Filling the Void: Deep Learning-based Reconstruction of Sampled Spatiotemporal Scientific Simulation Data

As high-performance computing systems continue to advance, thegap between computing performance and I/O capabilities is widen-ing. This bottleneck limits the storage capabilities of increasinglylarge-scale simulations, which generate data at never-before-seengranularities while only being able to store a small subset of the rawdata. Recently, strategies for data-driven sampling have been pro-posed to intelligently sample the data in a way that achieves high datareduction rates while preserving important regions or features withhigh fidelity. However, a thorough analysis of how such intelligentsamples can be used for data reconstruction is lacking. We proposean AI-driven approach based on training neural networks to recon-struct full-scale datasets based on a simulation’s sampled output.Compared to current state-of-the-art reconstruction approaches suchas Delaunay triangulation, we demonstrate that deep learning-basedreconstruction has several advantages, including reconstruction qual-ity and time-to-reconstruct. We propose and evaluate strategies thatbalance the sampling rates with model training and data reconstruc-tion time to demonstrate how such AI approaches can be tailored forboth speed and quality, and develop a visual analytics interface forcomparing the reconstruction quality of grid-based datasets