news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Written by Ma Xuewei

Editor | Cooper

Preface

Data from the World Meteorological Organization (WMO) show that over the past 50 years, an average of one weather, climate or flood-related disaster has occurred every day.On average, each disaster causes approximately 115 deaths and approximately $202 million in economic losses.。

What is even more regrettable is that in recent years, climate change accelerated by human activities has made extreme weather and climate disasters such as heat waves, cold waves, heavy rainfall, and droughts occur more frequently.

therefore,Timely and accurateImproved weather forecasting and climate simulations can not only help save tens of thousands of lives every year, but also reduce the catastrophic impacts of extreme weather and climate events on human society and ecosystems.

now,NeuralGCM, an artificial intelligence (AI) model developed by Google Research and its collaborators, takes weather forecasting and climate simulation to a new level——

• The accuracy of NeuralGCM for 1-15 day forecasts is comparable to that of the European Centre for Medium-Range Weather Forecasts (ECMWF), which has the most advanced traditional physical weather forecast model in the world;

• For the 10-day forecast accuracy, NeuralGCM performs as well as or better than other existing AI models;

• After adding sea surface temperature, the 40-year climate projections from NeuralGCM are consistent with the global warming trend found from ECMWF data;

• NeuralGCM also outperforms existing climate models in predicting cyclones and their tracks.

It is worth mentioning that NeuralGCM not only reaches or even exceeds the accuracy of existing traditional numerical weather forecast models and other machine learning (ML) models; it is also "far ahead" in speed and can be used inGenerate 22.8 days of atmospheric simulation in 30 seconds of computing time; and can save orders of magnitude of computational effort compared to traditional models.

Video | NeuralGCM simulates the atmosphere faster than the most advanced physical models, while generating forecasts of comparable accuracy. This video compares the number of simulated days of the atmosphere generated by NeuralGCM with two physical models, NOAA X-SHiELD and NCAR CAM6, within 30 seconds of computing time. Among them, NOAA X-SHiELD is a high-resolution (0.03°) physical model that must be run on a supercomputer; NCAR CAM6 is a pure atmospheric physics model with a lower resolution (1.0°). Due to its lower computational cost, it is a more commonly used choice for scientists. Although NeuralGCM runs at a lower resolution (1.4°), its accuracy is comparable to higher-resolution models. (Source: Google Research)

The related research paper, titled “Neural general circulation models for weather and climate”, has been published in the authoritative scientific journal Nature.

Together, these results indicate that NeuralGCM can generate deterministic synoptic, weather, and climate ensemble forecasts, showing adequate stability in long-term weather and climate simulations.

The research team believes that this end-to-end deep learning is compatible with the tasks performed by traditional general circulation models (GCMs, which characterize the physical processes of the atmosphere, ocean, and land and are the basis for weather and climate predictions) and can enhance large-scale physical simulations that are critical to understanding and predicting the Earth system.

In addition, NeuralGCM's hybrid modeling approach can also be applied to other scientific fields, such as materials discovery, protein folding, and multi-physics engineering design.

What is the real effect?？

Reducing uncertainty in long-term forecasts and estimating extreme weather events is key to understanding climate mitigation and adaptation.

ML ModelsIt has always been considered as an alternative means of weather forecasting, with the advantage of saving computing costs, and has even reached or exceeded the level of atmospheric circulation models in deterministic weather forecasting.Often underperforms GCM in long-term forecasts。

In this work, the research team combined machine learning and physical methods to design NeuralGCM, which uses ML components to replace or correct traditional physical parameterization schemes in GCM. It consists of the following key parts:

• Differentiable dynamic core: This core is responsible for solving the discretized dynamic equations, simulating large-scale fluid motion and thermodynamic processes, affected by gravity, Coriolis forces and other factors. The dynamic core uses horizontal pseudospectral discretization and vertical sigma coordinates, and is implemented using the JAX library with support for automatic differentiation. It simulates seven forecast variables: horizontal wind vorticity, horizontal wind divergence, temperature, surface pressure and three water species (specific humidity, ice cloud water content and liquid cloud water content).

• Study Physics Module: This module uses the single column approach in GCM, using only the information of a single atmospheric column to predict the impact of unresolved processes within that column. It uses a fully connected neural network with residual connections and shares weights across all atmospheric columns. The inputs to the neural network include the forecast variables in the atmospheric column, total incoming solar radiation, sea ice concentration, and sea surface temperature, as well as the horizontal gradients of the forecast variables. The output of the neural network is the forecast variable trend, scaled by the unconditional standard deviation of the target field.

• Encoders and Decoders: Since ERA5 data is stored in pressure coordinates, while the power core uses the sigma coordinate system, encoders and decoders are needed to convert. These components perform linear interpolation between pressure levels and sigma coordinate levels and use the same neural network architecture as the learned physics module for correction. The encoder can remove gravity waves caused by the initialization shock, thus avoiding contamination of the prediction results.

Figure ｜ NeuralGCM model architecture. NeuralGCM combines traditional fluid dynamics solvers with neural networks for small-scale physics. These components are composed of differential equation solvers that advance the system sequentially in time. (Source: Google Research)

The results show that NeuralGCM demonstrates strong capabilities in weather forecasting, comparable to the most advanced models on ultra-short-term, short-term, and medium-term time scales.

Ultra-short-term forecast (0-1 days)

• Generalization: Compared to GraphCast, NeuralGCM performs better on untrained weather conditions because it uses a local neural network to predict physical processes in the vertical column of the atmosphere.

Short-term forecast (1-10 days)

• Accuracy: In the short-term forecast of 1-3 days, NeuralGCM-0.7° and GraphCast perform best, accurately tracking the changes in weather patterns.

• Physical consistency: NeuralGCM’s predictions are clearer than other machine learning models, avoiding ambiguous predictions that are physically inconsistent.

• Interpretability: By diagnosing precipitation minus evaporation, NeuralGCM results are more interpretable and convenient for water resources analysis.

• Geostrophic wind balance: NeuralGCM simulates the vertical structure of the geostrophic and geostrophic winds and their ratios more accurately than GraphCast.

Medium-term forecast (7-15 days)

• Ensemble forecast: The ensemble mean RMSE, RMSB and CRPS errors of NeuralGCM-ENS at 1.4° resolution are lower than those of ECMWF-ENS, indicating that it can better capture the possible mean weather state.

• Calibration: The ensemble forecasts of NeuralGCM-ENS, like those of ECMWF-ENS, have a divergence-skill ratio of about 1, which is a necessary condition for calibrating the forecasts.

In addition to its outstanding performance in weather forecasting, NeuralGCM also demonstrates strong capabilities in climate simulation, including seasonal cycle simulation, tropical cyclone simulation, and historical temperature trend simulation. As follows:

Seasonal cycle simulation

• Accuracy: NeuralGCM is able to accurately simulate the seasonal cycle, including the annual cycle of global precipitable water and global total kinetic energy, as well as key atmospheric dynamics such as the Hadley cell and meridional mean winds.

• Comparison with global cloud-resolving models: NeuralGCM has smaller biases in precipitable water and lower temperature biases in the tropics compared to the global cloud-resolving model X-SHiELD.

Tropical cyclone simulation

• Tracks and numbers: Even at a coarse resolution of 1.4°, NeuralGCM produces similar tropical cyclone tracks and numbers to ERA5, while the global cloud-resolving model X-SHiELD underestimates the number of tropical cyclones at 1.4° resolution.

Video | NeuralGCM predicts the path of global tropical cyclones in 2020. The forecast results are consistent with the number and intensity of actual cyclones that year shown in the ECMWF Reanalysis Version 5 (ERA5) dataset. (Source: Google Research)

Historical temperature trend simulation

• AMIP simulations: NeuralGCM-2.8° performed 40 years of AMIP simulations and the results showed that all simulations accurately captured the global warming trend observed in the ERA5 data and that the interannual temperature trends were strongly correlated with the ERA5 data, indicating that NeuralGCM can effectively simulate the impact of SST forcing on climate.

• Comparison with CMIP6 models: NeuralGCM-2.8° has smaller temperature biases over the 1981-2014 period compared to the CMIP6 AMIP model, and this result holds even after removing the global temperature bias of the CMIP6 AMIP model.

Figure | NeuralGCM's accuracy and ability to capture global warming on a ten-year time scale. Comparison of the performance of NeuralGCM and AMIP in predicting global average temperatures from 1980 to 2020. (Source: Google Research)

Although NeuralGCM has shown great capabilities in weather and climate prediction, itThere are still some limitations。

first,NeuralGCMs have limited ability to predict future climateNeuralGCM is currently unable to predict future climates that differ significantly from historical climates. When the sea surface temperature (SST) increases significantly (e.g. +4K), the response of NeuralGCM is inconsistent with expectations and climate drift occurs.

Secondly,NeuralGCMs have insufficient ability to simulate unobserved climateSimilar to other machine learning climate models, NeuralGCM also faces the challenge of simulating unobserved climates, such as future climates or climates that differ greatly from historical data. This requires the model to have stronger generalization capabilities and more advanced training strategies, such as adversarial training or meta-learning.

Then,NeuralGCM also has physical constraints and numerical stability issuesFor example, the spectral distribution of NeuralGCM is still more ambiguous than that of ECMWF physical forecasts, and there is an underestimation in simulating tropical extreme events. This requires further research and improvement of the model's physical process parameterization and numerical methods to improve the model's physical consistency and numerical stability.

at last,Lack of coupling with other Earth system componentsCurrently, NeuralGCM only simulates the atmospheric system, while the climate system is a complex interacting system that includes the ocean, land, ice and snow, and the biosphere. To perform a more comprehensive climate simulation, NeuralGCM needs to be coupled with these components and consider the interactions between them. This requires the development of new model architectures and training strategies to achieve multi-physics coupled simulations.

Traditional weather forecasting and climate simulation are being subverted by AI

NeuralGCM is not a pioneer in weather prediction and climate simulation。

In the past few years, technology companies and universities including Huawei, Google and Tsinghua University have made significant progress in this direction.

In July 2023, Huawei Cloud developedPangu-Weather ModelIt was published in Nature. It used 39 years of global reanalysis weather data as training data. Its prediction accuracy is comparable to that of the world's best numerical weather forecast system IFS, and is more than 10,000 times faster than the IFS system at the same spatial resolution.

Another paper published in Nature at the same time introducedNowcastNet, which comes from a research team led by Michael Jordan, a leading figure in the field of machine learning and professor at the University of California, Berkeley, and Wang Jianmin, a professor at Tsinghua University. The model can combine physical laws and deep learning to make real-time precipitation forecasts.

In November 2023, Google DeepMind launched a weather prediction model based on machine learning——GraphCast, at a global resolution of 0.25°, the model can predict hundreds of weather variables for the next 10 days within one minute, significantly outperforming traditional weather forecasting methods and performing well in predicting extreme events. The relevant research paper has been published in the authoritative scientific journal Science.

In March this year, an AI model also developed by the Google Research team defeated the most advanced global flood warning system. It was trained using 5,680 existing gauges and can predict daily runoff in ungauged basins within a 7-day forecast period.

Today, traditional weather forecasting and climate simulation are being disrupted by AI. In the future, AI will further accelerate the speed and accuracy of weather forecasting, benefiting all mankind.

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

https://www.science.org/doi/10.1126/science.adi2336

https://www.nature.com/articles/s41586-024-07145-1

https://www.nature.com/articles/s41586-023-06185-3

https://www.nature.com/articles/s41586-023-06184-4

｜Click to follow and remember to mark the star｜