This study presents a novel framework for automated generative process synthesis, addressing the complexity of simultaneously optimizing discrete topologies and continuous operating variables. To overcome conventional superstructure limitations, we propose a dual-loop architecture integrating generative transformers with rigorous process simulation. The framework utilizes SFILES to encode process structures into textual sequences, which are mapped onto a continuous embedding manifold via a transformer-based model. The optimization is executed through a dual-loop approach: the outer loop leverages Particle Swarm Optimization (PSO) to explore the embedding space for candidate topologies, while the inner loop serves as an automated evaluation agent that performs rigorous simulation-based optimization within Aspen Plus®. Validated on cyclohexane, ethylene glycol, and CO2 capture systems, the framework successfully rediscovered established industrial configurations and identified novel, superior alternatives. By bridging generative modeling with high-fidelity physics-based simulation, this work establishes a robust computational framework for the automated exploration and structural evolution of chemical processes.

(Yeong Woo Son and Chan Kim contributed equally to this study.)