par Zdybał, Kamila;Parente, Alessandro ;Sutherland, James C.
Référence 13th US National Combustion Meeting (College Station, Texas, US)
Publication Non publié, 2023-01-18
Communication à un colloque
Résumé : The first step in reduced-order modeling (ROM) workflows is finding a low-dimensional representation of a highly-dimensional system. The second step of ROM often requires training a nonlinear regression model to predict quantities of interest (QoIs) from the reduced representation. In the context of reacting flows, much of the research on training ROMs thus far has tackled those two steps separately. While they both come with their challenges, a good-quality low-dimensional system representation generally facilitates building a regression model. In this work, we leverage this inherent link between dimensionality reduction and nonlinear regression. We propose an approach where dimensionality reduction and nonlinear regression are considered jointly within an autoencoder-like neural network architecture. The dimensionality reduction (encoding) is affected by forcing accurate regression (decoding) of the QoIs. We show that such a joint architecture leads to improved low-dimensional representations as the two steps communicate with each other through backpropagation. In order to make quantitative assessments of projection quality, we use the recently proposed cost function that quantifies the quality of low-dimensional data representations from the perspective of non-uniqueness or steep gradients in QoIs. Using the cost function, we find that a nonlinear decoding promotes minimizing non-uniqueness in representing QoIs on a projection, as compared to a linear decoding (e.g., as is done in principal component analysis). We apply our regression-aware autoencoder to reacting flow datasets constructed from a steady laminar flamelet model for various fuels: hydrogen, syngas, methane, and ethylene. The original dimensionalities of the thermo-chemical state-space range from 9 to 53 and are efficiently reduced to 2 or 3. The relevant QoIs are the important state variables, such as temperature and major chemical species, and highly nonlinear projected source terms required by the reduced model. We demonstrate that the proposed approach can serve as an effective replacement of standalone dimensionality reduction techniques whenever nonlinear regression is anticipated in the downstream use.