A many-objective surrogate optimization model driven by hybrid pilot-test data, molecular reconstruction, and crude oil direct cracking reaction mechanism

• Published in: Chemical engineering journal (Lausanne, Switzerland : 1996) Vol. 507; p. 160389
• Main Authors: Zhou, Xin, Zhang, Zhibo, Wang, Changyuan, Wu, Lianying, Yan, Hao, Zhao, Hui, Liu, Yibin, Chen, Xiaobo, Yang, Chaohe
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2025
• Subjects: Cracking reaction mechanisms; Deep learning; Many-objective optimization; Molecular reconstruction; Surrogate model; LPG; Surrogate model; BTX; DRT; GHG; IRR; RRCT2; RRCT1; DMCD; DDO; Deep learning; GDP; NPV; RRT; SFR; SBO; PHFT; CGO; Many-objective optimization; LCA; MOO; Molecular reconstruction; MOEA/D; PED; Cracking reaction mechanisms; MCO; RRT2; POSS; NSGA; RRT1
• ISSN: 1385-8947
• DOI: 10.1016/j.cej.2025.160389

The Graphic Abstract illustrates a surrogate model using a deep residual network (or other deep learning networks) driven by hybrid pilot-test data and molecular reaction mechanism for many-objective optimization of crude oil catalytic pyrolysis. [Display omitted] •A novel optimization model including pilot-test data and cracking mechanism is proposed.•The molecular-level cracking model combined with lumped cracking model is employed.•The many-objective optimization algorithm MOEA/D is applied in the optimization.•The gasoline-oriented process shows better economic & environmental character. A computationally efficient surrogate model leveraging deep learning and molecular reconstruction has garnered significant attention for addressing complex, large-scale optimization challenges. This study presents a deep residual network-based surrogate model that integrates deep learning with molecular reconstruction and cracking reaction mechanisms to handle many-objective optimization problems. The model focuses on optimizing across various dimensions—economic, societal, livelihood, and environmental—in the catalytic cracking process of crude oil. Initially, a hybrid database combining extensive process data and molecular reaction mechanisms is created. The study then examines the interactions between cracking reaction mechanisms and key operational variables, leading to the development of a hybrid model that merges deep learning with mechanistic insights. To manage the many-objective optimization challenges, the MODE/A algorithm is employed. Two scenarios were evaluated from a lifecycle perspective: creating GDP orientation (CGO and maximizing chemical orientation (MCO). The results indicate that the CGO process utilizes 29 tons of crude oil and generates 46.77 tons CO2 less than the MCO process for every USD 1 million of GDP produced. This research framework offers a comprehensive strategy for enhancing the efficiency and effectiveness of direct catalytic cracking processes.