(63c) A Simultaneous Material Screening and Process Optimization Approach for Carbon Capture Using Machine Learning

Recent advances in High-fidelity (HF) simulation-based optimization have led to the development of many algorithms that can handle high-dimensional spaces with mixed-integer inputs. However, there are challenges that remain when the inputs consist of both continuous and discrete variables and the sampling costs are limited. As an alternative, surrogate models, or Machine Learning (ML) models are employed to model HF simulation data and expedite the optimization search. Due to the complexity associated with dynamic HF simulations, many researchers have employed surrogate modeling to relate the inputs and the target outputs produced the simulation (Kim & Boukouvala, 2020) (Xiao et al., 2021).

One case study where this type of mixed-integer, simulation-based optimization problem arises is in design of adsorption processes for carbon capture using solid adsorbents. Recent studies show that the choice of adsorbents strongly affects the process performance, implying process optimization and adsorbent selection should be considered alongside (Subramanian Balashankar & Rajendran, 2019). This requires formulating a complex simulation-based optimization problem containing partial differential algebraic equations and a large search space with both continuous and discrete variables. In this study, we present a ML-based approach for simultaneous process optimization and adsorbent selection. Firstly, we propose a technique to design modular CO₂ capture systems for coal-fired power plants using Vacuum Pressure Swing Adsorption (VPSA) coupled with thermally modulated fiber composite adsorbents. We then use surrogate-based optimization to determine the optimal adsorbent as well as process conditions to meet at least 90% recovery and 95% purity constraints (Kim, Landa, Ravutla, Realff, & Boukouvala, 2022). We compare feasibility and computational cost savings of simultaneous optimization of adsorbent selection and operation using surrogate-based optimization versus a brute force optimization of operation for each material separately. For this study we explore a pool of 75 different adsorbent materials and 5 process conditions for optimization. Finally, we use the large amount of local-optima collected by our solver and utilize ML techniques to develop “adsorbent-to-optimal process” correlations. Specifically, we will show that using principal component analysis we can generate a reduced-space that represents the material properties, while coupling that with process inputs we can build classification and regression models to predict the feasibility and process outputs of different materials.

References

Kim, S. H., & Boukouvala, F. (2020). Surrogate-Based Optimization for Mixed-Integer Nonlinear Problems. Computers & Chemical Engineering.

Kim, S. H., Landa, H. O. R., Ravutla, S., Realff, M. J., & Boukouvala, F. (2022). Data-driven simultaneous process optimization and adsorbent selection for vacuum pressure swing adsorption. Chemical Engineering Research and Design, 188, 1013-1028. doi:https://doi.org/10.1016/j.cherd.2022.10.002

Subramanian Balashankar, V., & Rajendran, A. (2019). Process Optimization-Based Screening of Zeolites for Post-Combustion CO2 Capture by Vacuum Swing Adsorption. ACS Sustainable Chemistry & Engineering, 7(21), 17747-17755. doi:10.1021/acssuschemeng.9b04124

Xiao, J., Mei, A., Tao, W., Ma, S., Bénard, P., & Chahine, R. (2021). Hydrogen Purification Performance Optimization of Vacuum Pressure Swing Adsorption on Different Activated Carbons. Energies, 14(9), 2450. Retrieved from https://www.mdpi.com/1996-1073/14/9/2450