Combining Surrogate Modeling, Superstructures and Neural-Aided Optimization for Design of Innovative Intensified Processes

The energy transition requires certain chemical processes to be redesigned to adapt to sustainable operation (biobased inputs, intermittency, decentralized production, etc.). These changes necessitate a reshape of the process structure. Concurrently, intensified multifunctional units make it possible to take advantage of the synergies between transformation functions (reactions, transfers, separations, etc.) within a single piece of equipment. Including these units in a process may lead to a simplified process. Therefore, design of energy-efficient processes necessitates to evaluate new process structures.

Process synthesis methods have considerably evolved from heuristics-based to mathematical-programming based optimization (MINLP) and evolutionary approaches. Numerous tools are available to widen the scope of attainable structures. Implementation of superstructures in CAPD software is possible through conditional splitters (switches) and optimization libraries. Extensions enable to combine process-engineering criteria and product-engineering formulations. User-defined functions may consider fast surrogate models, either to accelerate optimization or to substitute time-consuming models. Multi-objective optimization with constraints are helped by neural networks to accelerate convergence in the range of decision variables.

This work illustrates these opportunities with examples. Efficiency of micro-CHP systems is increased by compact heat-exchanger reactors: constrained multi-objective optimization is slowed down by the simulation of these units. Iterative fitting of a neural-based surrogate model of the objectives and constraints enables to explore the optimality domain and focus within the area of interest. Combined heat and power thermodynamic cycles can be taylor-designed by merging product and process design approaches: several fluids are screened while the optimal flowsheet is determined by superstructure approach including bleeding, reheating or regeneration as structural alternatives. For the synthesis of Dimethyl Ether, multifunctional reactors (coupling reaction, heat-exchange and separation) are relevant to reach very high yields, but their integration in a CAPE environment is delicate: Physics-Informed Neural Networks (PINNs) fasten the optimization step for further comparison with conventional technologies.