Segmentation of turbulent computational fluid dynamics simulations with unsupervised ensemble learning

Computer vision and machine learning tools offer an exciting new way for automatically analyzing categorizing information from complex computer simulations. Here we design ensemble framework that can independently robustly categorize dissect simulation data output contents of turbulent flow patterns into distinct structure catalogs. The segmentation is performed using unsupervised clustering algorithm, which segments physical structures by grouping together similar pixels in images. accuracy robustness the resulting segment region boundaries are enhanced combining multiple simultaneously-evaluated operations. stacking object evaluations image mask combination This statistically-combined (SCE) different cluster masks allows us to construct reliability metrics each pixel associated without any prior user input. By comparing similarity occurrences ensemble, also assess optimal number clusters needed describe data. Furthermore, relying on ensemble-averaged spatial boundaries, SCE method enables reconstruction more accurate robust interest (ROI) clusters. We apply algorithm 2-dimensional snapshots magnetically-dominated fully-kinetic plasma flows where ROI geometrical measurements intermittent known as current sheets.