By Scott Hamilton

A group of researchers in Taiwan at the Meteorological Information Center (of Taiwan Weather Bureau) have been tasked with the development of new weather-related products and tools, discovered a new method of weather forecast models that will increase model performance by a factor of 200,000 times.

Weather forecasting tool development is a long, drawn out process that requires extensive testing to ensure accurate predictions. Much of the development effort is spent in testing and validation, rerunning past models on the new system and comparing the results with both the original predictions and the actual weather on the forecast days. It is important that the models are able to predict the same weather given the same inputs, unless the new model shows higher accuracy in prediction based on historical records.

There are a very limited number of tools available for weather forecast modeling and most of them are developed by the researchers to gather data and predict outcomes for the function the particular researcher is most interested in. This leads to the need for rapid development. The python programming language is used for most of this development due to the powerful public libraries that can be easily reused. Python is also easy to use for rapid development due to the simplicity of the language.

The motivation behind developing this new and faster model was the need for practical evaluation of the models. In order to validate a weather model, it requires running through decades of weather data using historical data to attempt to predict the next day, week and month of weather. This is compared to the actual weather and used to verify the models. Every new weather modeling technique has to go through the process of re-forecasting historical weather events before being trusted to predict the next weather event.

The team began to look at similarities between things like data collected on wind speed and direction in a region and geological topology maps of the region; this led to the discovery that many of the geological libraries of code could be reused to find weather front boundaries. These are highly accelerated processes using graphic processing units to detect features, like mountain ranges, and were easy to convert for detecting storm fronts. Utilizing the GPU based functions resulted in a 200,000 times faster detection of storm fronts from wind patterns. This was not the only place combining this knowledge of improved weather forecast models, but the performance improvements were astounding and helped to overcome the lack of computing speed necessary for rapid modeling during intense weather events.

For more information on the project, you can go to Until next week, stay safe and learn something new.

Scott Hamilton is a Senior Expert in Emerging Technologies at ATOS and can be reached with questions and comments via email to

Share via
Copy link
Powered by Social Snap