Limited-projection volumetric tomography for time-resolved turbulent combustion diagnostics via deep learning

摘要

Time-resolved volumetric tomography (VT) has been applied extensively for turbulent flow/combustion diagnostics, due to its great capacity in reconstructing three dimensional scalar/vector fields. However, it usually suffers from high computational costs of conventional iterative methods in processing thousands of tomographic frames, and also the requirement of multiple high-speed camera/intensifiers to ensure sufficient spatial sampling, resulting in high experimental costs. In this work, we aim to take the full advantage of the recent progress in deep learning algorithms and develop an inversion method which not only reduces the processing time of a single frame down to the milliseconds level but also the number of projections required without sacrificing the imaging quality. Two distinct frameworks of convolutional neural network were designed and tested for VT reconstructions of turbulent flames for the first time. The results from proof-of-concept experiments implementing computed tomography of chemiluminescence (CTC) confirmed the feasibility of our method. Our data-driven approach can expedite the reconstruction process by a factor of similar to 10(5) compared with conventional iterative methods (e.g., algebraic reconstruction technique). This work is expected to be valuable for all tomographic modalities which are seeking expedited reconstruction and reduced costs. (C) 2020 Elsevier Masson SAS. All rights reserved.