Intra-frame Compression of Molecular Dynamics Simulations of Water
Keegan Carruthers-Smith, Patrick Marias, James Edward Gain · 2009
Molecular dynamics (MD) simulations generate vast amounts of data. A typical 100-million atom MD simulation produces approximately 5 gigabytes of data per frame consisting of atom types, coordinates and velocities. The main contribution of the report is the specification of an MD compressor which targets simulations with large amounts of water in it. Water is targeted since many MD simulations contain significant amounts of water molecules. It uses an approach based on predictive point cloud compressors, but with predictors tailored towards models of water. A point cloud is a collection of points in 3D space. The method improves on existing point cloud compressors [Gumhold et al., 2005, Devillers and Gandoin, 2000] and MD compressor [Omeltchenko et al., 2000] when applied to MD data with significant amounts of water molecules. There are six MD simulations the compression schemes are tested and compared against. Our water compression scheme performs best in general. At 8-bit quantisation the mean and average compression rate for water compression is 17.7%. The next best compression rate was gzip with a mean and average compression rate of 19.5% and 18.5% respectively. At 12-bit quantisation the mean and average compression rate for water compression is 29.8%. The next best scheme is gzip which has a mean and average compression rate of 37% and 36.5% respectively. The report also presents techniques for compressing permutations. It presents a technique which has a worst case performance that is approximately the lower bound on the performance of a general permutation encoder. It also presents a technique which probabilistically performs better than the lower bound. When permutation information is discarded, our water compression scheme still performs best.