Scale-up of complex molecular reaction system by hybrid mechanistic modeling and deep transfer learning

The scale-up of chemical processes involves substantial changes in reactor size, operational modes, and data characteristics, leading to significant challenges in predicting product distribution across scales. This study presents a unified modeling framework that integrates the mechanistic model with deep transfer learning to accelerate chemical process scale-up. The framework is demonstrated through a case study on naphtha fluid catalytic cracking. A molecular-level kinetic model was developed from laboratory-scale experimental data, and a deep neural network was designed and trained to represent complex molecular reaction systems. To address the challenge of discrepancies in data types at various scales, a property-informed transfer learning strategy was developed by incorporating bulk property equations into the neural network. This approach enabled automated prediction of pilot-scale product distribution with minimal data. Moreover, process conditions of the pilot plant were optimized using a multi-objective optimization algorithm. Process scale-up is a persistent challenge in chemical industries. Here, the authors integrated the mechanistic model with transfer learning to accelerate process scale-up, and product distribution was auto-predicted from the laboratory to the pilot.