Compressing convolutional neural networks using POD for the reconstruction of nonlinear tomographic absorption spectroscopy

Abstract

Nonlinear tomographic absorption spectroscopy (NTAS) is an imaging technique that inherits the advantages of both absorption spectroscopy and tomography such as species specificity, quantitative measurements, high sensitivity and capability in resolving non-uniformities of the target flow field. It has found many applications such as engine measurements where the flow field is highly turbulent. However, the solution of the NTAS problem is non-trivial as a set of nonlinear equations need to be solved. The inversion problem is typically converted into an optimization problem and solved with the simulated annealing algorithm which is criticized for its low computational efficiency. Convolutional neural networks havebeen successfully introduced to speed up the reconstructions of NTAS. However, due to the large amount of training data, CNN has a high demand for memory to store the network parameters, which also requires a large amount of time to be optimized during the learning process. In this work, we proposed a scheme to compress the output layer via dimension reduction using proper orthogonal decomposition. The simulative study performed here demonstrated that the number of network parameters and the training time can be reduced dramatically by a factor of ∼100 and ∼10 respectively. This improvement makes NTAS a more preferable technique for combustion diagnostics for which time and spatially resolved measurements are highly desired.