Source-linked AI summary
Deep learning to represent sub-grid processes in climate models
Stephan Rasp, Michael S. Pritchard, Pierre Gentine
TL;DR
Climate models need better representations of atmospheric sub-grid processes, while cloud-resolving simulations remain computationally costly. This paper trains a deep neural network on cloud-resolving model data and finds that it produces stable climate simulations close to the training climate, while struggling with out-of-sample climates.
Problem
Climate models require improved representations of atmospheric sub-grid processes, motivating data-driven alternatives to manually designed parameterizations.
Method
A single deep neural network learns all atmospheric sub-grid processes from cloud-resolving model data and replaces traditional parameterizations in a global model.
Results
The neural-network model produces a stable mean climate close to its training climate, captures precipitation extremes and tropical waves, and approximately conserves energy.
Takeaways & Limitations
Deep learning offers a potential data-driven approach for developing climate and weather models from high-resolution simulations.
Takeaways & Limitations
The neural network struggles with temperatures beyond its training range, although training on extreme climates enables interpolation between them.
Abstract
from arXiv · showhide
The representation of nonlinear sub-grid processes, especially clouds, has been a major source of uncertainty in climate models for decades. Cloud-resolving models better represent many of these processes and can now be run globally but only for short-term simulations of at most a few years because of computational limitations. Here we demonstrate that deep learning can be used to capture many advantages of cloud-resolving modeling at a fraction of the computational cost. We train a deep neural network to represent all atmospheric sub-grid processes in a climate model by learning from a multi-scale model in which convection is treated explicitly. The trained neural network then replaces the traditional sub-grid parameterizations in a global general circulation model in which it freely interacts with the resolved dynamics and the surface-flux scheme. The prognostic multi-year simulations are stable and closely reproduce not only the mean climate of the cloud-resolving simulation but also key aspects of variability, including precipitation extremes and the equatorial wave spectrum. Furthermore, the neural network approximately conserves energy despite not being explicitly instructed to. Finally, we show that the neural network parameterization generalizes to new surface forcing patterns but struggles to cope with temperatures far outside its training manifold. Our results show the feasibility of using deep learning for climate model parameterization. In a broader context, we anticipate that data-driven Earth System Model development could play a key role in reducing climate prediction uncertainty in the coming decade.
Model and neural network setup
The study uses a global aquaplanet climate model with approximately 2-degree horizontal resolution, 30 vertical levels, and a 30-minute timestep, replacing traditional sub-grid parameterizations with a neural network. The neural network runs substantially faster while reproducing key aspects of the training model’s climate, conserving energy approximately, and showing limited generalization to out-of-sample climates.
- Model setup: SPCAM uses fixed, zonally invariant sea-surface temperatures with an equator-to-pole gradient, a full diurnal cycle without seasonal variation, approximately 2-degree horizontal spacing, 30 vertical levels, and a 30-minute timestep.The model is configured as an aquaplanet.
- Neural network setup: NNCAM’s physical parameterization is around 20 times faster than SPCAM’s and 8 times faster than NNCAM’s radiation-inclusive configuration.Prediction cost does not increase when the network is trained with higher-resolution data.
- Climate reproduction: NNCAM closely reproduces SPCAM’s mean sub-grid heating, moistening, temperature, humidity, wind, radiative fluxes, and precipitation, while avoiding CTRLCAM’s double ITCZ bias.NNCAM’s ITCZ heating and precipitation peaks are slightly sharper than SPCAM’s, and its larger temperature deviations occur mainly in the stratosphere.
- Climate variability: NNCAM closely matches SPCAM’s precipitation distribution, including its tail, and captures realistic tropical waves despite smoother heating and moistening variability.The network has reduced variance, especially near 900 hPa and in the boundary layer, and its instantaneous predictions lack some SPCAM variability.
- Physical consistency: NNCAM approximately conserves column moist static energy without explicit conservation training, while globally integrated energy and moisture remain stable without noticeable drift in multi-year simulations.The stability also holds in a 50-year NNCAM test simulation.
- Generalization: NNCAM trained only on the reference climate fails in much warmer climates, whereas NNCAM-ref+4K performs well across extreme and intermediate climates.NNCAM-ref+4K also captures increased precipitation extremes and amplified, accelerated MJO and Kelvin waves under warming.
Supplement
The supplement details the model setup, neural-network architecture, training procedure, and experimental design. It reports substantial computational speed-ups and identifies deep network complexity as important for stable, realistic prognostic simulations.
- Computational performance: NNCAM’s physical parameterization is 20 times faster than SPCAM and 8 times faster than CTRLCAM, yielding total model speed-ups of 10× and 4×, respectively.Radiation was run every GCM time step in both SPCAM and CTRLCAM to generate training data and enable comparison.
- Experimental design: Experiments used prescribed aquaplanet sea-surface temperatures, including global warming perturbations up to 4 K and a zonally asymmetric wavenumber-one perturbation.All simulations shared initial conditions, spun up for one year, and used the subsequent five years for analysis; training data came from SPCAM’s second year.
- Neural-network design: The neural network uses nine fully connected layers with 256 nodes each and 567,361 learnable parameters.It uses LeakyReLU activations and was trained with Adam for 18 epochs using batches of 1024.
- Neural-network design: Deep neural networks were essential for stable and realistic prognostic simulations, whereas one- and two-hidden-layer networks produced unstable modes and unrealistic artifacts.A four-layer network was the minimum complexity that provided good results for this configuration.