Machine learning-based design and techno-economic assessments of adsorption processes for CO2 capture

  • Author / Creator
    Subraveti, Sai Gokul
  • Cyclic adsorption processes are widely considered for various industrial gas separations, including CO2 capture. The flexibility to configure a variety of process cycles is an attractive process design feature of these processes. Despite such flexibility for process design, computationally expensive and time-consuming mathematical models used to simulate and optimize cyclic adsorption processes often limit the design to few cycle configurations. Moreover, the potential of adsorption processes for CO2 capture is often poorly understood due to the lack of reliable techno-economic assessments. This thesis focuses on developing models of varying complexity to advance the understanding of adsorption processes for CO2 capture. In the first part of the thesis, viable machine learning models are developed to simulate and optimize pressure swing adsorption (PSA) and vacuum swing adsorption (VSA) processes for CO2 capture. To this end, two hybrid optimization approaches that incorporate techniques such as artificial neural networks, partial least squares regression are proposed to accelerate the computational speeds of multi-objective optimization for a fixed PSA cycle ten times. Next, physics-based deep neural network methodology is developed to synthesize and simulate vacuum swing adsorption (VSA) processes. As a first step, a simple chromatography system is considered where the neural network model is developed to simulate the spatiotemporal dynamics of generic pulse injections in chromatography columns. In neural network training, residuals of governing partial differential equations are incorporated into the loss function. As a result, the learning process required only small amounts of training data due to the additional knowledge of physics. The results showed that the neural network predictions of column dynamics for an arbitrary pulse injection were remarkably accurate. Following this, the framework is extended to synthesize and simulate VSA cycles. Here, individual neural networks are trained to learn the spatiotemporal dynamics of each constituent step. For training the models, conservation laws of mass and momentum are incorporated into the loss function. The results demonstrated that neural networks were capable of synthesizing and simulating four different VSA cycles. The second part of the thesis focuses on developing a rigorous techno-economic optimization model for the systematic design of PVSA processes for CO2 capture. The methodology incorporated a detailed process model, vacuum pump dynamics, rational scale-up, and cost model consistent with best practices, combined with a stochastic optimization routine to optimize process variables for determining the minimum CO2 avoided cost. This methodology was first applied to post-combustion CO2 capture from steam methane reformer flue gas by considering a four-step VSA process and three different adsorbents: Zeolite 13X and metal-organic frameworks, UTSA-16 and IISERP MOF2. The results showed that the four-step VSA process with IISERP MOF2 performed the best among the adsorbents considered; however, it still obtained 10% higher CO2 avoided cost compared to the baseline monoethanolamine (MEA) based absorption process. Finally, the techno-economic optimization methodology is extended to optimize both adsorbent and process variables to determine the lowest possible CO2 avoided costs of two PVSA cycles, namely, four-step and six-step dual reflux cycles. The techno-economic investigation is carried out at different flue gas flow rates and CO2 compositions to identify the potential of adsorption processes for post-combustion CO2 capture. Compared to MEA based absorption process, PVSA is attractive for flue gas streams with high CO2 compositions ≥7.5%. The ideal adsorbents needed to achieve the cost limits have fairly linear CO2 adsorption isotherms and zero N2 adsorption.

  • 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.