Deep neural network inversion for 3D laser absorption imaging of methane in reacting flows

Abstract

Mid-infrared laser absorption imaging of methane in flames is performed with a learning-based approach to the limited view-angle inversion problem. A deep neural network is trained with superimposed Gaussian field distributions of spectral absorption coefficients, and the prediction capability is compared to linear tomography methods at a varying number of view angles for simulated fields representative of a flame pair. Experimental 3D imaging is demonstrated on a methane–oxygen laminar flame doublet ( &lt;<!-- cm ) backlit with tunable radiation from an interband cascade laser near 3.16 µm. Spectrally resolved data at each pixel provide for species-specific projected absorbance. 2D images were collected at six projection angles on a high-speed infrared camera, yielding an aggregate of 27,648 unique lines of sight capturing the scene with a pixel resolution of ∼ 70 µm. Mole fraction measurements are inferred from the predicted absorption coefficient images using an estimated temperature field, showing consistency with expected values from reactant flow rates. To the authors’ knowledge, this work represents the first 3D imaging of methane in a reacting flow.

CSTNet: A Dual-Branch Convolutional Neural Network for Imaging of Reactive Flows Using Chemical Species Tomography

“…In this work, we adopt X min = 0.01, X max = 0.12, T min = 318 K, and T max = 1300 K. The dataset is generated with 19 Subsequently, path integrated absorbance for the j th beam in the training and validation sets is standardized. The process of standardization has two benefits.…”

“…Previous work employed CNNs in CST and showed that their models could achieve a similar accuracy level as simulated annealing [17] and the reduction in network parameters [18]. In addition, CNN has been applied in a proofof-concept experiment to reconstruct the 3-D distribution of methane concentration using mid-infrared CST [19]. Although these endeavors are promising for the industrial application of CNN in CST, the following three issues remain to be addressed as a matter of urgency.…”

Flow Field Tomography with Uncertainty Quantification using a Bayesian Physics-Informed Neural Network

“…Convolutional neural nets (CNNs) have been particularly successful in the realm of image processing and were therefore adapted for tomographic reconstruction by numerous groups. For instance, Huang et al [27], [28], [29] conducted 2D laser absorption tomography and 3D chemiluminescence tomography via CNNs, and Wei et al [30], [31] used CNNs to reconstruct 3D mole fraction and temperature fields from laser absorption measurements of methane and ethylene flame doublets. These examples employed the common supervised training paradigm, in which phantom distributions (c exact ) paired with synthetic measurements (b exact = Ac exact ) are used to train the network.…”

“…These examples employed the common supervised training paradigm, in which phantom distributions (c exact ) paired with synthetic measurements (b exact = Ac exact ) are used to train the network. Phantoms for training have been obtained from previous reconstructions [28], [29], random Gaussian fields [27], [30], and large-eddy simulations [31], and random errors were usually added to b train to make the reconstruction procedure robust to noise. While this approach can yield accurate estimates of the QoI when the training set accurately represents the target physics, the application of DNN-based reconstruction to targets that exhibit unique structures is not reliable.…”

Weight Encode Reconstruction Network for Computed Tomography in a Semi-Case-Wise and Learning-Based Way

“…Deep learning, as its extraordinary capability of feature abstracting and non-linear fitting 13 , can potentially deal with the issues mentioned above. Some attempts have been made to adopt deep learning algorithms in some reconstruction tasks like the sliced absorption spectroscopy, and chemiluminescence reconstruction [14][15][16] . However, such end-to-end methods (using the projections as the input and reconstructed object as the outputs) potentially deduct the generalization of the model.…”