Batched neural-network 4D-Var improves ECMWF hybrid forecasts out of sample
Original title: Batched NN 4D-Var: a performant hybrid model framework for the Integrated Forecasting System
Sequential online training cut errors by more than 10% in-period but failed to generalize; batching did.
Researchers train a neural network inside ECMWF’s Integrated Forecasting System so 4D-Var learns state-dependent model-error corrections from observations. A targeted variant estimates only network parameters from a precomputed analysis; a batched extension averages parameter increments over randomly ordered assimilation windows. At intermediate resolution, sequential online learning reduced errors by more than 10% in some training-period cases but did not generalize. Batched training improved out-of-sample forecast skill, with larger batches converging faster and more robustly, without artificial smoothing.
Unconfirmed: Skill gains are the authors’ ECMWF experiments; independent operational verification is not reported.
- Only neural-network parameters are estimated; the atmospheric state comes from a precomputed analysis.
- Batching averages increments across windows, inspired by mini-batch stochastic optimization.
- Experiments use ECMWF’s IFS at an intermediate resolution.
- Diagnostics say corrections keep realistic forecast variability.
- Products
- Integrated Forecasting System
- Organizations
- European Centre for Medium-Range Weather Forecasts
From the original article
1 min read at arxiv.orgHybrid forecasting systems combining machine learning and numerical weather prediction offer a promising route toward improved forecast skill while preserving physical consistency. In this study, we investigate online learning of state-dependent model-error corrections within the Integrated Forecasting System using 4D-Var.
By Alban Farchi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.