Impacts of first- and second-order optimization in deep learning of turbulent fluctuations from gas puff imaging

Abstract

Following the results of [1], which demonstrates a novel method to translate 2-dimensional measurements of HeI line radiation on turbulent scale into local plasma fluctuations via an integrated deep learning framework, this manuscript investigates the results when applying two separate techniques for optimization: Adam and L-BFGS. Fundamentally, the two approaches apply the same set of constraints and loss functions that combine neutral transport physics and collisional radiative theory for the 33D − 23P (587.6 nm line) transition in atomic helium whilst training the networks. The impact of these first- and second-order optimization techniques are investigated to examine their influence on numerical convergence and stability when seeking to analyze turbulent dynamics via gas puff imaging in experimental plasmas.