Neural general circulation models

Climate Computer Summit

University of Illinois Urbana-Champaign

30 September 2024

Stephan Hoyer

with many others from Google, ECMWF & MIT

The AI revolution has arrived for weather forecasting

The AI revolution has arrived for weather forecasting

The AI revolution has arrived for weather forecasting

The AI revolution has arrived for weather forecasting

AI weather forecasts are skillful, but not yet fully physically realistic

Hybrid modeling may offer the best of both worlds

Physics-based

Pure ML

Traditional NWP

Climate models

GraphCast

Pangu-Weather

Hybrid models

NeuralGCM

Hybrid modeling may offer the best of both worlds

Physics-based

Pure ML

Traditional NWP

Climate models

GraphCast

Pangu-Weather

Very little code�Based on data

Optimized for forecast accuracy

Hybrid models

NeuralGCM

Hybrid modeling may offer the best of both worlds

Physics-based

Pure ML

Traditional NWP

Climate models

GraphCast

Pangu-Weather

Hybrid models

NeuralGCM

Very little code�Based on data

Optimized for forecast accuracy

Complex, but interpretable�Based on physics

Designed to generalize

NeuralGCM is the first differentiable hybrid model for the Earth’s atmosphere

NeuralGCM combines a spectral dynamical core (written in Python/JAX) with neural network learned physics.

​

Models are trained on 3-5 day weather forecasts of the ERA5 reanalysis.

NeuralGCM achieves state-of-the-art results both for weather forecasting and climate simulation

1. Competitive 1-15 day ensemble weather forecasts with ECMWF

NeuralGCM achieves state-of-the-art results both for weather forecasting and climate simulation

1. Competitive 1-15 day ensemble weather forecasts with ECMWF

2. Realistic decadal simulations, competitive with atmosphere only (AMIP) climate models

NeuralGCM near-term climate forecasts have less bias than a global storm resolving model

Precipitable water bias for 2020 [mm]

140 km Neural GCM�RMSE = 1.09 mm

3 km GFDL X-SHiELD�RMSE = 1.74 mm

NeuralGCM near-term climate forecasts have less bias than a global storm resolving model

Precipitable water bias for 2020 [mm]

140 km Neural GCM�RMSE = 1.09 mm

3 km GFDL X-SHiELD�RMSE = 1.74 mm

200 sim years / day

1 Google TPU core

0.05 sim years / day

13,824 CPU cores

NeuralGCM near-term climate forecasts have less bias than a global storm resolving model

Precipitable water bias for 2020 [mm]

140 km Neural GCM�RMSE = 1.09 mm

3 km GFDL X-SHiELD�RMSE = 1.74 mm

200 sim years / day

1 Google TPU core�~$0.08 / simulated year

0.05 sim years / day

13,824 CPU cores�~$80,000 / simulated year

Differentiability makes it possible to optimize GCMs directly to match observations – such as precipitation

IMERG (ground truth)

Differentiability makes it possible to optimize GCMs directly to match observations – such as precipitation

​

IMERG (ground truth)

NeuralGCM

X-SHiELD

ERA5

GFDL

What lessons does NeuralGCM offer for climate simulation?

1. Do not underestimate the power of benchmarks (see: WeatherBench).�If we can measure what “good” means, AI will be competitive.

​

What lessons does NeuralGCM offer for climate simulation?

1. Do not underestimate the power of benchmarks (see: WeatherBench).�If we can measure what “good” means, AI will be competitive.
2. AI is an accelerator, not a replacement.
1. For data-rich phenomena: AI enables speed and accuracy
2. For data-poor phenomena: we can still use physical models

​

What lessons does NeuralGCM offer for climate simulation?

1. Do not underestimate the power of benchmarks (see: WeatherBench).�If we can measure what “good” means, AI will be competitive.
2. AI is an accelerator, not a replacement.
1. For data-rich phenomena: AI enables speed and accuracy
2. For data-poor phenomena: we can still use physics
1. Re-writing a GCM in Python is feasible, with many benefits beyond AI (differentiability, GPU/TPU support, parallelization, modern code…).