Our mission at reask is to use the best available science and technology to build the next generation of natural catastrophe risk products. We are a tech company with a passion for modelling and understanding the natural world.
Lately we have been working to develop models that predict the extremely destructive winds created by Tropical Cyclones (TCs) — also known as Hurricanes or Typhoons. Recent examples include Hurrican Dorian (2019) which is regarded as the worst natural disaster to hit the Bahamas, and Typhoon Faxai (2019) which passed through the centre of Tokyo.
Our approach to modelling in general, and the one we’re applying to TC winds, is to use Machine Learning (ML) algorithms to discover significant features and relationships within Earth System data. The models can then be used to very efficiently infer how a hazard, such as wind, will behave in different or future situations. The beautiful thing about this approach, if it is done well, is that the data becomes the main source of truth. The model is not “made up” by a human, instead it is derived directly from nature.
So a key question is, where can we get the large amount of atmospheric data needed to train ML wind models? We are fortunate that weather and space agencies in the US, Japan and Europe release a lot of high-quality environmental data. However there was, until now, no comprehensive collection of global high resolution surface-level TC winds.
To fill this gap we have developed InCyc — a global database of physically realistic TC simulations. We have used the WRF1 numerical weather model to perform simulations of close to 200 carefully chosen TCs across six global basins. The outcome is over 50 Tb of 1km resolution gridded atmospheric data which includes variables such as wind velocity, temperature, pressure, humidity and precipitation. The image below shows in colour all of the TCs simulated to date.
The purpose of InCyc is to capture a large range of physically realistic TC phenomena. It does not exactly reproduce past events. However, since the simulations are initialised with, and driven by historical weather they generally do a good job of matching key observations such as storm track and intensity. For example the animation below shows how our simulation of Hurricane Katrina (2005) compares to the IBTrACS best track data.
Using InCyc we have developed an over-water TC wind-field model that captures important real-world behaviour such as how Typhoons and Hurricanes change shape as they move Northward and transition into extra-tropical storms2. This is a weakness of existing models.
We have also used InCyc to build the over-land component of our model. By doing this we are able to capture detailed land-surface interactions that other models struggle with. For example, the video below shows wind vectors from an InCyc simulation of Typhoon Jebi (2018) as it crosses over Osaka. It’s possible to see the effect from the mountainous topography surrounding the city and the effect of land-use as strong winds pass from Osaka Bay into the city.
There are a huge number of potential applications for InCyc, especially given it’s global coverage. For example it could be used to drive storm-surge models in the Ganges Delta, a region not well-serviced by the catastrophe modelling industry. Another use might be modelling or researching TC rainfall, especially given the recent flooding events such as Hurricanes Harvey (2017). If you have other ideas for how it could be used please do contact us, we are particularly interested in potential academic research collaborations.
“This work was supported by resources provided by the Pawsey Supercomputing Centre with funding from the Australian Government and the Government of Western Australia.”
- 1.Skamarock WC, Klemp JB, Dudhia J, et al. A Description of the Advanced Research WRF Model Version 4. UCAR/NCAR; 2019. doi:10.5065/1DFH-6P97
- 2.Loridan T, Crompton RP, Dubossarsky E. A Machine Learning Approach to Modeling Tropical Cyclone Wind Field Uncertainty. Mon Wea Rev. August 2017:3203-3221. doi:10.1175/mwr-d-16-0429.1