Truace
Subscribe
60/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Hybrid forecast cuts typhoon track errors 30%–40% in a 10-storm study

Study title: An online spectral nudging-based correction system: improving physical model forecasts by incorporating large-scale circulations derived from machine learning models

Evidence type
Controlled or validated comparison
Sample
n = 10
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2557cite or check it

Study title: An online spectral nudging-based correction system: improving physical model forecasts by incorporating large-scale circulations derived from machine learning models

Researchers at the China Meteorological Administration built a hybrid system that steers the large-scale circulation of its CMA-GFS physical forecast model toward forecasts from the AI model FuXi during the model run. In 10 typhoons from 2024, the hybrid's track errors were about 30%–40% lower than the physical model's alone. On day 5, the physical model's mean track error was about 365 km, versus 206 km for FuXi.

Read the study abstract

The development of traditional numerical weather prediction (NWP) relies on continuous advances in observation technology, data assimilation methods, numerical and parameterization algorithms, and the steady growth of computational resources, resulting in lengthy development cycles and relatively slow improvements in forecast skill. In recent years, machine learning (ML)-based weather forecasting models have advanced rapidly, and in some aspects, outperform traditional physical models, particularly in forecastin…

An online spectral nudging-based correction system: improving physical model forecasts by incorporating large-scale circulations derived from machine learning models
"Released to Public: Three Typhoons in Pacific on August 7, 2006 by Jeff Schmaltz (NASA Image)" by pingnews.com is marked with Public Domain Mark 1.0. To view the terms, visit https://creativecommons.org/publicdomain/mark/1.0/.

The gain

In a proof-of-concept test on 10 western Pacific typhoons from 2024, steering China's physical forecast model toward AI forecasts cut typhoon track errors by about 30%–40% and extended precipitation forecast lead time by about 2 days.

In brief

The approach tries to keep the strengths of both systems. FuXi forecast large-scale circulation better, while the physical model kept more realistic typhoon intensity and fine-scale rainfall detail. In batch tests for July 2024 and January 2025, the hybrid's large-scale skill stayed close to FuXi's, and precipitation skill on day 5 matched the physical model's day 3 skill.

The work is a proof of concept. Every run used ERA5 reanalysis data as the starting point, not the CMA's own analysis, so it does not yet show how the system would perform in operations.

Main points

  1. The hybrid adds a correction term to the CMA-GFS equations so its large-scale circulation is pulled toward FuXi forecasts every 6 hours, using a truncation wavenumber of 21 (wavelengths above about 2,000 km).
  2. In batch tests, FuXi's 500-hPa geopotential height forecast lead time beat CMA-GFS by about 2 days in July 2024 and 1.5 days in January 2025, and the hybrid's skill stayed largely consistent with FuXi's.
  3. On day 5, CMA-GFS typhoon track error averaged about 365 km versus 206 km for FuXi; the hybrid matched FuXi, a 30%–40% improvement over CMA-GFS.
  4. FuXi underestimated typhoon intensity, with errors of about 30 hPa in central pressure and 15–20 m/s in maximum wind, against about 20 hPa and 10 m/s for CMA-GFS; the hybrid performed like CMA-GFS and did better at longer lead times.
  5. In a heavy rain case in Inner Mongolia on July 25–26, 2025, CMA-GFS put the rain centre too far south at 50–75 mm and FuXi gave under 50 mm; the hybrid reached 100–150 mm against observed 100–200 mm.
  6. The hybrid reduced sudden drops in day-5 forecast skill that CMA-GFS showed on four dates (July 17, January 10, 14 and 23).
  7. Nudging is applied mainly between 600 hPa and 200 hPa and not to humidity; the authors report negative contributions above 10 hPa that need further analysis.

The full story

The CMA-GFS runs at 0.125° resolution with 87 vertical levels, and its operational lead time for 500-hPa geopotential height is about 7 days in summer and 9 days in winter, behind the ECMWF, the UK Met Office and the US centres. FuXi, developed at Fudan University, produces 15-day forecasts at 0.25° resolution and was used without fine-tuning. Because FuXi outputs only 13 pressure levels, nudging was weighted toward the mid and upper troposphere.

Batch experiments ran daily for July 2024 and January 2025, each for 10 days. With ERA5 starting data, CMA-GFS had an Northern Hemisphere lead time of about 8.5 days in July 2024 and 9.3 days in January 2025. Equitable threat scores for rainfall across China improved at all thresholds, and the gain grew with lead time. Day-5 hybrid scores were comparable to CMA-GFS day-3 scores.

The typhoon test used 10 storms from 2024 that lasted more than 5 days, including Gaemi, Yagi, Krathon and Man-yi, all run at 0.25° resolution from ERA5 data. The hybrid kept the physical model's more realistic intensity while gaining FuXi's track accuracy. Next steps are a cycling 4D-Var system and possibly swapping in the CMA's own FengQing AI model.

Sources

1 on file
  1. Geoscientific Model Developmentdoi:10.5194/gmd-19-8673-2026
    Peer-reviewed2026-09-17

Talk it over

Four questions to start onWho gets this benefit, and who was left out of the study or story?

Start with a question

Hybrid forecast cuts typhoon track errors 30%–40% in a 10-storm studyhttps://truvace.com/item/an-online-spectral-nudging-based-correction-system-improving-physical-model-fore

  1. Who gets this benefit, and who was left out of the study or story?

    The gain In a proof-of-concept test on 10 western Pacific typhoons from 2024, steering China's physical forecast model toward AI forecasts cut typhoon track errors by about 30%–40% and extended precipitation forecast lead time by about 2 days.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a controlled comparison with a sample of 10, from Geoscientific Model Development.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat All experiments started from ERA5 reanalysis data because FuXi loses skill with CMA-GFS analyses (lead time fell 0.7 days in the Northern Hemisphere and 2.0 days in the Southern Hemisphere in July 2024), and the full 4D-Var system is not yet complete.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate