Truace
Subscribe
60/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

AI weather models cut Inner Mongolia temperature errors by 13–16% but miss heavy rain, study finds

Study title: Comparative Verification of IFS-Control, AIFS, and WeatherNext 2 Mean for Short-Range (6–84 h) Temperature and Precipitation Forecasts over Inner Mongolia

Evidence type
Observational study
Sample
n = 117
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2514cite or check it

Study title: Comparative Verification of IFS-Control, AIFS, and WeatherNext 2 Mean for Short-Range (6–84 h) Temperature and Precipitation Forecasts over Inner Mongolia

A verification study using observations from 117 stations in Inner Mongolia compared three forecast products over the June–August 2026 flood season at lead times of 6 to 84 hours: the conventional IFS-Control, the AI-based AIFS, and the 64-member ensemble mean of WeatherNext 2. For 2 m temperature, mean absolute error was about 15.7% lower for AIFS and 13.1% lower for WeatherNext 2 than for IFS-Control.

Read the study abstract

Whether advances in global weather forecasting translate into local forecast skill requires regional verification. Using observations at 117 stations in Inner Mongolia during the June–August 2026 flood season, we compared three publicly distributed forecast fields at 6–84 h lead times: the Integrated Forecasting System control forecast (IFS-Control), the Artificial Intelligence Forecasting System (AIFS), and WeatherNext 2’s 64-member ensemble mean (WN2). Because WN2 is an ensemble mean and the others are single…

55Gain side

How the two sides scoreClose call

56Problem side

The gain

In a verification against 117 Inner Mongolia stations over the June–August 2026 flood season, temperature mean absolute error was about 15.7% lower for AIFS and 13.1% lower for WeatherNext 2 than for the IFS-Control forecast.

The problem

In the same station-based study, both AI products under-forecast heavy-rain frequency and extremes, and IFS-Control clearly beat AIFS for 6-hour heavy rain (13 mm or more).

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

Rain forecasts were mixed. Both AI products beat IFS-Control for 6-hour moderate rain (4 mm or more). For 6-hour heavy rain (13 mm or more), IFS-Control clearly outperformed AIFS, while its edge over WeatherNext 2 did not hold up under a stricter multiple-comparison correction.

The temperature gains came mainly from smaller correlation-error terms, while IFS-Control had the smallest systematic bias. Both AI products under-forecast heavy-rain frequency and extremes, which matters for flood-season warnings. A limited member experiment suggests ensemble averaging may contribute to this, though it cannot be separated from model differences.

Main points

  1. The study used observations from 117 stations in Inner Mongolia during June–August 2026.
  2. Forecasts were verified at 6–84 hour lead times for 2 m temperature and precipitation.
  3. Temperature mean absolute error was about 15.7% lower for AIFS and 13.1% lower for WeatherNext 2 than for IFS-Control.
  4. IFS-Control had the smallest systematic temperature bias; the AI gains came mainly through smaller correlation-error terms.
  5. Both AI products outperformed IFS-Control for 6-hour moderate rain (4 mm or more).
  6. IFS-Control clearly beat AIFS for 6-hour heavy rain (13 mm or more); its advantage over WeatherNext 2 did not survive stricter correction.
  7. Both AI products under-forecast heavy-rain frequency and extremes.

The full story

The researchers asked whether global advances in AI weather forecasting carry over to local skill, which they say requires regional verification. They compared three publicly distributed forecast fields: the IFS control forecast, AIFS, and the 64-member ensemble mean of WeatherNext 2.

Because WeatherNext 2 is an ensemble mean and the other two are single deterministic forecasts, the comparison is between operational products, not the underlying models. Verification used Murphy error decomposition and block-bootstrap sensitivity tests.

For temperature, the AI products had lower mean absolute error, mainly through smaller correlation-error terms, although IFS-Control had the smallest systematic bias. For precipitation, the picture depended on intensity: AI products led for moderate rain at 6 hours, while IFS-Control led for heavy rain, clearly over AIFS and less certainly over WeatherNext 2.

Both AI products under-forecast heavy-rain frequency and extremes. A limited member experiment suggests ensemble averaging may play a part, but not separably from differences between models.

What this doesn’t fix

The findings apply only to the product versions, region and season evaluated, and the comparison pits an ensemble mean against single deterministic forecasts, so it cannot isolate the underlying models.

Sources

1 on file
  1. Atmospheredoi:10.3390/atmos17111053
    Peer-reviewed2026-10-27

Talk it over

Four questions to start onThis record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

Start with a question

AI weather models cut Inner Mongolia temperature errors by 13–16% but miss heavy rain, study findshttps://truvace.com/item/comparative-verification-of-ifs-control-aifs-and-weathernext-2-mean-for-short-ra

  1. This record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

    The gain In a verification against 117 Inner Mongolia stations over the June–August 2026 flood season, temperature mean absolute error was about 15.7% lower for AIFS and 13.1% lower for WeatherNext 2 than for the IFS-Control forecast.

    The problem In the same station-based study, both AI products under-forecast heavy-rain frequency and extremes, and IFS-Control clearly beat AIFS for 6-hour heavy rain (13 mm or more).

    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 an observational study with a sample of 117, from Atmosphere.

    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 The findings apply only to the product versions, region and season evaluated, and the comparison pits an ensemble mean against single deterministic forecasts, so it cannot isolate the underlying models.

    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