We are applying an algorithm inspired by the phenomenon of spontaneous synchronization to improve the accuracy of climate models.
Physics parametrizations are an essential component of weather and climate models, and represent small scale processes that models cannot resolve. Models sometimes suffer from large biases when compared to observations. Surface winds simulated by the weather prediction model of ECMWF (Integrated Forecast System, IFS) is one such case. The IFS has shown a persistent bias in the direction and magnitude of surface winds, especially over the ocean, when compared to scatterometer satellite data (Belmonte Rivas and Stoffelen 2019, ECMWF Technical Memo 866). Such a bias could potentially have a large impact on the climate simulations of EC-Earth, whose atmospheric component is a version of IFS, by consistently driving oceanic currents with a biased direction and magnitude. A solution to this problem is to tune parameters in the physics schemes of the IFS. Adjusting parameters in the vertical mixing of momentum solves the issue, but this comes at the expense of the medium range prediction skill. Does a better set of parameters exist which fixes the surface winds without compromising on (or even improving) the medium range forecasting skill? Unfortunately, the tuning process currently requires long runs for each test of a set of parameter values, greatly limiting the extent to which models can be tuned (Hourdin et al. 2017).
Online Synchronization Parameter Estimation (OSPE) is an algorithmic method which is currently being developed at KNMI which hopes to solve this problem. It is an algorithm based on the idea of synchronization, where two dynamical systems can, when adequately coupled, over time reach a near-identical “synchronous” state. In this video five metronomes coupled through a freely moving plate on which they are standing are able to influence one another as to reach synchronous states.
Data assimilation already does an excellent job at “synchronizing” the atmospheric state of weather models with that obtained from observations of the Earth’s real atmosphere. However, the model state always has a sizeable error since our models are not a perfect representation of the Earth system, but also because parameter values used in the model are not optimal. The idea of OSPE is to also synchronize the parameters of the model to values which minimize the model error. The parameter values are updated continuously and simultaneously as the model is running, making this type of tuning much more efficient than conventional methods. With OSPE, more ground can be covered by searching for optimal sets of a larger number of parameters.
As part of this MSO project, OSPE will first be applied to the problem of the surface winds of the IFS. So far, the method was tested on a 3-level atmospheric model at truncation 21, and was able to accurately recover parameters of a perfect model based on noisy data, see Fig. 1 and 2. Tuning the solar irradiance constant of the IFS to a self-produced simulation has also shown to be robust and return accurate estimates, see Fig. 3. The remaining question is whether the method is robust enough to train on noisy and complex data of the real atmosphere, and is efficient enough to train parameters to which the atmospheric state is not very sensitive.