Application-specific compression of large MD data preserving physical characteristics

Application areas like physics or thermodynamics often require simulations of very large data sets, up to the order of 10
¹²
particles or even larger, to obtain results relevant for realistic industrial processes. Persisting such data is too costly, prohibiting interactive visual analysis in a classical post-processing fashion. Thus, analysis is restricted to statistical aggregation or visual in-situ exploration, both requiring an inkling of the results beforehand. We alleviate this issue by applying an application-optimized lossy compression. Reducing the size while at the same time preserving relevant physical characteristics of the data allows for accessibility on workstations and practical long-term storage. The compression is achieved by generating a density volume that is processed using wavelet decomposition, quantization and run-length encoding. Our reconstruction of particle data ensures the restoration of physically relevant properties. It employs a model based on stochastic distributions complemented by further adjustments. We evaluate the precision of the reconstruction for several data sets and a wide range of compression variants to show the effectiveness and user-adjustable trade-offs of the presented method.