(200c) Optimal Design of Flowsheets Via Surrogate-Based Sub-Process Models: A Case Study of CO2 Capture and Utilization

Flowsheet design enables a top-down overview of processes. The focus is not on a single process, but on the interactions between processes, which leads to more optimal overall solutions. Chemical process models commonly include non-linear equations due to complex kinetics and thermodynamics, whereas discrete choices of alternative processes or, for example, the tray number of a distillation column, introduces discrete variables. Accordingly, the mathematical formulation of a flowsheet is generally a complex mixed integer nonlinear programming (MINLP). The larger the flowsheet is, the more difficult is the numerical problem of solving it. The use of simplified models facilitates the numerical solution, but introduces uncertainty due to the loss of details. For example, the catalyst information is missing in the equilibrium reactor model, and shortcut distillation column model cannot deal with azeotropic mixtures [1]. To reduce uncertainty, rigorous process models can be implemented in commercial software, e.g. Aspen, Pro II and gPROMS. In a large flowsheet, sub-processes may be simulated in multiple packages and then their results are transferred to one platform for the overall simulation. However, the optimization tends to be challenging due to difficulties with initialization, time-expensive evaluations and problems of convergence [1]. Additionally, the speed is slowed down by data export-storage-import between different simulators.

Surrogate models trained from rigorous models can solve the above-mentioned issues. Machine learning algorithms (Surrogate models, Linear, Polynomial, Neural Networks, and Artificial Neural Networks) can build a direct correlation between process input and output based on sufficient simulations via rigorous models [2-4]. Artificial Neural Networks (ANNs) are said to be universal approximations [3-4]. These surrogate models can then be implemented in the same platform for further optimization. Additionally, life cycle impact (LCI) analysis can be incorporated into objective function for environmentally-friendly solutions. The steps of this methodology are as follows:

• uncertainty should be reduced to an acceptable range through sufficient simulations of rigorous models;
• smart sampling methods can be used to identify the effective inputs for simulations;
• Gaussian process or ANNs can be used to establish surrogate models;
• LCI analysis can introduce environmental aspect of sustainability into the optimization objectives;
• multi-objective optimization can be introduced to present the trade-off between different objectives.

A case study of optimal flowsheet design for carbon capture and utilization (CUU) is presented. CO₂ capture is performed in vacuum swing adsorption or MEA-based absorption process, while the captured CO­₂ is then converted to valuable products via typical chemical processes (this case study is focused on methanation, methanol synthesis and Fischer-Tropsch synthesis). Each sub-process is trained into a data-driven model incorporated with LCI analysis, and then optimization is performed for the design of sustainable flowsheet.

References

1. Recker, S. Systematic and Optimization-based Synthesis and Design of Chemical Processes, Ph.D. Dissertation, RWTH Aachen, Aachen, Germany, 2017.
2. Gonzalez-Garay, A., Guillen-Gosalbez, G. SUSCAPE: A framework for the optimal design of SUStainable ChemicAl ProcEsses incorporating data envelopment analysis. Chemical Engineering Research and Design2018, 137, 246-264.
3. Schweidtmann, A. M., Mitsos, A. Global deterministic optimization with artificial neural networks embedded. arXiv preprint arXiv:1801.07114. 2018.
4. Wilson, Z. T., Sahinidis, N. V. The ALAMO approach to machine learning. Computers & Chemical Engineering 2017, 106, 785-795.