Unified Machine learning based design of adsorption separation processes

  • Author / Creator
    Kasturi Nagesh Pai
  • Pressure swing adsorption (PSA) processes are an industrially mature low energy consumption pathway for gas separations. Due to their performance being linked to the separation media, they provide an additional degree of freedom for process design. They are difficult to accurately model due to the propagation of sharp heat and mass transfer fronts. These are unsteady-state processes and, as such, have to be run cyclically till a cyclic steady state to measure performance. The highly multivariate nature of the process inputs also makes them hard to optimize and control. With the advent of metal-organic chemistry and the almost infinite number of possible adsorbents, and the possibility of tailor-made adsorbents already being highlighted in the literature, there is a need for fast and accurate means there is a need for fast and accurate models. Industrial gas separations such as CO2 capture and O2 concentration can significantly benefit from finding the suitable adsorbent in an optimized process. The traditional process design and optimization frameworks require considerable computation resources, and thus primary screening of adsorbents is performed using simplified metrics. There is no explicit agreement over the link between process performance and separation media.
    In this thesis, the modelling, simulation, optimization, screening, and experimental validation of pressure swing adsorption processes are investigated using machine learning. The CoRE database of adsorbents is screened in a multi-scale framework for post-combustion CO2 capture. The GCMC simulations are used to predict CO2 and N2 equilibria. Using this data, a detailed process model was used to evaluate over 1500 adsorbents, and the results showed no statistical correlation to common screening metrics. A machine learning model generated using a decision tree algorithm was also presented to screen adsorbents for CO2 capture. Various machine learning algorithms are investigated for their ability to accelerate the optimization of PSA processes. Two optimization frameworks, Surrogate Opt. and CSS Opt., were presented to speed up PSA optimization. A neural network model was trained to learn from the final cyclic steady-state profiles, and the detailed model was used to initialize at CSS. The models were validated using lab-scale experimental data.
    A general adsorbent agnostic machine-assisted process learner and emulator (MAPLE) was developed to simulate the detailed process model. The unique aspect of this model is that adsorbent-specific parameters were inputs to the model along with process inputs. This means that once trained; the model can predict the performance of any type 1 adsorbent for a given process. The trained model had high accuracy with R2 ADJ ≥ 0.99 for all the outputs, such as CO2 purity, recovery. The modelling and optimization framework (MAPLE Opt.) was validated for a CO2 capture case study using data from the scholastic literature. The question of process performance limits for PVSA based CO2 capture was analyzed using the MAPLE model. Various optimization case studies investigated performance limits of real-world as well as hypothetical best adsorbents at different feed compositions. The showed that the innovation gap between real-world materials and the hypothetical best was very conservative. It was also shown that significant energy saving is possible at higher CO2 feed compositions using PSAs. The experimental validation of the MAPLE model was performed using an O2 concentration case study. The adsorbent agnostic MAPLE model was used to optimize the performance of LiX and 13X of a Skarstrom cycle. The results from the optimization were used to operate a bench-scale 2-bed lab scale rig to verify the performance. The results show that a model trained with hypothetical adsorbent equilibria can target performance in an experiment. This experimentally validated machine learning framework provides an alternative fast modelling pathway for PVSA process design and optimization.

  • Subjects / Keywords
  • Graduation date
    Fall 2021
  • Type of Item
  • Degree
    Doctor of Philosophy
  • DOI
  • 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.