This thesis improves error-bounded lossy compression of scientific data by replacing general numerical prediction with methods that use the structure of the data.

Overview

Scientific simulations and observations now generate data at volumes that make storage, transfer, and analysis the bottleneck, and lossless compression is ineffective on it. Error-bounded lossy compressors answer this, but the established ones are data agnostic and rely on general numerical approximation, which caps the ratios they reach under strict error bounds. Three compressors are proposed: GraphCom, which models a grid as a graph; LLMComp, which treats compression as autoregressive token modeling; and LSDComp, which adds spatiotemporal structure awareness, reaching up to 47 percent higher compression ratios than State of the art methods under various error bounds. A final study finds that repurposing existing deep learning weather forecasters as predictors does not raise dataset-level compression ratio, because residual structure, not prediction accuracy, leads entropy-coding efficiency.

Presenters

Brief Biography

Muhannad Alhumaidi is a Ph.D. student in Computer Science at King Abdullah University of Science and Technology (KAUST), advised by Prof. Panos Kalnis. He holds a B.S. in Computer Science from Purdue University in 2020 and an M.S. in Computer Science from KAUST in 2022. His research has been published in IEEE Transactions on Knowledge and Data Engineering and at IEEE BigData.