March 2020

Don’t overlook internal variability when modelling past weather

Reflections from a recently published paper on internal model variability.

Our daily weather forecasts, rain or shine, are derived from interpreting sets of simulations all for the same time period. Known as ‘ensembles’, each simulation contains an ever so slightly different (perturbed) initial condition: the observed state of the Earth’s surface and atmosphere. From here, using modelling to propagate forward in time, these small changes in initial conditions can lead to drastic differences between simulations – the chaos effect. Realised by Edward Lorenz in the 1960s, you may be familiar with expressions like “a butterfly flapping its wings may lead to a tornado elsewhere” or equally “… a nice sunny day”.

Within the atmospheric science community, it is common for us to model past weather events to help understand physical processes (hence improve modelling) or model extreme events in fine resolutions (like recent UK flooding). Several excellent community-based regional climate models (RCMs) exist permitting these types of studies; regional in the sense events happened, therefore we already have global reanalysis inputs (observation based) that RCMs can refine (known as dynamic downscaling). This saves on impractical amounts of computing that would otherwise be required to simulate the whole world in high resolution. That being said exciting developments are fast being made in cloud computing.

Although chaos is well-known, unlike forecasting the future, RCMs take regular updates (usually every 6 hours) on past weather conditions at their boundaries (e.g. a limited domain cantered over Europe). It is thought that this should largely constrain internal model variability (IMV) or noise; the overall climatology within the RCM will be comparable to that taken from the global reanalysis input. However, as we demonstrate in our findings this is not always the case. RCMs after all still contain code approximations to non-linear physical processes.

We ran a large ensemble using the Weather Research and Forecasting (WRF) model, finding sizeable differences in modelled temperatures (over 8oC at times) and other meteorological fields in a fixed analysis window. The range in variability was found to be driven by conditions at the RCM boundaries: slow input winds allow the model to be more independent. This means that chaos dominates, and fast input winds effectively flush the model with conditions from the large global input.

Take urban heat island (UHI) studies by way of example to the magnitude of this impact. UHI studies typically focus on periods of calm weather as this is known to lead to the largest UHI intensities. As we demonstrate in the image below, these are also the periods of the largest internal model variability between simulations. Here, each image is for the same time slice yet from a different ensemble members. We see large differences in temperature suggesting UHI results based on single runs should be heavily questioned. Considering internal model variability is rarely considered by the RCM community we hope our study will help improve best-practices and be reflected in future research.

Grid of images showing temperatures over London
Example of temperature over Greater London (black outline) from different ensemble members (#) 1, 5, 9, 13, 17, 21, 25, 29 and 33. All images are at the same single time slice, midnight 20th August 2018.


The full paper is available here: Bassett, R., Young, P.J., Blair, G., Samreen, F., & Simm, W. (2020). A large ensemble approach to quantifying internal model variability within the WRF numerical model. Journal of Geophysical Research: Atmospheres,  125.

Author: Richard Bassett