The Emergence of AI in Weather Prediction

In the last five years, a significant shift has occurred in weather forecasting with the advent of AI technologies. Researchers have utilized reanalysis weather data from the last half-century to develop AI models capable of forecasting weather conditions. These models analyze historical weather patterns from training data, allowing them to predict future atmospheric states based on current conditions. By employing this iterative process, they can provide forecasts extending several days ahead.

In the video, "AI models outperform traditional weather forecasting, but won't replace them yet," experts discuss how AI is changing the landscape of weather predictions, highlighting the advantages and limitations of AI models compared to traditional methods.

AI weather models typically operate at a spatial resolution of 25 kilometers, which matches the resolution of their training data. High-resolution numerical models, however, function at a 9-kilometer resolution. For forecasts extending up to 10 days, the AI models usually maintain a temporal resolution of six hours.

Current Developments in AI Weather Models

The last two years have seen a surge in the development of AI weather models, with the WeatherBench initiative providing a comprehensive evaluation of these technologies. We focus on models that have been tested in operational settings, initialized with the same data used in numerical weather forecasts.

PanguWeather, created by Huawei, employs a 3D Earth-specific transformer framework. Meanwhile, GraphCast, developed by Google DeepMind, utilizes a graph neural network architecture. Other notable models, such as NVIDIA's FourCastNet and Ryan Keisler's Graph Network, have also been evaluated within this framework, although they currently require a five-day delay for operational readiness.

Metrics for Evaluating Weather Forecasts

In the realm of machine learning, metrics like accuracy and precision are commonplace. However, weather forecasting requires a broader range of metrics due to the diverse needs of its users. For instance, farmers might seek hourly rainfall estimates, while tourists may prioritize temperature forecasts. Solar energy predictions rely heavily on cloud cover data.

Weather forecasts are assessed using scorecards, with the WeatherBench 2 project focusing on crucial surface weather variables that directly impact daily activities. Different meteorological quantities are compared using root mean squared error (RMSE) against actual weather observations and established weather models.

The ECMWF high-resolution forecast serves as the current benchmark. RMSE values are compared against this reference model as well as AI-generated forecasts, with blue indicating superior performance and red signifying inferior results. The operational GraphCast model, for example, demonstrates better predictions for temperature, surface pressure, and near-surface wind speed compared to the ECMWF model, although it falls short in short-term precipitation forecasts.

The Importance of High-Level Atmospheric Predictions

In addition to surface-level assessments, meteorologists also analyze weather models at higher atmospheric levels, which are critical for medium-range forecasts extending up to 15 days. The panel showcasing geopotential at 500 hPa (approximately 5.5 kilometers above ground) indicates that GraphCast consistently surpasses the ECMWF model.

Users can explore and compare various models on the WeatherBench project website.

AI Performance in Extreme Weather Events

One remarkable area where AI weather models excel is in tracking typhoons. Accurate predictions of typhoon landfall are crucial for effective evacuation planning. PanguWeather was the first AI model to report superior accuracy over traditional numerical predictions in this context.

However, deterministic AI models may struggle with predicting rare but extreme events. Traditional forecasts often utilize ensembles of up to 50 members, allowing for a probabilistic approach to forecasting.

The Computational Efficiency of AI Models

Traditional numerical weather prediction is computationally intensive, with organizations like the ECMWF relying on supercomputers to produce high-resolution forecasts multiple times daily. In contrast, AI models can generate forecasts in under a minute using a single TPU or GPU, reducing energy consumption by a factor of 12,000.

Although the inference cost is minimal, the computational resources necessary for generating training data and training the models should not be overlooked. PanguWeather was trained for 16 days on 192 NVIDIA Tesla V100 GPUs, while GraphCast required 28 days on a cluster of 32 Cloud TPU v4 devices, with estimated costs around $38,000.

The Future of Weather Prediction: AI vs. Traditional Methods

While AI weather models have demonstrated superior performance in many areas, there are compelling reasons to continue investing in traditional numerical weather prediction methods. High-resolution numerical forecasts are essential for providing training data and initialization for AI models. Additionally, meteorologists strive to deepen their understanding of atmospheric science, which has historically enhanced forecasting accuracy.

Climate change has a significant impact on weather patterns, influencing the data that AI models rely on. Traditional methods are less susceptible to these shifts. Furthermore, numerical forecasting typically offers better spatial and temporal resolution, which is crucial for predicting localized extreme weather events, such as the severe flooding experienced in Western Europe in 2021.

Future Directions for AI Weather Models

With numerous labs now developing their own AI weather models, advancements are anticipated in the near future. The availability of high-resolution training data presents a technical challenge, as increasing resolution from 25 kilometers to 12.5 kilometers requires integrating eight times more data points during training.

Currently, AI weather models provide fewer meteorological variables than traditional forecasts. For instance, GraphCast does not predict cloud cover, a crucial variable for photovoltaic power predictions.

Weather services are already implementing AI models on a daily basis, continuously evaluating and comparing them against established methodologies. Users can access forecasts generated by various AI models on the ECMWF website.

Summary of Findings

AI weather models are now competitive with traditional numerical models, leveraging extensive historical weather data to deliver reliable forecasts. It is likely that these AI models will become integral to operational weather forecasting in the near future, although traditional numerical methods remain vital due to their physical accuracy and foundational role in model training.

As affordable and accessible AI weather models proliferate, they hold the potential to democratize weather forecasting, making it more widely available.

In the video, "Why AI Creates Better Weather Forecasts," experts delve into the reasons why AI technology can enhance weather predictions, illustrating the benefits and challenges associated with these advancements.

Links

GraphCast: Lam et al, Learning skillful medium-range global weather forecasting, Science 382, 1416–1421 (2023). [free preprint]

PanguWeather: Bi et al, Accurate medium-range global weather forecasting with 3D neural networks. Nature 619, 533–538 (2023).