AI Surpasses Traditional Models: Rapid and Accurate Prediction of Weather and Climate Change

Artificial intelligence is fundamentally changing the landscape of weather forecasting and climate simulation. This cutting-edge technology is bringing revolutionary changes to traditional methods of weather prediction and climate research. The application of AI not only improves the accuracy and efficiency of forecasts but also provides new perspectives and tools for climate change research. This technological advancement is reshaping the way we understand and predict atmospheric phenomena, opening up new development prospects for meteorology and climate science.

Data from the World Meteorological Organization (WMO) shows that over the past 50 years, on average, a weather, climate or water-related disaster has occurred every day, causing approximately 115 deaths and $202 million in economic losses per event.

More alarmingly, in recent years, climate change accelerated by human activities has led to an abnormal increase in extreme weather and climate disasters such as heat waves, cold spells, heavy precipitation, and droughts.

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.

Now, an artificial intelligence (AI) model called NeuralGCM, developed by the Google Research team and its collaborators, has taken weather forecasting and climate simulation to a new level:

  • NeuralGCM's accuracy for 1-15 day forecasts is comparable to that of the European Centre for Medium-Range Weather Forecasts (ECMWF), which has the world's most advanced traditional physical weather forecasting model.
  • For 10-day advance forecasts, NeuralGCM performs as well as or better than existing AI models.
  • With the inclusion of sea surface temperature, NeuralGCM's 40-year climate prediction results are consistent with global warming trends found in ECMWF data.
  • NeuralGCM also outperforms existing climate models in predicting cyclones and their trajectories.

Notably, NeuralGCM not only matches or exceeds the accuracy of existing traditional numerical weather prediction models and other machine learning (ML) models, but it is also significantly faster, capable of generating 22.8 days of atmospheric simulation in 30 seconds of computation time. It can also save orders of magnitude in computational resources compared to traditional models.

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

These results collectively demonstrate that NeuralGCM can generate deterministic weather, weather and climate ensemble forecasts, showing sufficient stability for long-term weather and climate simulations.

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

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

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

ML models have long been considered an alternative means of weather prediction, with the advantage of saving computational costs. They have even reached or exceeded the level of atmospheric circulation models in deterministic weather forecasting. However, they often underperform atmospheric circulation models in long-term forecasting.

In this work, the research team designed NeuralGCM by combining machine learning and physical methods, using ML components to replace or correct traditional physical parameterization schemes in GCMs. It consists of the following key parts:

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

  2. Learning physics module: This module uses the single-column method in GCMs, using only information from a single atmospheric column to predict the influence of unresolved processes within that column. It uses a fully connected neural network with residual connections, sharing weights across all atmospheric columns. The neural network's inputs include forecast variables in the atmospheric column, total incident solar radiation, sea ice concentration and sea surface temperature, as well as horizontal gradients of forecast variables. The neural network's output is the forecast variable trend, scaled by the unconditional standard deviation of the target field.

  3. Encoder and decoder: Since ERA5 data is stored in pressure coordinates while the dynamic core uses a sigma coordinate system, encoders and decoders are needed for conversion. 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 eliminate gravity waves caused by initialization shock, thereby avoiding contamination of prediction results.

Results show that NeuralGCM demonstrates powerful capabilities in weather prediction, comparable to state-of-the-art models at ultra-short-term, short-term, and medium-term time scales. For example:

Ultra-short-term prediction (0-1 day):

  • Generalization ability: Compared to GraphCast, NeuralGCM performs better under untrained weather conditions because it uses local neural networks to predict physical processes in atmospheric vertical columns.

Short-term prediction (1-10 days):

  • Accuracy: In short-term predictions of 1-3 days, NeuralGCM-0.7° and GraphCast perform best, accurately tracking changes in weather patterns.
  • Physical consistency: Compared to other machine learning models, NeuralGCM's predictions are clearer, avoiding physically inconsistent blurry predictions.
  • Interpretability: By diagnosing precipitation minus evaporation, NeuralGCM's results are more interpretable, facilitating water resource analysis.
  • Geostrophic wind balance: Compared to GraphCast, NeuralGCM more accurately simulates geostrophic winds and their vertical structure and ratios.

Medium-term prediction (7-15 days):

  • Ensemble forecasting: NeuralGCM-ENS at 1.4° resolution has lower ensemble average RMSE, RMSB, and CRPS errors than ECMWF-ENS, indicating its ability to better capture possible average weather states.
  • Calibratability: NeuralGCM-ENS's ensemble forecasts, like ECMWF-ENS, have a dispersion-skill ratio of about 1, which is a necessary condition for calibrated forecasts.

In addition to excellent performance in weather prediction, NeuralGCM also demonstrates strong capabilities in climate simulation, including seasonal cycle simulation, tropical cyclone simulation, and historical temperature trend simulation. For example:

Seasonal cycle simulation:

  • Accuracy: NeuralGCM can accurately simulate seasonal cycles, including annual cycles of global precipitable water and global total kinetic energy, as well as key atmospheric dynamics such as the Hadley circulation and zonal mean winds.
  • Comparison with global cloud-resolving models: Compared to the global cloud-resolving model X-SHiELD, NeuralGCM has smaller biases in precipitable water and lower temperature biases in tropical regions.

Tropical cyclone simulation:

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

Historical temperature trend simulation:

  • AMIP simulation: NeuralGCM-2.8° conducted a 40-year AMIP simulation. Results show that all simulations accurately capture the global warming trend observed in ERA5 data, and the interannual temperature trends have a strong correlation with ERA5 data, indicating that NeuralGCM can effectively simulate the impact of sea temperature forcing on climate.
  • Comparison with CMIP6 models: Compared to CMIP6 AMIP models, NeuralGCM-2.8° has smaller temperature biases during 1981-2014, even after removing the global temperature bias of CMIP6 AMIP models.

Although NeuralGCM demonstrates powerful capabilities in weather and climate prediction, it still has some limitations:

  1. Limited ability to predict future climate: NeuralGCM currently cannot predict future climates that are significantly different from historical climates. When sea surface temperature (SST) increases significantly (e.g., +4K), NeuralGCM's response is inconsistent with expectations and climate drift occurs.

  2. Insufficient ability to simulate unobserved climates: Like other machine learning climate models, NeuralGCM also faces challenges in simulating unobserved climates, such as future climates or climates that differ significantly from historical data. This requires models to have stronger generalization capabilities and more advanced training strategies, such as adversarial training or meta-learning.

  3. Physical constraints and numerical stability issues: For example, NeuralGCM's spectral distribution is still blurrier than ECMWF physical forecasts, and there are still some numerical stability issues in simulating extreme weather events.