(426e) Dimensionality Reduction of Chemical Kinetics Based on Extent-of-Reaction in a Physics-Inspired Machine Learning Framework

The accurate description of chemical reaction mechanisms and the estimation of their associated kinetic parameters is of paramount importance for the design and optimization of chemical processes. This work extends the application of physics-inspired machine learning techniques to inverse problems in chemical kinetics, as previously proposed by kinetics-informed neural networks (KINNs) in the context of initial-value problems (IVP). We show that singular value decomposition (SVD) provides a basis for the solution of inverse problems in uncorrelated subspaces, mitigating the issue of ill-posed covariances for the inverse problem formulated as maximum-likelihood estimators (MLE). Furthermore, we introduce the conversion of the original KINNs formalism to extents-of-reaction, which enables the problem of finding states that satisfy underlying governing equations to be solved separately from that of determining physically consistent rate equations. Such a representation in conjunction with SVD allows for more efficient determination of rate equations given sparse chemical data for systems described by boundary-value problems (BVP), e.g., differential packed-bed reactors. These approaches combined pave ways for the construction of a more general machine-learning based framework for the solution of chemical kinetics inverse IVPs and BVPs