Neural General Circulation Models for Weather and Climate

TL;DR

This paper presents a hybrid model that combines a differentiable solver for atmospheric dynamics with machine-learning components and shows that it can generate forecasts of deterministic weather, ensemble weather and climate on par with the best machine-learning and physics-based methods.

Abstract

General circulation models (GCMs) are the foundation of weather and climate prediction. GCMs are physics-based simulators which combine a numerical solver for large-scale dynamics with tuned representations for small-scale processes such as cloud formation. Recently, machine learning (ML) models trained on reanalysis data achieved comparable or better skill than GCMs for deterministic weather forecasting. However, these models have not demonstrated improved ensemble forecasts, or shown sufficient stability for long-term weather and climate simulations. Here we present the first GCM that combines a differentiable solver for atmospheric dynamics with ML components, and show that it can generate forecasts of deterministic weather, ensemble weather and climate on par with the best ML and physics-based methods. NeuralGCM is competitive with ML models for 1-10 day forecasts, and with the European Centre for Medium-Range Weather Forecasts ensemble prediction for 1-15 day forecasts. With prescribed sea surface temperature, NeuralGCM can accurately track climate metrics such as global mean temperature for multiple decades, and climate forecasts with 140 km resolution exhibit emergent phenomena such as realistic frequency and trajectories of tropical cyclones. For both weather and climate, our approach offers orders of magnitude computational savings over conventional GCMs. Our results show that end-to-end deep learning is compatible with tasks performed by conventional GCMs, and can enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system.

Figures (50)

• Figure 1: Structure of the NeuralGCM model. (a) Overall model structure, showing how forcings $F_t$, noise $z_t$ (for stochastic models), and inputs $y_t$ are encoded into the model state $x_t$. Model state is fed into the dynamical core, and alongside forcings and noise into the learned physics module. This produces tendencies (rates of change) used by an implicit-explicit ODE solver to advance the state in time. The new model state $x_{t+1}$ can then be fed back into another time step, or decoded into model predictions. (b) Inset of the learned physics module, which feeds data for individual columns of the atmosphere into a neural network used to produce physics tendencies in that vertical column.
• Figure 2: Weather forecasting accuracy scores for deterministic and stochastic models. (a) RMSE and (c) RMSB for ECMWF-ENS, ECMWF-HRES, NeuralGCM-$0.7^{\circ}$, NeuralGCM-ENS, GraphCast lam2022graphcast and Pangu bi2023accurate on the main WeatherBench variables, as a percent of the error of ECMWF-ENS. Deterministic and stochastic models are shown in solid and dashed lines respectively. (e) CRPS relative to ECMWF-ENS and (g) skill-spread ratio for ENS and NeuralGCM-ENS models. Spatial distributions of (b) RMSE, (d) Bias, (f) CRPS and (h) spread-skill ratio for NeuralGCM-ENS and ECMWF-ENS models for $10$-day forecasts of specific humidity at 700 hPa. Spatial plots of RMSE and CRPS are relative to a probabilistic climatology with an ensemble member for each of the years 1990-2019. Grey areas indicate regions where climatological surface pressure on average is below 700hPa.
• Figure 3: Case study of a medium-range weather forecast. All forecasts are initialized at 2020-08-22T12z, chosen to highlight Hurricane Laura, the most damaging Atlantic hurricane of 2020. (a) Specific humidity at 700 hPa for 1-day, 5-day and 10-day forecasts over North America and the North-East Pacific ocean from ERA5, ECMWF-HRES, NeuralGCM-$0.7^{\circ}$, ECMWF-ENS (mean), NeuralGCM-ENS (mean), GraphCast lam2022graphcast and Pangu bi2023accurate. (b) Forecasts from individual ensemble members from ECMWF-ENS and NeuralGCM-ENS over regions of interest, including predicted tracks of Hurricane Laura from each of the 50 ensemble members (Appendix \ref{['apx:sec:TC_tracking']}). The track from ERA5 is plotted in black.
• Figure 4: Simulation of climate with NeuralGCM. (a) Global mean temperature for ERA5 for 2020 (orange), climatology (defined as the averaged temperature between 1990-2019; green), and for NeuralGCM-$1.4^{\circ}$ for 2020 for 35 simulations initialized every 10 days during 2019 (thick blue represents the ensemble mean; thin blue lines indicate different initial conditions). (b) Yearly global mean temperature for ERA5 (orange), mean over 22 CMIP6 AMIP experiments for 1981-2014 (purple; model details are found in Appendix \ref{['apx:sec:AMIP_models_used']}), and NeuralGCM-$2.8^{\circ}$ for 22 AMIP-like simulations with prescribed SST initialized every 10 days during 1980 (thick blue represents the ensemble mean, and thin blue lines indicate different initial conditions). (c) The root mean square bias (RMSB) of the 850hPa temperature averaged between 1981-2014 for 22 NeuralGCM-$2.8^\circ$ AMIP runs (labeled NGCM), 22 CMIP6 AMIP experiments (labeled AMIP) and debiased 22 CMIP6 AMIP experiments (labeled as AMIP*; biased was removed by removing the 850hPa global temperature bias). In the box plots, the red line represents the median; the box delineates the first to third quartiles; the whiskers extend to 1.5 times the interquartile range (Q1 - 1.5IQR and Q3 + 1.5IQR), and outliers are shown as individual dots. (d) Vertical profiles of tropical (20S-20N) temperature trends for the period 1981-2014. The orange line shows ERA5 reanalysis, and the black dots show the trends calculated from Radiosonde Observation Correction using Reanalyses haimberger2008toward. The blue dots shows the mean trends for NeuralGCM-$2.8^{\circ}$ 22 AMIP-like runs and purple dots are the mean trends from CMIP6 AMIP runs (see appendix \ref{['apx:sec:AMIP_models_used']} for full list of models used), where the grey (black) whiskers show the 25th and 75th percentiles for NeuralGCM-$2.8^{\circ}$ (CMIP6 AMIP runs). (e,f,g) Tropical cyclone tracks for (d) ERA5, (e) NeuralGCM-$1.4^{\circ}$ and (f) X-SHiELD. (h) Mean precipitable water for ERA5 and the precipitable water bias in (i) NeuralGCM-$1.4^{\circ}$; initialized 90 days before mid-January 2020 similarly to X-SHiELD, (j) X-SHiELD and (j) climatology (averaged between 1990-2019). In panels d-i quantities are calculated between mid-January 2020 and mid-January 2021 (when X-SHiELD model data is available) and all models were regridded to a 256x128 Gaussian grid before computation and tracking.
• Figure 5: Visualization of the data flow in the learned physics module of NeuralGCM.