Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2503.23953 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866916667842363392 |
|---|---|
| author | Kent, Chris Scaife, Adam A. Dunstone, Nick J. Smith, Doug Hardiman, Steven C. Dunstan, Tom Watt-Meyer, Oliver |
| author_facet | Kent, Chris Scaife, Adam A. Dunstone, Nick J. Smith, Doug Hardiman, Steven C. Dunstan, Tom Watt-Meyer, Oliver |
| contents | Machine learning weather models trained on observed atmospheric conditions can outperform conventional physics-based models at short- to medium-range (1-14 day) forecast timescales. Here we take the machine learning weather model ACE2, trained to predict 6-hourly steps in atmospheric evolution and which can remain stable over long forecast periods, and assess it from a seasonal forecasting perspective. Applying persisted sea surface temperature (SST) and sea-ice anomalies centred on 1st November each year, we initialise a lagged ensemble of winter predictions covering 1993/1994 to 2015/2016. Over this 23-year period there is remarkable similarity in the patterns of predictability with a leading physics-based model. The ACE2 model exhibits skilful predictions of the North Atlantic Oscillation (NAO) with a correlation score of 0.47 (p=0.02), as well as a realistic global distribution of skill and ensemble spread. Surprisingly, ACE2 is found to exhibit a signal-to-noise error as seen in physics-based models, in which it is better at predicting the real world than itself. Examining predictions of winter 2009/2010 indicates potential limitations of ACE2 in capturing extreme seasonal conditions that extend outside the training data. Nevertheless, this study reveals that machine learning weather models can produce skilful global seasonal predictions and heralds a new era of increased understanding, development and generation of near-term climate predictions. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2503_23953 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Skilful global seasonal predictions from a machine learning weather model trained on reanalysis data Kent, Chris Scaife, Adam A. Dunstone, Nick J. Smith, Doug Hardiman, Steven C. Dunstan, Tom Watt-Meyer, Oliver Atmospheric and Oceanic Physics Machine learning weather models trained on observed atmospheric conditions can outperform conventional physics-based models at short- to medium-range (1-14 day) forecast timescales. Here we take the machine learning weather model ACE2, trained to predict 6-hourly steps in atmospheric evolution and which can remain stable over long forecast periods, and assess it from a seasonal forecasting perspective. Applying persisted sea surface temperature (SST) and sea-ice anomalies centred on 1st November each year, we initialise a lagged ensemble of winter predictions covering 1993/1994 to 2015/2016. Over this 23-year period there is remarkable similarity in the patterns of predictability with a leading physics-based model. The ACE2 model exhibits skilful predictions of the North Atlantic Oscillation (NAO) with a correlation score of 0.47 (p=0.02), as well as a realistic global distribution of skill and ensemble spread. Surprisingly, ACE2 is found to exhibit a signal-to-noise error as seen in physics-based models, in which it is better at predicting the real world than itself. Examining predictions of winter 2009/2010 indicates potential limitations of ACE2 in capturing extreme seasonal conditions that extend outside the training data. Nevertheless, this study reveals that machine learning weather models can produce skilful global seasonal predictions and heralds a new era of increased understanding, development and generation of near-term climate predictions. |
| title | Skilful global seasonal predictions from a machine learning weather model trained on reanalysis data |
| topic | Atmospheric and Oceanic Physics |
| url | https://arxiv.org/abs/2503.23953 |