Retrieving soot volume fraction and temperature from broadband flame emission measurements via artificial-neural-network-based Abel inversion

This paper presents a methodology for combustion diagnostics that integrates artificial neural networks with the Abel inversion technique to retrieve soot temperature and volume fraction in laminar axisymmetric flames from emission signals captured using off-the-shelf RGB color cameras. The proposed approach minimizes the difference between signals captured by the camera with specific combinations of soot temperature and volume fraction, as described by the line-of-sight and spectrally integrated radiative transfer equation. The neural network approach is first tested on synthetic datasets generated from detailed numerical simulations of coflow flames with known local properties, along with their corresponding RGB emission fields. Then, the network is used to process experimental color images of similar coflow flames, providing direct predictions of the temperature and soot distribution without requiring an explicit Abel inversion or training data. Results are validated against well-established diagnostic techniques, including traditional broadband emission methods and line-of-sight attenuation measurements. The proposed neural-network-based approach produces accurate and smoother estimations of soot temperature fields compared to traditional broadband emission diagnostics, reducing measurement noise and artifacts. Soot volume fraction estimates using broadband emission signals also demonstrate strong agreement with well-established but more complex techniques that rely on additional instrumentation for measuring light attenuation. Overall, the proposed technique offers an accurate non-intrusive method for soot characterization using only an RGB camera with a known spectral response. The simplicity and efficiency of this approach paves the way for its integration into compact devices for real-time soot monitoring, which could be highly beneficial in industrial combustion applications. • Neural network retrieves soot fields from RGB flame images with high accuracy. • Method avoids explicit Abel inversion and light extinction measurements. • Soot temperature and volume fraction are estimated directly from RGB image channels. • Technique validated using both simulated and experimental coflow flames. • RGB-based method enables low-cost, non-intrusive combustion monitoring.