Google has introduced NeuralGCM, a hybrid weather forecasting model that mixes machine learning with traditional forecasting techniques and offers surprising benefits.
The accuracy of weather forecasts has improved dramatically, but traditional techniques require enormous computing resources to run increasingly complex algorithms.
General circulation models (GCMs) form the premise of climate and weather forecasts that inform you whether you'll need an umbrella tomorrow.
GCMs are physics-based simulators that use mathematical equations based on the laws of physics to simulate how air, water and energy move on the planet.
Typical GCMs divide the Earth's surface right into a grid of as much as 100 kilometers, just like an enormous chessboard. The algorithm processes each square incrementally to predict how atmospheric conditions are more likely to change.
The equations behind GCMs are incredibly complex and employ among the largest supercomputers on this planet.
Machine learning (ML) models for weather forecasting have shown significant potential, but are primarily data-based.
While an ML weather prediction model has an excellent idea of ​​historical weather data, it lacks the elemental understanding of the physical laws of the atmosphere which can be modeled in a GCM.
ML models are fast and may provide accurate short-term forecasts, but often struggle with long-term stability and rare extreme weather events or future climate scenarios.
NeuralGCM, developed by a team at Google Research, combines the accuracy and long-term predictive capabilities of traditional GCMs with the improved resolution, efficiency and speed of ML models.
NeuralGCM is offered totally free and we’re excited to see how scientists construct on it.
For more details, see my blog post describing the work and our open source code:https://t.co/AfqLagoYDshttps://t.co/MSlFg6pgfj
— Stephan Hoyer (@shoyer) July 22, 2024
The article states that NeuralGCM's accuracy is comparable to or higher than that of current GCM models. NeuralGCM is “the primary machine learning-based model to supply accurate ensemble weather forecasts, with a greater CRPS than the newest physics-based models.”
CRPS is a price that compares the anticipated weather with the actual weather.
The researchers claim: “NeuralGCM is competitive with machine learning models for one- to ten-day forecasts and with the European Centre for Medium-Range Weather Forecasts ensemble forecast for one- to fifteen-day forecasts.”
While NeuralGCM achieves comparable prediction results to GCMs, it’s orders of magnitude less computationally intensive and far less complex.
The paper doesn’t specify how big NeuralGCM is, but offers Google's ML weather forecast model GraphCast for comparison.
GraphCast consists of about 5,417 lines, while the National Oceanic and Atmospheric Administration (NOAA) atmospheric model FV3 has about 376,578 lines of code.
The researchers say NeuralGCM enables “computational resource savings of three to 5 orders of magnitude.”
To put this in context, the paper explains that “NeuralGCM-1.4° simulates 70,000 simulation days in 24 hours using a single tensor processing unit, in comparison with 19 simulated days on 13,824 central processing cores using X-SHiELD,” a high-resolution weather forecast model.
The researchers say their results show that their model has impressive climate modeling capabilities. The paper states: “NeuralGCM models trained with 72-hour forecasts are able to performing realistic simulations over multiple years.”
Combining machine learning with traditional physics models, resembling those utilized by Google in weather forecasting, “has the potential to rework simulation for a wide selection of applications, resembling materials science, protein folding, and multiphysics engineering design.”
Data centers have come under heavy criticism resulting from resource-hungry AI and its potential impact on the climate.
NeuralGCM is an excellent example of how AI can have a positive impact on the environment by replacing or augmenting inefficient traditional processes to cut back computing power requirements.
