(629g) Pressure Swing Adsorption Cycle Synthesis Utilizing Artificial Neural Networks As Surrogate Models | AIChE

(629g) Pressure Swing Adsorption Cycle Synthesis Utilizing Artificial Neural Networks As Surrogate Models


Leperi, K. - Presenter, Northwestern University
Yancy-Caballero, D., Northwestern University
Snurr, R., Northwestern University
You, F., Cornell University
With fossil fuels expected to be a significant portion of the world’s energy mix for the near future, it is important to minimize CO2 emissions from current power plants through carbon capture and sequestration (CCS).1 In post-combustion CCS, CO2 is separated from the power plant flue gas emissions, containing mainly of N2 and CO2. Of the technologies currently available for CCS, pressure swing adsorption (PSA) is one of the more promising due to low energy requirements and short cycle times compared to other adsorption based technologies. However, one challenge that exists in using PSA for CCS is designing the cycle to match newly developed materials for this application. Although the steps in all PSA cycles can be classified into six different possibilities (pressurization, feed, depressurization, light reflux, heavy reflux and pressure equilibration), the arrangement of the steps and interactions between steps lead to hundreds of potential different combinations. The objective of this work is to develop a new approach that is capable of synthesizing PSA cycles to capture the CO2 from flue gas at the required purities and recoveries while minimizing energy requirements and maximizing adsorbent productivities.

In this work, we present a new framework for synthesizing the PSA cycle with the lowest predicted CO2 capture costs. In this framework, we train artificial neural networks (ANNs) using Bayesian regularization methods as surrogate models for the various PSA steps. The ANNs are trained on simulation data collected from our PSA model consisting of a system of partial differential algebraic equations incorporating mass and energy balances, pressure drop across the column, competitive multi-site Langmuir isotherms and the linear driving force model.2 With the ANN surrogate models, we propose a mixed integer nonlinear programming (MINLP) model to determine the ideal ordering and duration of steps in order to minimize the energy requirements and maximize the adsorbent productivity. We evaluate this model with several adsorbents, including Ni-MOF-74, UTSA-16 and zeolite 13X to compare the adsorbents under optimized cycle conditions.

  1. MIT. The Future of Coal; 2007. http://web.mit.edu/coal/.
  2. Leperi KT, Snurr RQ, You F. Optimization of Two-Stage Pressure/Vacuum Swing Adsorption with Variable Dehydration Level for Postcombustion Carbon Capture. Ind. Eng. Chem. Res. 2016;55:3338-3350.