Preface
Data from the World Meteorological Organization (WMO) indicates that over the past 50 years, on average, one disaster related to weather, climate, or water hazards has occurred every day, with each disaster resulting in an average of about 115 deaths and approximately $202 million in economic losses.
More lamentably, in recent years, climate change accelerated by human activities has made extreme weather and climate disasters such as heatwaves, cold waves, heavy precipitation, and droughts occur more frequently.
Therefore, timely and accurate weather forecasting and climate simulation can not only help save tens of thousands of lives each year but also reduce the catastrophic impact of extreme weather and climate events on human society and ecosystems.
Nowadays, the artificial intelligence (AI) model NeuralGCM, developed by the Google Research team and its collaborators, has elevated 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 possesses the world's most advanced traditional physical weather forecasting models.The accuracy of the 10-day advance forecast of NeuralGCM is comparable to, or even better than, that of other existing AI models;
Advertisement
After incorporating sea surface temperatures, the 40-year climate prediction results of NeuralGCM are consistent with the global warming trend found in ECMWF data;
NeuralGCM also surpasses existing climate models in predicting cyclones and their trajectories.
It is worth mentioning that NeuralGCM not only achieves or exceeds the accuracy of existing traditional numerical weather forecasting models and other machine learning (ML) models; it is also "far ahead" in terms of speed, capable of generating 22.8-day atmospheric simulations within a 30-second calculation time; and can save an order of magnitude in computing power compared to traditional models.
The relevant research paper, titled "Neural general circulation models for weather and climate," has been published in the authoritative scientific journal Nature.
These results collectively indicate that NeuralGCM can generate deterministic weather, weather and climate ensemble forecasts, and has shown sufficient 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 represent the physical processes of the atmosphere, oceans, and land, and are the foundation for weather and climate prediction), and can enhance the large-scale physical simulations that are crucial for understanding and predicting the Earth system.
In addition, the hybrid modeling method of NeuralGCM can also be applied to other scientific fields, such as material discovery, protein folding, and multi-physics engineering design.
What about the real effect?
Reducing the uncertainty of long-term forecasts and estimating extreme weather events are key to understanding climate mitigation and adaptation.
ML models have always been considered an alternative means of weather prediction, with the advantage of saving computational power costs, and have even reached or exceeded the level of general circulation models in deterministic weather forecasting, but their performance in long-term forecasts is often not as good as that of general circulation models.In this work, the research team has designed NeuralGCM by combining machine learning and physical methods, replacing or correcting the traditional physical parameterization schemes in GCMs with ML components. It consists of the following key parts:
Differentiable dynamical core: This core is responsible for solving the discretized dynamical equations, simulating large-scale fluid motion and thermodynamic processes, influenced by gravity, the Coriolis force, and other factors. The dynamical core uses horizontal pseudo-spectral discretization and vertical sigma coordinates, and is implemented with the JAX library, supporting automatic differentiation. It simulates seven prognostic variables: horizontal wind vorticity, horizontal wind divergence, temperature, surface pressure, and three types of water substances (specific humidity, ice cloud water content, and liquid cloud water content).
Learning physical module: This module uses the single column approach in the 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 between all atmospheric columns. The inputs to the neural network include the prognostic variables in the atmospheric column, total incident solar radiation, sea ice concentration, sea surface temperature, and horizontal gradients of the prognostic variables. The output of the neural network is the trend of the prognostic variables, scaled by the unconditional standard deviation of the target field.
Encoder and decoder: Since the ERA5 data is stored in pressure coordinates, and the dynamical core uses a sigma coordinate system, an encoder and decoder are needed for transformation. These components perform linear interpolation between pressure levels and sigma coordinate levels, and use the same neural network architecture as the learned physical module for correction. The encoder can eliminate the gravity waves caused by initialization shocks, thereby avoiding contamination of the forecast results.
The results show that NeuralGCM has demonstrated strong capabilities in weather forecasting, comparable to the state-of-the-art models on ultra-short, short, and medium time scales. As follows:Ultra-short-term forecasting (0-1 day)
Generalization ability: Compared to GraphCast, NeuralGCM performs better under untrained weather conditions because it uses local neural networks to predict the physical processes in the atmospheric vertical column.
Short-term forecasting (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: Compared with other machine learning models, the predictions of NeuralGCM are clearer, avoiding physically inconsistent vague predictions.
Interpretability: By diagnosing precipitation minus evaporation, the results of NeuralGCM are more interpretable, facilitating water resource analysis.Geostrophic Wind Balance: Compared to GraphCast, NeuralGCM more accurately simulates the geostrophic wind and the vertical structure of the geostrophic wind and its ratios.
Medium-range Forecast (7-15 days)
Ensemble Forecast: The ensemble average RMSE, RMSB, and CRPS errors of NeuralGCM-ENS at a resolution of 1.4° are all lower than those of ECMWF-ENS, indicating its ability to better capture the average state of possible weather.
Calibratability: The ensemble forecast of NeuralGCM-ENS has a spread-skill ratio of about 1, just like ECMWF-ENS, which is a necessary condition for calibrated forecasts.
In addition to its excellent performance in weather forecasting, NeuralGCM also demonstrates strong capabilities in climate simulation, as shown in the simulation of seasonal cycles, tropical cyclone simulation, and historical temperature trend simulation. For example:
Seasonal Cycle SimulationTranslate the following text into English:
Accuracy: NeuralGCM can 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 circulation and the zonal mean wind.
Comparison with global cloud-resolving models: Compared with the global cloud-resolving model X-SHiELD, NeuralGCM has a smaller bias in precipitable water and has lower temperature bias in the tropics.
Tropical cyclone simulation
Trajectory and number: Even at a coarse resolution of 1.4 degrees, NeuralGCM can produce tropical cyclone trajectories and numbers similar to ERA5, while the global cloud-resolving model X-SHiELD underestimates the number of tropical cyclones at 1.4 degree resolution.
Historical temperature trend simulationAMIP Simulation: The NeuralGCM-2.8° conducted a 40-year AMIP simulation, the results of which indicate that all simulations accurately captured the global warming trend observed in the ERA5 data, and the interannual temperature trends showed a strong correlation with the ERA5 data, demonstrating that the NeuralGCM can effectively simulate the impact of sea surface temperature forcing on the climate.
Comparison with CMIP6 Models: Compared to the CMIP6 AMIP models, the NeuralGCM-2.8° had smaller temperature biases during the period from 1981 to 2014, and this result still holds true even after the global temperature biases of the CMIP6 AMIP models were eliminated.
Despite the strong capabilities demonstrated by NeuralGCM in weather and climate prediction, it still has some limitations.
Firstly, the NeuralGCM's ability to predict future climates is limited. The NeuralGCM currently cannot predict future climates that are significantly different from the historical climate. When the sea surface temperature (SST) increases significantly (for example, by +4K), the response of the NeuralGCM is inconsistent with expectations and exhibits climate drift phenomena.
Secondly, the NeuralGCM's ability to simulate unobserved climates is insufficient. Like other machine learning climate models, the NeuralGCM also faces the challenge of simulating unobserved climates, such as future climates or climates that differ significantly 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, there are also issues with physical constraints and numerical stability in the NeuralGCM. For example, the spectral distribution of the NeuralGCM is still more blurred than that of the ECMWF physical forecast, and there is an underestimation in the simulation of tropical extreme events. This requires further research and improvement of the model's physical process parameterization and numerical methods to enhance the model's physical consistency and numerical stability.Finally, there is a lack of coupling with other components of the Earth system. Currently, NeuralGCM only simulates the atmospheric system, while the climate system is a complex interactive system that includes the ocean, land, ice and snow, and the biosphere, among others. 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-physical field coupling simulation.
Traditional weather forecasting and climate simulation are being disrupted by AI.
In the field of weather forecasting and climate simulation, NeuralGCM is not a "pioneer."
In the past few years, technology companies and universities such as Huawei, Google, and Tsinghua University have made significant progress in this direction.In July 2023, the Pangu-Weather model, developed by Huawei Cloud, was published in Nature. It used 39 years of global reanalysis weather data as training data, and its prediction accuracy is comparable to the world's best numerical weather prediction system, IFS, and it 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 introduced NowcastNet, which was led by the research team of Michael Jordan, a giant in the field of machine learning and a professor at the University of California, Berkeley, and Wang Jianmin, a professor at Tsinghua University. This model can combine physical laws and deep learning to make real-time precipitation forecasts.
In November 2023, Google DeepMind launched a machine learning-based weather prediction model called GraphCast. At a global resolution of 0.25 degrees, this model can predict hundreds of weather variables for the next 10 days within a minute, which is significantly better than traditional meteorological forecasting methods, and it also performs well in predicting extreme events. The relevant research paper has been published in the authoritative scientific journal Science.
In March of this year, an AI model developed by the Google Research team defeated the most advanced global flood warning system. It used training with the existing 5680 measuring instruments and can predict daily runoff in unmeasured watersheds within a 7-day forecast period.
Nowadays, traditional weather forecasting and climate simulation are being disrupted by AI. In the future, AI will further accelerate the speed and accuracy of meteorological forecasting, benefiting all of humanity.