Technology & Science
HUAWEI’s AI Model Surpasses Supercomputers in Global Weather Forecasting
The success of HUAWEI's model is yet another example of China's unique approach to large models. Instead of training a general-purpose one, AI models are developed and trained with a specific application in mind, using industry-relevant data.

July 7, 2023

A team of researchers from HUAWEI in China has developed an AI model called Pangu-Weather, which has been shown to outperform the world’s best numerical weather prediction (NWP) system in global weather forecasting. For nearly a century, numerical weather prediction has been the dominant approach for forecasting weather, but as models have become more complex, numerical methods have hit bottlenecks. Huawei Cloud scientists have pioneered a new way forward with artificial intelligence.

Huawei Cloud launched a research initiative two years ago to explore AI for weather forecasting and found that existing AI methods lagged well behind numerical forecasting in accuracy, particularly for predicting extreme weather events. The researchers identified two main reasons for the gap: AI models were based on 2D neural networks that struggle with 3D weather data, and AI methods lack the mathematical constraints of numerical modeling, causing errors to accumulate.

To overcome these challenges, Huawei Cloud developed a 3D Earth-Specific Transformer called 3DEST to handle complex 3D weather data. They also proposed a hierarchical temporal aggregation strategy to minimize iteration errors. The result is their Pangu-Weather model, which trains separate AI systems to predict weather at intervals of 1 hour, 3 hours, 6 hours, and 24 hours.

The Huawei Cloud R&D team trained Pangu-Weather on 39 years of global reanalysis data and tested it on various variables such as temperature, pressure, wind speed, and precipitation, comparing it to the European Centre for Medium-Range Weather Forecasts (ECMWF), considered the best NWP system globally. Published in Nature, the results demonstrate Pangu-Weather’s superior deterministic forecast accuracy compared to ECMWF on all tested variables, particularly in medium-range forecasts. The AI model also outperforms ECMWF in extreme weather forecasts, ensemble forecasts, and tropical cyclone tracking.

Pangu-Weather is significantly faster than conventional NWP methods and can run on a single GPU device. The researchers suggest its potential applicability in other domains, such as air quality forecasting, climate modeling, and ocean dynamics. The researchers acknowledge Pangu-Weather’s limitations, including dependence on high-quality reanalysis data and the lack of physical constraints. They emphasize the AI model’s potential to complement NWP methods and provide alternative solutions for weather forecasting.

Huawei Cloud’s Pangu-Weather model has attracted attention for its accurate forecasting, such as predicting Typhoon “Mawar” changing paths five days in advance. The World Meteorological Organization (WMO) plans to incorporate AI elements into their 2024-2027 strategic plan to promote meteorological technology development and improve global forecasting capabilities.

The future of AI weather forecasting lies in leveraging massive meteorological data, enhancing computing power to handle ultra-high resolution data, and utilizing AI models with extremely high computing power. As new data, resources, and techniques become available, AI is poised to transform weather prediction and provide life-saving early warning capabilities, especially for developing nations. Overall, the breakthrough of Pangu-Weather demonstrates the power of AI to solve complex problems, extract new insights from big data, and advance science in ways that could benefit humanity.

The success of HUAWEI’s model is yet another example of China’s unique approach to large models. Instead of training a general-purpose AI model, such as ChatGPT, and hoping it can solve all problems and become AGI in the future, AI models are developed and trained with a specific application in mind, using industry-relevant data. They demonstrate better performance in complex tasks for targeted industries than generalist models. Compared to generalist large models, this type of specialist model also requires much less computing power to operate. In this case, just a single GPU would suffice.

This approach results in a very different landscape for AI in China, one that is considerably more fragmented, with more players carving out a niche for themselves, rather than being dominated by a handful of technology giants.


Editor-in-Chief, The China Academy
Share This Post