What happened
Researchers at the Meteorological Research Institute of the Japan Meteorological Agency (JMA) and three universities tested Pangu-Weather on typhoons in the western North Pacific. They ran Huawei's pretrained model without fine-tuning from the real-time analyses of JMA's Global Spectral Model (GSM). They compared its forecasts of typhoon tracks up to five days ahead for all 64 named typhoons of 2021 to 2023 with the GSM's forecasts from the same analyses. The authors report that Pangu-Weather's mean position errors were smaller at every lead time and that the differences were statistically significant. The improvement was 8%, 19%, 18%, 15% and 9% for one- to five-day forecasts. A five-day run took under a minute on one GPU, against about 19 minutes for the GSM. Intensity errors were larger than the GSM's, and forecast busts such as Typhoon Haikui in 2023 still occurred. The Journal of the Meteorological Society of Japan published the study online on 27 February 2025. The authors propose adding the model's tracks to the consensus that underlies JMA's track forecasts.
Sources
-
Peer-reviewed study, by authors with no tie to the system.
Results indicate that the ML forecasts exhibit smaller position errors compared to the NWP model, alleviate the westward bias around Japan, and retain its forecast accuracy for TCs with unusual paths
Details
AI system
- Workflow
-
The pretrained model ran without fine-tuning from JMA's real-time GSM analyses. Forecasts from 0000 and 1200 UTC were verified against JMA best tracks up to five days ahead.
Follow-up
- Conditions for replication
-
An operational analysis must be available on time to start the model. Intensity forecasts need other tools, since the model's typhoons were too weak.
Entry details
- Sector
-
Weather forecasting / tropical cyclones.
- Purpose
-
Huawei Cloud built Pangu-Weather to produce fast global medium-range forecasts with a deep learning model trained on ERA5 reanalysis.
Not yet checked: model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, beneficiaries, reach, distribution and access, duration, every measurement question, every money question, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.
Cite this record
Peter H. Vartanian. (2026). Pangu-Weather tracks typhoons better than Japan's global model (AIBD-0760). Artificial Intelligence Benefits Database, edition of 31 October 2026. https://petervartanian.xyz/aibd/records/AIBD-0760.html (archived at https://doi.org/10.7910/DVN/AADSII, version 4.0)