Machine learning-based monitoring of complex reactive systems

• Author / Creator

  Puliyanda, Anjana Thimmaiah

• Processing of complex feedstocks for the production of value-added chemicals and fuels is industrially important. The lack of a priori knowledge of the innumerable species and the reaction pathways governing their conversion, has posed challenges to monitoring these processes. Although, data-driven models have been used, their lack of interpretability and an end-to-end modeling framework has limited the efficiency of diagnostic decisions in process monitoring. On the other hand, systems where the mechanistic knowledge of the species and their reactions are arrived at from first-principles simulations, face computational challenges in the deployment of such models for the process design. This thesis focuses on the following two aspects: (i) developing inferential machine learning models to enhance the interpretability of data-driven models, and (ii) developing predictive machine learning models to reduce the computational cost of first-principles simulations, in modeling chemical systems.

  The first aspect of developing inferential machine learning models focuses on the identification of species, reaction pathways, and kinetic parameter estimation from spectroscopic data of the system, with application to the visbreaking of bitumen. Spectroscopic curve resolution methods that are structure-preserving, interpretable, and jointly parse data from multiple sensors, to extract latent features for species identification have been presented with an increasing degree of sophistication as follows:(i) self-modeling multivariate curve resolution (SMCR), (ii) joint non-negative matrix factorization (JNMF) as a data fusion analogue of SMCR where regularization constraints act like chemical information sieves to handle complementary, orthogonal and redundant features in the latent factorization of multi-sensor data and (iii) joint non-negative tensor factorization (JNTF) as a structure-preserving higher order analogue of JNMF. Next, Bayesian structure learning among the extracted spectral features has been used to causally infer plausible reaction pathways that have been validated by domain knowledge. Finally, the latent factorization and causal inference models have been used as an engine to interpret the modes identified by training hidden semi-Markov models on spectra. This captures the time scales and dynamics of reaction mechanisms with changing temperatures, for the realtime monitoring of reactive systems purely from spectroscopic data. Projections of spectroscopic data onto the temporal mode of data collection via latent factorization, are interpreted as concentrations. Kinetic models constrained by physical laws and the reaction adjacency matrix deduced from the Bayesian network structure are implemented using chemical neural ODEs trained on the temporal concentrations. The prediction accuracy is seen to depend on the ability of latent factorization to handle process noise.

  The second aspect of training predictive machine learning models, focuses on not only reducing the computational cost of the ab initio molecular dynamics (AIMD) simulations of chemical systems, but also the cost in itself of developing such models. This has been demonstrated with application to the transglycosylation of cellobiose, to assess whether or not the solvent molecules reorganize significantly in going from the reactant to the product configurations. A self-supervised 3D convolutional neural network autoencoder is trained to extract features from the reactant and product simulation trajectories, the probability distributions across the difference between which is used to assess if the solvent reorganization is significant. Cellobiose systems at lower temperatures are found to reorganize to a greater extent than those at higher temperatures, consistent with the decrease in the activation free energy barrier as temperature increases. Similarity between the reactant configuration features of other chemical systems with those extracted from that of the cellobiose systems, is then used as a basis to inform the extent of reorganization in the product profiles, without having to explicitly run AIMD simulations for the same.

• Subjects / Keywords

  • Cheminformatics
  • Reaction network hypotheses
  • Reaction pathway identification
  • Constrained latent factorization
  • Spectroscopic data fusion
  • Bayesian networks
  • Reaction monitoring
  • Hidden semi-Markov models
  • Chemical reaction neural ODEs
  • Self-supervised solvent screening
  • Molecular dynamics simulation manifolds
  • Bitumen visbreaking
  • Cellobiose transglycosylation
  • Acid-catalyzed fructose conversion
  • Data-driven modeling
  • Machine learning

• Graduation date

  Spring 2022

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-j852-9b81

• License

  This thesis is made available by the University of Alberta Libraries 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.