Predictions for the 2021 North American Winter Storm indicate that ML models might perform better than numerical models for some extreme events: wind-chill, a metric that measures how cold “temperatures feel” in windy conditions, is forecast with smaller errors by two ML models than by HRES
Overall it is hard to tell whether the ML-based models are better than HRES in this case, in one area-aggregated metric it looks as if HRES has the advantage.
* ML models do surprisingly well even on an event of unprecedented magnitude as the 2021 Pacific Northwest heatwave. The figure 👇 shows climatological anomalies for forecasts and the ground-truth data sets ERA5 AND HRES-fc0. The anomalies are computed using the ERA5 climatology of WeatherBench 2.
Our study is motivated by the difficulty of quantifying how well ML methods predict extreme events (rare in both training and test sets). If performance measures are summarised into overall scores, information on the extremes might be hidden. Extreme weather events cause large impacts, however.
A preprint I’ve been working on together with Olivier Pasche, Zhongwei Zhang, @zscheischlerjak.bsky.socialarxiv.org/abs/2404.17652. We compared ML-based weather models and ECMWF’s HRES in case studies on 3 recent high-impact extreme events! Summary & a few thoughts 🧵👇
🔥 PhD position alert! 🌍 Are you interested in studying fire risk and its drivers in Europe under climate change via large datasets? Join our new research group on compound climate extremes! Details & applications👇 recruitingapp-5128.de.umantis.com/Vacancies/28...
Join us in Wageningen in January for the first AgMIP Machine Learning (AgML) workshop! www.wur.nl/en/research-...
New release of the Land Use Change Alerts (LUCA) app, which I have been helping CTrees.orgctrees.org/products/lucaglobal-forest-structure.projects.earthengine.app/view/luca-vi...