TY - JOUR
T1 - 110th Anniversary
T2 - Surrogate Models Based on Artificial Neural Networks to Simulate and Optimize Pressure Swing Adsorption Cycles for CO2 Capture
AU - Leperi, Karson T.
AU - Yancy-Caballero, Daison
AU - Snurr, Randall Q.
AU - You, Fengqi
N1 - Funding Information:
This work was financially supported by the National Science Foundation (CBET-1604890).
Publisher Copyright:
© 2019 American Chemical Society.
PY - 2019/10/2
Y1 - 2019/10/2
N2 - Carbon capture technologies are expected to play a key role in the global energy system, as it is likely that fossil fuels will continue to be dominant in the world's energy mix in the near future. Pressure swing adsorption (PSA) is a promising alternative among currently available technologies for carbon capture due to its low energy requirements. Still, the design of the appropriate PSA cycle for a given adsorbent material is a challenge that must be addressed to make PSA commercially competitive for carbon capture applications. In this work, we propose and test a model reduction-based approach that systematically generates low-order representations of rigorous PSA models. These reduced-order models are obtained by training artificial neural networks on data collected from full partial differential algebraic equation (PDAE) model simulations. The main contribution of this paper is the development of surrogate models for every possible step in PSA cycles: pressurization, adsorption, and depressurization steps in cocurrent and counter-current operation. Three different PSA cycles (three-step, Skarstrom, and five-step cycle) for postcombustion carbon capture applications were employed for training purposes, and two adsorbents, Ni-MOF-74 and zeolite 13X, were chosen to evaluate the surrogate models under optimized cycle conditions. A good agreement was observed between the results of the ANN models and the PDAE simulations. Average mean square errors, for dimensionless state variables, of 1.7 × 10-8, 5.8 × 10-8, and 9.9 × 10-7 were obtained for the three PSA cycles analyzed in this work, and the highest relative error, regarding the CO2 purity and recovery, was 1.42%. These results suggest that the use of machine learning techniques to develop PSA surrogate models is feasible and that these models can be implemented in optimization environments to synthesize PSA cycles.
AB - Carbon capture technologies are expected to play a key role in the global energy system, as it is likely that fossil fuels will continue to be dominant in the world's energy mix in the near future. Pressure swing adsorption (PSA) is a promising alternative among currently available technologies for carbon capture due to its low energy requirements. Still, the design of the appropriate PSA cycle for a given adsorbent material is a challenge that must be addressed to make PSA commercially competitive for carbon capture applications. In this work, we propose and test a model reduction-based approach that systematically generates low-order representations of rigorous PSA models. These reduced-order models are obtained by training artificial neural networks on data collected from full partial differential algebraic equation (PDAE) model simulations. The main contribution of this paper is the development of surrogate models for every possible step in PSA cycles: pressurization, adsorption, and depressurization steps in cocurrent and counter-current operation. Three different PSA cycles (three-step, Skarstrom, and five-step cycle) for postcombustion carbon capture applications were employed for training purposes, and two adsorbents, Ni-MOF-74 and zeolite 13X, were chosen to evaluate the surrogate models under optimized cycle conditions. A good agreement was observed between the results of the ANN models and the PDAE simulations. Average mean square errors, for dimensionless state variables, of 1.7 × 10-8, 5.8 × 10-8, and 9.9 × 10-7 were obtained for the three PSA cycles analyzed in this work, and the highest relative error, regarding the CO2 purity and recovery, was 1.42%. These results suggest that the use of machine learning techniques to develop PSA surrogate models is feasible and that these models can be implemented in optimization environments to synthesize PSA cycles.
UR - http://www.scopus.com/inward/record.url?scp=85072891972&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85072891972&partnerID=8YFLogxK
U2 - 10.1021/acs.iecr.9b02383
DO - 10.1021/acs.iecr.9b02383
M3 - Article
AN - SCOPUS:85072891972
SN - 0888-5885
VL - 58
SP - 18241
EP - 18252
JO - Industrial and Engineering Chemistry Research
JF - Industrial and Engineering Chemistry Research
IS - 39
ER -