May 28, 2026, 13:37 a.m. ET | ⏱️12–14 minutes

By Daniel Brooks

A noteworthy technological turning point

Weather forecasting affects everyone's daily life — from deciding whether to bring an umbrella tomorrow to preparing for extreme events like typhoons and heatwaves.

For decades, the world's weather agencies have relied mainly on a method called "numerical weather prediction."

In simple terms, it divides Earth's atmosphere into millions of small grid cells and uses supercomputers to solve complex physics equations, calculating how the weather will change.

This approach has been refined over nearly a century and is very mature, but it has also run into clear bottlenecks: huge computational demands, long processing times, and increasingly slow improvements in accuracy.

Around 2020, some technology companies began trying to use artificial intelligence for weather forecasting.

AI takes a different approach — it does not solve physics equations. Instead, it learns patterns of atmospheric evolution from vast amounts of historical weather data and then makes predictions quickly.

Early models were fast, but their accuracy often lagged behind traditional methods.

Then in 2023, a research team from Huawei Cloud published results in the journal Nature. Their "Pangu-Weather" model achieved, for the first time, accuracy on several key metrics that surpassed the operational system of the European Centre for MediumRange Weather Forecasts (ECMWF), while increasing prediction speed by more than 10,000 times.

Many meteorologists see this as an important turning point.

This article, based on published papers and publicly available industry reports, provides an objective introduction to PanguWeather's technical features, realworld performance, and limitations, and explores what it might mean for weather forecasting and for scientific computing more broadly.

1. The dilemma of traditional forecasting: why a new approach is needed

Traditional numerical weather prediction works by discretizing the atmosphere into a threedimensional grid and solving, at each grid point, partial differential equations that describe air motion, heat exchange, water vapor changes, and other processes.

As computing power increased, grid cells could be made smaller and physical processes could be represented in greater detail, steadily improving prediction accuracy.

However, in recent years the growth of computing power has slowed while the complexity of physical models continues to increase.

According to a previous ECMWF report, the improvement in global mediumrange forecast accuracy is roughly one extra day of useful lead time per decade.

At the same time, a single global forecast covering 7–10 days — even on the most powerful supercomputers — takes several hours to complete. This makes it difficult to update forecasts frequently or to run large "ensemble forecasts" with many different starting conditions to assess uncertainty.

These problems — "computational explosion" and "timeliness bottlenecks" — are pushing scientists and engineers to look for new paths.

Artificial intelligence, especially deep learning, is seen as a promising direction because of its patternrecognition ability and speed.

2. Early attempts at AI weather forecasting and the accuracy gap

Before Pangu, several organizations had tried using deep learning for weather prediction.

Google DeepMind's GraphCast and NVIDIA's FourCastNet, for example, achieved good results. Once trained, these models could produce a global forecast in just seconds to minutes — orders of magnitude faster than traditional methods.

Nevertheless, early AI models generally suffered from an "accuracy gap": for mediumrange forecasts (5–7 days out), their errors were often larger than those of traditional numerical models.

Two main reasons explain this.

First, most models were based on twodimensional neural networks, which struggle to handle the complex threedimensional structure of the atmosphere — for example, the interactions of temperature and wind between different altitude levels.

Second, AI models use an "iterative forecasting" approach: they feed the prediction for the current time step as input for the next step, so small errors accumulate and amplify, causing the forecast to drift significantly after a few days.

Therefore, for AI to truly match or exceed traditional methods, key innovations in architecture and error control are necessary.

3. Three key technical breakthroughs of the Pangu-Weather model

The Huawei Cloud team proposed three core innovations in their paper to address the accuracy problem.

Breakthrough 1: a 3D neural network that understands the atmosphere's "three-dimensional structure"

The atmosphere is not a flat layer — it is a threedimensional fluid that extends tens of kilometers from the ground to the upper air, with many vertical pressure levels. Variables at different heights (temperature, humidity, wind speed) are strongly coupled.

Pangu uses an architecture called the "3D EarthSpecific Transformer," whose core component is a threedimensional convolutional layer — it can extract features in both horizontal and vertical directions simultaneously.

The input data cover 13 vertical pressure levels and 13 weather parameters (including geopotential, humidity, wind speed, temperature, and sealevel pressure), with a horizontal resolution of 0.25° × 0.25° (roughly 28 kilometers at the equator) and a temporal resolution of 1 hour.

This three-dimensional structure allows the model to learn vertical physical processes that two-dimensional models miss, such as the development of convective clouds and the interaction between upperlevel and lowerlevel jets.

Breakthrough 2: a hierarchical temporal aggregation strategy to control error accumulation

To reduce the error amplification caused by many iteration steps, Pangu does not train just one "1-hour step" model. Instead, it trains four submodels with different forecast intervals: 1 hour, 3 hours, 6 hours, and 24 hours.

When making a prediction, the system automatically chooses the fewest iteration steps — for example, to predict 24 hours ahead, it can use the 24hour model directly rather than stepping 24 times with the 1hour model.

The paper also mentions the use of a "temporal consistency loss function" during training, which forces the predictions from different time steps to be statistically consistent. This acts as a soft constraint, preventing longrange forecasts from drifting away from physical plausibility.

Thanks to this design, Pangu accumulates errors much more slowly than previous AI models for forecasts from 1 to 7 days.

Breakthrough 3: largescale data and computing power

The model was trained on 43 years (1979–2021) of ECMWF ERA5 reanalysis data — one of the most complete and highestquality global atmospheric history datasets available.

Training the four submodels for 100 epochs each required 192 NVIDIA V100 GPUs running for 16 days. The paper's authors note that even after 100 epochs, the models had not fully converged, meaning more computing power could further improve accuracy.

For inference, Pangu runs on a single V100 GPU and takes only 1.4 seconds to produce a 24hour global forecast — about 10,000 times faster than traditional numerical methods.

4. How Pangu performs compared to traditional methods

The paper systematically compared Pangu's predictions with ECMWF's IFS system, widely recognized as the world's most advanced operational numerical weather prediction system.

The results showed that for forecast lead times of 1 to 7 days, Pangu achieved higher accuracy than IFS on all ten key weather variables tested (including 500 hPa geopotential, 850 hPa temperature, 2meter temperature, and 10meter wind speed).

In other words, Pangu's predictions had smaller errors compared to actual observations.

Pangu also extended the "useful forecast lead time" by about 1–2 days. For example, if IFS's 7day forecast is no longer reliable, Pangu's 8 or 9day forecast might still be as good as IFS's 7day forecast.

For extreme weather events, Pangu also demonstrated capability.

During Typhoon Mawar in May 2023, an assessment by the China Meteorological Administration (CMA) showed that Pangu accurately forecast the typhoon's turning path in the western North Pacific east of Taiwan five days in advance.

At the 19th World Meteorological Congress, ECMWF Director Florence Rabier presented a case study of a -29°C cold wave in Finland, noting that Pangu recognized the severity of that extreme event earlier than some other models.

She also emphasized that AI methods consume few computational resources, providing an important opportunity for developing countries that lack large supercomputers to improve their forecasting capabilities.

As of May 2026, the AI weather forecasting field has become competitive.

On May 12, 2026, ECMWF officially launched its own AI forecasting system, AIFS v2, and stopped running external AI models including Pangu. ECMWF's DirectorGeneral said this marks the entry of a new generation of AI forecasts into operational service.

Many in the industry see this as a key signal that AI weather forecasting is moving from "lab validation" to "operational deployment."

5.Applications: from grid dispatch to renewable energy forecasting

Pangu is no longer confined to research.

For example, Shenzhen Energy Group, a utility based in Shenzhen, China, partnered with Huawei Cloud to develop a renewable energy power prediction platform based on Pangu.

Industry reports indicate that this platform integrates global weather background, regional downscaled data, and realtime observations from individual power stations. It improved weather forecast accuracy by 15% and increased wind power generation forecast accuracy by 10%.

This means weather forecasts can be used directly to guide grid dispatch — for instance, adjusting thermal power output before strong winds arrive, thereby reducing wind curtailment.

In addition, a regional model called "Zhiji" — jointly developed by Huawei Cloud and the Shenzhen meteorological service — increased spatial resolution to 3 kilometers and achieved operational AIdriven ensemble forecasting (running 31 different initial conditions simultaneously).

In validation covering more than 2,200 rain gauge stations in Guangdong Province during the 2024 flood season, Zhiji outperformed traditional numerical models in predicting the location and intensity of heavy rainfall.

These cases show that AI weather models are moving from "showing capabilities" to "solving real problems," and their value metrics are shifting from technical accuracy to economic benefits — such as reduced wind curtailment and property losses prevented by early warnings.

6. Limitations: current shortcomings of AI weather forecasting

Despite Pangu's impressive achievements, several independent studies have pointed out inherent limitations of AI weather forecasting.

The World Meteorological Organization (WMO) held a special session on AI weather prediction in Abu Dhabi in September 2025. The session report noted ongoing challenges including limited ability to generalize to extreme events, lack of explainability, data imbalance, and insufficient publicprivate collaboration.

A study led by the Karlsruhe Institute of Technology (KIT) in Germany and the University of Geneva systematically evaluated several major AI models including GraphCast and PanguWeather.

The findings: when faced with recordbreaking extreme weather events (such as unprecedented heatwaves, cold waves, and storms), the traditional physicsbased highresolution model (HRES) consistently outperformed the AI models.

The AI models showed significantly larger forecast errors for these events and systematically underestimated the intensity and frequency of extremes.

The researchers attributed this to a fundamental limitation of pure datadriven models: neural networks can only learn patterns that exist in their training data. When climate change pushes the atmosphere into states never seen before in history, AI models struggle to extrapolate reliably.

Physicsbased models, on the other hand, are rooted in fundamental laws; even when facing an unprecedented state, their calculations remain physically consistent.

Therefore, the prevailing view in the field is that AI models should not completely replace traditional physicsbased models.

A safer approach is to combine the two — use AI for fast, inexpensive guidance, and use physicsbased models as a safety check, especially for highstakes decisions such as evacuations before a landfalling typhoon.

7. Conclusion: the new role of physical laws — "learned" rather than "replaced"

PanguWeather, through its 3D neural network, hierarchical temporal aggregation strategy, and largescale training data, became the first AI model to exceed the deterministic forecast accuracy of a worldleading numerical weather prediction system, while increasing computational speed by four orders of magnitude.

But this does not mean "AI has surpassed the laws of physics."

A more accurate description is that Pangu changed how physical knowledge is used in weather forecasting: instead of manually writing partial differential equations and solving them one by one on a supercomputer, the model encodes physical laws as a kind of "prior structure" into the neural network, allowing the model to learn statistical patterns of atmospheric evolution from data.

In the future, weather forecasting will likely move toward a "physics + AI dualtrack" system.

As ECMWF showed in May 2026 by upgrading both its physicsbased model (IFS) and its AI model (AIFS) at the same time, the two approaches have different strengths and complement each other. The physics model provides reliability at the extremes, while the AI model offers speed and patterndiscovery ability.

Pangu's success is seen as a typical example of the "AI for Science" paradigm. Similar approaches are being tried in ocean modeling, drug molecule design, materials science, and other fields.

The core idea is: train large models on massive amounts of data, let the machine "learn" the underlying rules of a complex system, and then perform inference and prediction very quickly. This could change the research and development paradigm in many scientific fields — shifting from "computationally intensive" to "data and model-intensive."

Of course, AI weather forecasting is still in a fastdeveloping but not yet mature stage. Its longterm stability, its ability to adapt to unknown extreme climates, and how it will be seamlessly integrated into existing operational systems — all these need more time and more research to verify.

But what is certain is that it has opened a new door — inviting us to rethink: when machines learn the laws of physics, how will scientists and AI explore this complex natural world together?

References

[1] Bi, K., Xie, L., Zhang, H., et al. (2023). Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619, 533–538.

[2] ECMWF. (2026, May 12). ECMWF launches AIFS v2: A new generation of datadriven forecasting. ECMWF Newsletter.

[3] Karlsruhe Institute of Technology & University of Geneva. (2025). Systematic evaluation of AI weather models against physics-based high-resolution models under record-breaking extremes. (Preprint / conference presentation)

[4] World Meteorological Organization. (2025, September). Summary report on AI for weather forecasting: Challenges and opportunities. WMO Abu Dhabi Session.

[5] Huawei Cloud. (2023). PanguWeather: AI for global weather forecasting. GitHub repository & technical whitepaper.

About the Author

Daniel Brooks covers the intersection of technology, business, and industrial transformation. His reporting focuses on robotics, advanced manufacturing, cloud computing, and emerging technology markets. He aims to provide clear, evidence-based analysis of how technological innovation is reshaping industries worldwide.

Editor's Note

This article is based on publicly available research papers, industry reports, and statements from meteorological organizations as of May 2026. The author is a technology writer and industry observer, not a meteorologist or AI researcher. The goal is to provide an objective, evidencebased overview of PanguWeather's capabilities and limitations for a general audience of technology enthusiasts. All factual claims are supported by cited sources. The views expressed in the conclusion are the author's own analysis and do not represent any organization.