. When evaluated against the satellite dataset, ML_IMERG closely matches observations, notably reproducing the diurnal cycle of the satellite product. ML_IMERG generally captures extremes better than ERA5 while also reducing ERA5's overproduction of light precipitation. When evaluated against a third ground-and-radar-based dataset, ML_IMERG inherits the strengths of the satellite dataset which is superior to ERA5 in the summer months.

Figures (16)

• Figure 1: Diurnal cycle of precipitation rate anomalies relative to the local daily mean values for 2022-2024 over 50N-50S for ERA5 (red), IMERG (black) and our ML model trained on the IMERG data (blue). Solid lines are precipitation over land and dotted is precipitation over the ocean.
• Figure 2: Diagram of CNN U-net used for the ML precipitation model. ConvNeXt block (blue block) layout shown in bottom left. Dark green blocks represent layers evaluated by the encoder while the lighter green blocks represent layers evaluated by the decoder.
• Figure 3: Example 3-hour totals beginning at 00 UTC on at April 30, 2022 for (a) ML$_{\rm ERA5}$ and (b) ERA5. (c) Performance diagram for ML$_{\rm ERA5}$ 3-h totals (navy) and daily totals (light blue) showing skill by intensity threshold. (d) Temporal correlation of 3-h totals from ML$_{\rm ERA5}$ with ERA5 estimates.
• Figure 4: Skill metrics for ML$_{\rm ERA5}$ against ERA5 validation over 70N-70S and the test set (2022-2024). (a) Categorical fraction skill score for daily totals from ML$_{\rm ERA5}$ (light blue) and daily nominal skill thresholds (dashed grey) are plotted for the intensity categories on the top axis. 6-h accumulation categorical FSS from ML$_{\rm ERA5}$ (navy) and nominal skill thresholds (dashed black) are also plotted (bottom axis). (b) Equitable threat score for ML$_{\rm ERA5}$ by intensity threshold for 6-hour (navy, bottom axis) and 24-h totals (light blue, top axis). The dashed grey curve corresponds to the "perfect model" prediction of Goeber2008 for daily totals. (c) 1-SEEPS skill score for 24-h totals from ML$_{\rm ERA5}$ (blue) for all times in the test set. The dashed black line represents the approximate 1-SEEPS score of a "perfect model" Rodwell2010.
• Figure 5: Example 24-h totals for July 30, 2023 from (a) ML$_{\rm IMERG}$ , (c) IMERG validation, and (d) ERA5. Temporal correlations between IMERG and (b) ML$_{\rm IMERG}$ or (e) ERA5. Note that cooler tones show regions of low correlation.