Abstract
Abstract. Most geophysical models include many parameters that are not fully determined by theory, and can be “tuned” to improve the model's agreement with available data. We might attempt to automate this tuning process in an objective way by employing an optimisation algorithm to find the set of parameters that minimises a cost function derived from comparing model outputs with measurements. A number of algorithms are available for solving optimisation problems, in various programming languages, but interfacing such software to a complex geophysical model simulation presents certain challenges. To tackle this problem, we have developed an optimisation suite (“Cyclops”) based on the Cylc workflow engine that implements a wide selection of optimisation algorithms from the NLopt Python toolbox (Johnson, 2014). The Cyclops optimisation suite can be used to calibrate any modelling system that has itself been implemented as a (separate) Cylc model suite, provided it includes computation and output of the desired scalar cost function. A growing number of institutions are using Cylc to orchestrate complex distributed suites of interdependent cycling tasks within their operational forecast systems, and in such cases application of the optimisation suite is particularly straightforward. As a test case, we applied the Cyclops to calibrate a global implementation of the WAVEWATCH III (v4.18) third-generation spectral wave model, forced by ERA-Interim input fields. This was calibrated over a 1-year period (1997), before applying the calibrated model to a full (1979–2016) wave hindcast. The chosen error metric was the spatial average of the root mean square error of hindcast significant wave height compared with collocated altimeter records. We describe the results of a calibration in which up to 19 parameters were optimised.
Highlights
Geophysical models generally include some empirical parameterisations that are not fully determined by physical theory and which need calibration
This, is rarely the case for a geophysical modelling system, so we will restrict our attention to the field of differential free optimisation (DFO), in which the objective function f can be calculated, but its gradient is not available
The Cyclops Cylc-based optimisation suite offers a flexible tool for tuning the parameters of any modelling system that has been implemented to run under the Cylc workflow engine
Summary
Geophysical models generally include some empirical parameterisations that are not fully determined by physical theory and which need calibration. The calibration process has often been somewhat subjective and poorly documented (Voosen, 2016) but in a more objective approach has the aim of minimising some measure of error quantified from comparisons with measurement (Hourdin et al, 2017). We can turn this into an optimisation problem: namely, to find the minimum of an objective function f (x), where x represents the set of adjustable parameters, and f is a single error metric (e.g. the sum of RMS differences between measured and predicted values of a set of output variables) resulting from a model simulation with that parameter set. The algorithms are encoded in various languages (e.g. Fortran, C, Python, MATLAB) and usually
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
