Bayesian optimisation of parameters in the ECMWF IFS

Title
Bayesian optimisation of parameters in the ECMWF IFS
Technical memorandum
Date Published
12/2025
Secondary Title
Technical Memoranda
Number
934
Author
Annelize van Niekerk
Birgit Suetzl
Nina Raoult
Martin Janousek
Ivan Bastak Duran
Publisher
ECMWF
Abstract Calibrating uncertain parameters is a necessary step in the development of comprehensive numerical weather prediction models such as the ECMWF Integrated Forecasting System (IFS), particularly
when scientific changes have been made. We present a Bayesian optimisation tool and workflow that semi-automates parameter calibration within the fully coupled IFS using perturbed parameter ensemble forecasts and Gaussian Process (GP) emulators. The workflow uses the GP emulator to map from parameter values to a combined forecast-error metric. The tool then proposes new parameter values to sample with the IFS based on minimising the error metric using the GP emulator predictions, following a Bayesian optimisation approach. By iteratively sampling parameter combinations using the IFS, minimisation of the GP emulator yields several plausible optimal parameter sets. This helps avoid local minima and exposes parameter interdependencies.
We demonstrate the approach by calibrating five orographic drag parameters, after changes to the parametrization scheme led to initially degraded large-scale scores. We show that the method is able
to identify parameter sets that deliver neutral to improved upper-air skill and significant improvements in near-surface variables. The workflow is user configurable and applicable to many other
parameters in the IFS. While choices of metrics and priors remain subjective and compensating errors can mask regional trade-offs, Bayesian optimisation offers a practical, systematic alternative to
manual calibration and accelerates the path from scientific change to operationally acceptable performance.
URL https://www.ecmwf.int/en/elibrary/81705-bayesian-optimisation-parameters-ecmwf-ifs
DOI 10.21957/3febe9bc59
Download citation