news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Zhidongxi (public account:zhidxcom）
Compiled by Meng Qiang
Edited by Yunpeng

According to Zhidongxi on July 24, Google published a paper in Nature on July 23, introducing the NeuralGCM atmospheric model it developed in cooperation with the European Centre for Medium-Range Weather Forecasts (ECMWF). The model combines traditional physics-based modeling with machine learning to improve the accuracy and efficiency of the model in predicting weather and climate.

The paper shows that NeuralGCM's forecast accuracy for 1 to 15 days is comparable to that of ECMWF, which has the world's most advanced traditional physical weather forecast model; after adding sea surface temperature, NeuralGCM's 40-year climate forecast results are consistent with the global warming trend obtained from ECMWF data; NeuralGCM also surpasses existing climate models in predicting cyclones and their trajectories.

It is worth mentioning that NeuralGCM is also "far ahead" in speed, and can generate 22.8 days of atmospheric simulation in 30 seconds of computing time, and the computing cost is 100,000 times lower than that of traditional GCM. As the first climate model based on machine learning, NeuralGCM has brought weather prediction and climate simulation to a new level in terms of both prediction accuracy and efficiency.

Paper address: https://www.nature.com/articles/s41586-024-07744-y

1. Machine learning drives climate model transformation

The earth is warming at an unprecedented rate. Extreme weather has occurred frequently in recent years. The World Meteorological Organization said that 2023 will be the hottest year in history, and 2024 may be even hotter. Against the backdrop of frequent extreme weather, the importance of climate prediction is particularly prominent.

General circulation models (GCMs) are the foundation of weather and climate forecasts. They are traditional models used to simulate and predict the Earth's atmosphere and climate system based on physics. By simulating the physical processes of the Earth's atmosphere, oceans, land, and ice sheets, GCMs can provide long-term weather and climate forecasts. Although traditional climate models have been continuously improved over the decades, these models often produce errors and deviations due to scientists' incomplete understanding of how the Earth's climate works and how models are built.

Stephan Hoyer, a senior engineer at Google, said that traditional GCMs divide the Earth into cubes extending from the surface to the atmosphere, with sides usually 50-100 kilometers long, and predict weather changes in each cube over a period of time based on this. They make predictions based on the laws of physics by calculating the dynamic changes of air and moisture. However, many important climate processes, such as cloud formation and rainfall, vary on a scale ranging from millimeters to kilometers, which is much smaller than the size of the cubes currently used by GCMs, so they cannot be accurately calculated based on physics.

In addition, scientists lack a complete physical understanding of some processes, such as how clouds form. Therefore, these traditional models do not fully rely on physical principles, but use simplified models to generate approximations and parameterize weather dynamics, but this approach reduces the accuracy of GCMs.

Like traditional models, NeuralGCM divides the Earth's atmosphere into cubes and calculates the physical properties of large-scale changes such as air movement and moisture movement. However, for small-scale weather dynamics such as cloud formation, NeuralGCM does not use traditional parameterization, but uses neural networks to learn the physical properties of these weather dynamics from existing weather data.

Hoyer revealed that a key innovation of NeuralGCM is that Google wrote a numerical solver for large-scale changes in JAX from scratch, which enables the model to be adjusted "online" using gradient-based optimization. Another benefit of writing the entire model in JAX is that the model can run efficiently on TPUs and GPUs, while traditional climate models mostly run on CPUs.

2. Prediction accuracy is better than the current most advanced model

The paper shows that NeuralGCM's deterministic model (which outputs a single, deterministic prediction result) performs comparable to the current state-of-the-art models at a resolution of 0.7°, with a weather forecast accuracy of up to 5 days.

Since deterministic models only provide one prediction result, they may not fully represent the diversity of the future state of the climate system. Therefore, ensemble forecasts are introduced in climate prediction, which is to generate a series of possible weather scenarios based on a slightly different set of initial conditions. After integration, ensemble forecasts will produce probabilistic weather forecasts, which are usually more accurate than deterministic forecasts. The paper states that NeuralGCM's 1.4° resolution ensemble forecast model outperforms the current most advanced models in terms of forecast accuracy for 5 to 15 days.

In addition, NeuralGCM's climate prediction accuracy for longer time spans is also higher than that of the most advanced models. When predicting the temperature over the 40 years from 1980 to 2020, the average error of NeuralGCM's 2.8° deterministic model is only 0.25 degrees Celsius, which is one-third of the error of the Atmospheric Model Intercomparison Project (AMIP).

3. Complete a year of atmospheric dynamics simulation in 8 minutes

Hoyer said that NeuralGCM's calculation speed is several orders of magnitude faster than traditional GCM, and the calculation cost is also lower. NeuralGCM's 1.4° model is more than 3,500 times faster than the high-precision climate model X-SHiELD. In other words, it takes researchers 20 days to simulate a year's atmospheric dynamics using X-SHiELD, while it only takes 8 minutes using NeuralGCM.

In addition, researchers need to request access to a supercomputer with 13,000 CPUs to run X-SHiELD, while NeuralGCM only requires a computer with a single TPU. Hoyer said that the computational cost of using NeuralGCM for climate simulation is one hundred thousandth of that of using X-SHiELD.

Conclusion: Towards a more open, fast and efficient climate prediction model

The Google Research team has made the source code and model weights of NeuralGCM public on GitHub for non-commercial use. Hoyer said that Google hopes that researchers around the world can actively participate in testing and improving the model. NeuralGCM can be run on a laptop, so it is hoped that more climate researchers will use the model in their work.

Currently, NeuralGCM only simulates the Earth's atmosphere, and Google hopes to include other climate systems such as the ocean and carbon cycle in the model in the future. Although NeuralGCM is not a complete climate model yet, its emergence provides new ideas for climate prediction. In the future, we expect to see AI further improve the accuracy and speed of climate prediction.

Source: Google, Nature