SIParCS 2022 - Anil Alper
Parallel Algorithms to Recognize Spatial Patterns in Climate Analysis
We want to predict atmospheric conditions accurately to adapt to the local effects of climate change. The global atmospheric models are developed to simulate weather conditions across the world but they are used on a very large scale making them less useful for more local scale environments. Many downscaling methods are developed to better understand the local scale changes caused by climate change. Intermediate Complexity Atmospheric Research Model (ICAR) is among such models that downscales the global model data using physical and statistical processes. However, because of the simplifications, the downscaled data are systematically biased compared to observations. There are various statistical methods and complex algorithms to bias correct models like ICAR. Due to the complexity of such bias correction methods, we have to use high performance computing (i.e. parallel algorithms). In this project, we used one of many bias correction methods with a runtime of 3 hours without the use of high performance computing. After the GPU is introduced to the problem, a significant performance improvement is observed and the bias-corrected ICAR more closely matches the observed precipitation patterns. Future work can use the GPU approach for more computationally expensive bias correction methods that might be impossible to implement sequentially.
Mentors: Ethan Gutmann, Rachel McCrary
Slides and poster