Princeton University Thunch Talk

Computer vision and machine learning tools offer an exciting new way to automate analyzing and categorizing information from complex computer simulations. In my talk, I will discuss our current efforts in designing new computer vision tools for the automatic segmentation of large, high-resolution plasma turbulence simulations. In particular, there is a great interest in the astrophysical community to identify and characterize so-called current sheets -- 2D sheet-like structures of intense current flow near accreting black holes. The formation of sheets and subsequent magnetic reconnections due to the tearing of these sheets contribute to the local heating in plasma and serve as a mechanism for non-thermal particle acceleration. Our current sheet identification framework is based on a self-organizing map (SOM) algorithm. SOMs are an unsupervised machine-learning technique that is particularly efficient in clustering and classifying high-dimensional big data. In my talk, I will describe in detail our new implementation of a JIT-accelerated SOM algorithm and usage of the SOM method to efficiently cluster plasma simulations.