5,000 storms, Google trained an AI predictor, weather forecast ChatGPT moment?

AI Can Also Predict Typhoons! In terms of performance, Google has for the first time launched an AI typhoon prediction model that clearly surpasses mainstream physical models. This is expected to save tens of thousands of lives.

Yesterday, Google DeepMind and the Google Research team officially launched the interactive meteorological platform Weather Lab to share AI weather models.

In tropical cyclone path prediction, Google's new model has refreshed the state-of-the-art (SOTA) and is the first AI prediction model that clearly surpasses mainstream physical models in performance.

Paper link: https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/how-we-re-supporting-better-tropical-cyclone-prediction-with-ai/skillful-joint-probabilistic-weather-forecasting-from-marginals.pdf

Blog link: https://deepmind.google/discover/blog/weather-lab-cyclone-predictions-with-ai/

Project address: https://deepmind.google.com/science/weatherlab

Tropical cyclones are extremely dangerous, leaving only ruins in their wake.

Hurricane Milton captured by NOAA GOES-16 satellite on October 8, 2024

According to the World Meteorological Organization, tropical cyclones have caused $1.4 trillion in economic losses over the past 50 years, resulting in 1,945 disasters and claiming approximately 800,000 lives. Moreover, the number of people threatened is still rising.

These massive and rotating storms, also known as hurricanes or typhoons, typically form in warm ocean waters, driven by heat, humidity, and convection.

On warm oceans, when water vapor condenses, the release of energy initiates a positive feedback loop, allowing tropical cyclones to form

They are extremely sensitive to small atmospheric changes, making precise prediction of their trajectory and intensity a long-recognized challenge in meteorology. However, improving cyclone prediction accuracy can help protect affected communities through more effective disaster preparedness and timely evacuation.

Paper link: https://uhero.hawaii.edu/wp-content/uploads/2023/09/hurricane_forecasts-7.pdf

15 Days in Advance, 50 Weather Simulations

The Weather Lab platform is equipped with the latest AI tropical cyclone model based on stochastic neural networks, capable of predicting cyclone generation, movement paths, intensity changes, scale, and morphological characteristics—

Able to generate up to 50 possible scenario simulations up to 15 days in advance.

The following animation shows the AI model prediction results.

When hurricanes "Hong De" and "Galong" were active in the waters south of Madagascar, the model (blue trajectory) accurately predicted their movement paths.

The model also successfully captured the future trajectories of cyclones "Jude" and "Yvonne" in the Indian Ocean—

Robustly predicting the storm area that would eventually strengthen into a tropical cyclone nearly seven days in advance.

Real-time and Historical Cyclone Predictions

Weather Lab demonstrates real-time and historical cyclone predictions from different AI weather models, including physical model predictions from the European Centre for Medium-Range Weather Forecasts (ECMWF).

Currently, they have multiple AI weather models running in real-time: WeatherNext Graph, WeatherNext Gen, and the latest experimental cyclone model. Weather Lab also provides over two years of historical prediction data for experts and researchers to download and analyze, allowing external evaluation of the model's performance across all ocean basins.

The following animation shows the new model's prediction of the Alfred cyclone becoming a Category 3 cyclone in the Coral Sea.

The ensemble mean prediction (thick blue line) accurately foresaw Alfred cyclone rapidly weakening to a tropical storm and making landfall near Brisbane, Australia seven days later, while also accurately predicting the high-probability landing area on the Queensland coast.

Weather Lab users can explore and compare predictions from various AI and physical models. When these predictions are comprehensively considered, they can help meteorological agencies and emergency service experts better predict cyclone paths and intensities, better respond to different scenarios, disseminate risk information, and support decision-making in managing cyclone impacts.

When using this tool, please keep the following reminder in mind, especially when making decisions based on Weather Lab predictions:

Weather Lab is a research tool. The real-time predictions shown are generated by models still under development and are not official warnings.

For official weather forecasts and warnings, please consult local meteorological agencies or national meteorological services.

AI-Driven Cyclone Prediction

In physics-based cyclone prediction, meeting operational requirements through approximate processing is difficult for a single model to accurately predict both cyclone path and intensity.

This is because cyclone paths are guided by large-scale atmospheric flow, while cyclone intensity depends on complex turbulent processes within and outside its compact core.

Global low-resolution models perform best in predicting cyclone paths but cannot capture the fine processes controlling cyclone intensity, thus requiring assistance from regional high-resolution models.

This time, the Google team invented FGN, a new weather probabilistic modeling method that combines architecture, training, and inference methods, with faster speed and higher flexibility.

FGN models cognitive uncertainty and random uncertainty through different mechanisms (see Figure 1): the former is achieved through model ensemble, while the latter uses techniques related to random functions.

In tropical cyclone path prediction, FGN's average path prediction and path probability prediction significantly outperform existing models (𝑝 < 0.05), becoming the first AI prediction model that clearly surpasses mainstream physical models in performance.

Figure 1 | FGN generation process overview: workflow of generating a single-step prediction ensemble from an input frame pair (𝑋ₜ₋₂:ₜ₋₁)

At two levels, FGN introduces diversity, modeling randomness uncertainty (aleatoric uncertainty) and cognitive uncertainty (epistemic uncertainty) separately.

For a specific model M𝑗, random uncertainty is introduced in each prediction trajectory step by sampling a low-dimensional noise vector 𝜖ₜᵢ, used for parameter-shared conditional normalization during model forward propagation. This can be understood as perturbing neural network weights to obtain parameters 𝜃ₜᵢ, thus viewed as sampling neural network parameters.

To generate N ensemble members under random uncertainty, simply conditionally generate N different 𝜖ₜᵢ independently. Cognitive uncertainty is modeled by integrating outputs from multiple independently trained models M𝑗, each with its own set of parameters {𝜃*𝑗, Δ𝑗}, and generating subset ensemble members separately as described above.

Evaluation Results

The new experimental cyclone model can simultaneously consider path and intensity prediction, with internal evaluations showing it is currently the best in cyclone path and intensity prediction.

It is trained on two types of different data:

First, global reanalysis datasets reconstructed from millions of observational data;