A machine learning approach assisting soot radiation-based thermometry to recover complete flame temperature field in a laminar flame

Abstract

Soot radiation-based thermometry is a popular approach for flame temperature diagnostics. However, it usually suffers from flame temperature information lost in the regions of soot absent or less significant area. The originality of the present work lies in the development of a two-step Multi-Layer Perceptron (MLP) neural network method to assist soot radiation-based thermometry for flame temperature field retrieval. It completes the whole temperature field in a steady axis-symmetric Santoro laminar flame, which was originally measured by the Modulated Absorption Emission technique (MAE). Using temperature fields provided by numerical simulation of this standard Santoro flame, the two-step Multi-Layer Perceptron neural network model is trained and tested, and the temperature recovery sensitivity to the noises is theoretically investigated. Moreover, this two-step MLP approach for flame temperature completion is further proofed by experimental temperature results. And the two-step MLP prediction uncertainty is estimated as ± 85.5 K experimentally. The developed approach could assist other soot radiation-based thermometry, i.e., spectral soot emission, to provide the complete flame temperature fields.