Machine Learning Based Data Compression in High-Fidelity Numerical Simulation

NE has invested heavily in the development of a wide range of computational frame-works over the past few decades. This considerable investment in advanced high fidelity codes has allowed the Nuclear engineering community to greatly reduce the non-recurring costs associated with reactor develop- ment. However, as the fidelity of these simulations grows, so too do the difficulties associated with using the simulations in industrial settings. Issues associated with writing, storing and analyzing the extreme levels of high-fidelity simulation data are not unique to NE. In fact, NASA, the DOE and CERN have all expressed the need for software and tools that can marshal high-fidelity simulation data into a format that can be quickly and efficiently used to rapidly inform real-world design decisions. RNET and OSU are proposing the development of robust framework for in-situ, scalable and data-driven compression in high-fidelity numerical simulation. This framework will allow simulations to output data in efficient reduced representations that occupy less memory, speed up input/output, and enable faster knowledge extraction. At the heart of the framework will be the aforementioned DLS data compression algorithm, a novel algorithm that uses modern machine learning based feature extraction techniques and the generalized finite element method (GFEM) to compress high-fidelity data into accurate reduced representations that may be queried in real-time. The three main objectives of the Phase I project are: 1)implement a perform-ant hybrid (MPI ,OpenMP) C++ implementation of the DLS algoirthm utilizing spatial domain decomposition and local multi- threading; 2) demonstrate the exceptional value of DLS data compression by comparing it other state-of-the-art lossy data compression algorithms for spatiotemporal solution fields; and 3) produce system design documents for the database and rapid analysis portal to be implemented in Phase II. For the developers of numerical software, the library will pro- vide a fast, in-situ method for exporting scientific data from advanced numerical simulations. The customers falling into this group include the many government agencies using numerical simulation including the DOE, NASA, Air Force, Navy, Army, and Marines. The commercial sector is also a heavy user of mesh based ap- plications. Those applications include the aeronautical, the biomechanical, and the automotive industries. The underlying algorithms of the proposed library could be used for data compression and/or fusion of large-scale experimental data sets, or multi-physics datasets such as those common in fracture analysis and space vehicle design, among many others.