Reconstructing temperature fields from OH distribution and soot volume fraction in turbulent flames using an artificial neural network

We present a methodology based on artificial neural networks for reconstructing flame temperature fields from planar distributions of hydroxyl (OH) radicals and soot volume fraction in turbulent jet flames. A convolutional neural network was trained using planar images of flame temperature, soot volume fraction and OH simultaneously recorded with laser-based experimental methods. Then, the capacity and accuracy of the neural network on reconstructing flame temperatures were assessed for the flame conditions not only within the training domain but also out of it. The results showed that the supervised neural network can reconstruct instantaneous temperature fields to within ± 60 K for the flame conditions within the training domain, and to within ± 150 K for the flame conditions outside of the training domain. Probability density functions (PDF) of reconstructed temperature and the joint PDFs with OH signals and soot volume fractions also show good statistical agreement with experiments. This work has application in extending measurement techniques into regimes that are presently difficult to achieve, such as by obtaining the training data for temperature with established low-speed imaging methods and using the neural network method to predict temperature from high-speed imaging of the other two scalars.