Comparing generative process synthesis approaches with superstructure optimization for the conception of supercritical CO2 Brayton cycles

In process synthesis, while heuristic-based approaches are most often used for proposing relevant alternatives (that must then be thoroughly analyzed), this strategy may not be the most efficient. When it comes to the search for innovative processes, prior domain knowledge may be scarce or may not e...

Full description

Saved in:
Bibliographic Details
Published in:Computers & chemical engineering Vol. 201; p. 109255
Main Authors: Rocha Azevedo, Antonio, Nabil, Tahar, Loubière, Valentin, Privat, Romain, Neveux, Thibaut, Commenge, Jean-Marc
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.10.2025
Subjects:
Artificial intelligence
Evolutionary programming
Process synthesis
Superstructure
Process synthesis
Evolutionary programming
Artificial intelligence
Superstructure
ISSN:0098-1354
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In process synthesis, while heuristic-based approaches are most often used for proposing relevant alternatives (that must then be thoroughly analyzed), this strategy may not be the most efficient. When it comes to the search for innovative processes, prior domain knowledge may be scarce or may not effectively exploit the properties contributing to the process novelty. Generative synthesis approaches that can freely explore the search space and that do not rely on any previous knowledge, have been proposed in the literature. Yet, a lack of benchmarks on complex problems strongly hinders their use. In this work, we address this gap by comparing two generative approaches, based on Evolutionary Programming and Machine Learning, to a superstructure optimization (which serves as a baseline). They are applied to the synthesis of supercritical CO2 Brayton cycles. Despite starting with no field of expertise, the generative approaches not only manage to identify multiple known heuristics of the domain, but also a counter-intuitive and new way of increasing the efficiency of sCO2 cycles — by expanding the fluid at lower temperatures. The approaches’ use-cases are discussed, based on the amount of computational resources necessary, implementation difficulties and quality of the results. [Display omitted] •Two generative synthesis approaches are benchmarked for a complex problem.•Both approaches discover a new heuristic for increasing efficiency of sCO2 cycles.•Large volumes of data are generated, which can be useful for data mining and ML.•The evolutionary method found good cycles despite starting from an empty flowsheet.•A better fine-tuning strategy is developed for training the ML-based approach.
ISSN:0098-1354
DOI:10.1016/j.compchemeng.2025.109255