Leveraging spectroscopic sensor measurements for development of models for reactions involving complex feedstocks

• Author / Creator

  Srinivasan, Karthik

• Chemically heterogeneous feedstocks are being increasingly used in process industries due to depletion of conventional feedstocks, to meet environmental demands and to recover
  value added products from wastes. Chemical modeling of reactive transformations of such complex feedstocks involves tracking the trajectories of multiple reactive species, typically
  through spectroscopic sensor measurements, to obtain atom level knowledge of the reacting species and can be challenging, especially without any human insight. Interpretation of spectroscopic signatures is an art and traditionally demands a level of domain expertise. Reaction models developed using sensor measurements also require domain expertise and
  are typically generated by suggesting model compounds for groups of substrates. Including human insight, however, leads to bias in modeling and does not allow for efficient exploration
  of the chemical space for all possible reactions in the system. Furthermore, updating these expert-guided models based on new operational data is quite cumbersome. This
  thesis aims to explore the usage of spectroscopic sensor measurements for automation of reaction and kinetic modeling of complex reaction systems by employing machine learning
  and chemometric methods. The methodologies developed are presented on hydrothermal liquefaction (HTL) of biomass as a case study by utilizing experimental Fourier Transform
  Infrared (FTIR) and Proton Nuclear Magnetic Resonance (1H-NMR) spectroscopy.
  Different methodologies for the identification of reaction networks from spectroscopic data are presented in decreasing order of human intervention required. Spectroscopic curve
  resolution techniques have been employed at different degrees of sensor data fusion to obtain interpretable and structurally consistent latent features of the reaction system. Signal level data fusion has been performed through Self Modeling Curve Resolution, while a higher order Joint Non-Negative Tensorial Factorization scheme has been applied at a contextual level to jointly analyze FTIR and 1H-NMR spectroscopic data. Expert knowledge has been used in determination of reactive compounds and the subsequent reaction networks. In a step towards automation, extraction of functional group signatures of the reactive species has been performed through application of convolutional operations on the resolved FTIR spectrum and partial molecular fingerprints for each reactive species have
  been identified. A reaction network identification methodology that maps spectroscopic signatures to candidate molecules is presented. The network generation is constrained based
  on the causal structure inferred using Bayesian structure learning and domain knowledge, and employs algorithmically extracted reaction rules obtained through Atom-Atom Mapping
  (AAM) of reactions from a database. A one-shot molecular generation methodology is presented as the next step in automation thus subverting the need for Bayesian structure
  learning and spectroscopic deconvolution. Employing a graph neural network based hetero-autoencoder and generative adversarial networks, the molecular generation routine
  generates molecules constrained by the FTIR spectrum. Localized reaction networks for the process are identified by recursive application of reaction templates. The reaction networks identified have been found to be concordant with reactive transformations recorded in the literature for HTL of biomass. Mathematical modeling of the kinetics of the system
  based on temporal projection of latent features of the spectroscopic deconvolution has been performed by employing chemical reaction neural networks constrained based on the
  adjacency information obtained via Bayesian structure learning of the resolved spectrum. Benchmarking studies comparing these neural Ordinary Differential Equations with constrained
  alternating least squares and basis reduction techniques such as SIND-y is also presented for a synthetic system.

• Subjects / Keywords

  • Reaction networks
  • Machine learning
  • Generative modeling
  • Chemoinformatic

• Graduation date

  Fall 2024

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-x681-ns89

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.