Objective: ] Chemical plant layout design involves inherently high levels of hazards, strong constraint coupling, and conflicting multi objective requirements, making it a long-standing challenge in process systems engineering. Traditional layout planning methods rely heavily on expert knowledge, deterministic rules, and heuristic optimization, which are often insufficient for dealing with high-dimensional, nonlinear, and nonconvex design spaces. To address these limitations and improve the intelligence and automation of hazardous chemical plant layout design, this study proposes a generative optimization framework named SAC-GEN, based on deep reinforcement learning and agent-environment interactive learning. The goal is to enable the autonomous generation of safe, compliant, and cost-effective plant layouts, especially for facilities involving flammable and explosive materials. Methods: The proposed SAC-GEN framework integrates the Soft Actor-Critic (SAC) algorithm as the core decision-making kernel, leveraging its entropy-regularized policy to ensure stable learning and effective exploration in continuous action spaces. A domain-informed simulation environment was constructed to embed fundamental chemical engineering design knowledge into the reinforcement learning process. A multilayer state representation scheme was developed to describe equipment geometric characteristics, relative spatial relationships, explosion hazard levels, pipeline connectivity, and layout feasibility. To ensure realistic, practical layout outcomes, multiple engineering constraint mechanisms were incorporated, including equipment boundary restrictions, nonoverlapping spatial feasibility checks, and layout invalidation penalties. A dynamic reward-shaping strategy, combining safety performance, spatial rationality, and economic indicators was designed to guide the agent toward balanced, high-quality layouts under trade-off conditions. For safety modeling, a TNT-equivalent explosion model was used to calculate the blast impact radius of hazardous units, and a quantitative risk diffusion model was implemented to simulate the attenuation of explosion energy across the plant area. In addition, a domino-effect propagation mechanism was developed to capture secondary explosions triggered by equipment-to-equipment impact. In this mechanism, the ignition sequence evolves dynamically based on spatial adjacency and blast-wave impact magnitude, enabling evaluation of both individual explosion consequences and cascading failure risks frequently observed in chemical industrial accidents. For economic evaluation, a hybrid pipeline routing algorithm based on axis-aligned (HV) path planning was constructed to compute material transfer paths, pipeline lengths, and connection complexity between units. This algorithm provides a practical economic indicator for layout rationality. By integrating these mechanisms, SAC-GEN achieves an intelligent mapping from safety-economic design objectives to spatial layout solutions. Results: The framework was validated through a case study on a 100 m × 80 m sulfuric acid alkylation unit. Ten representative equipment items, including a reactor, tanks, distillation columns, a compressor, and separation systems, were modeled with corresponding explosion hazard levels and logistics flow connections. The reinforcement learning model autonomously generated multiple feasible layout solutions, which were then analyzed from three perspectives: minimum explosion- risk, minimum pipeline- length, and overall optimization. Results showed that the safety-oriented layout substantially reduced the blast impact zone and eliminated domino risks by spatially isolating high-energy units. The economy-oriented layout reduced the total pipeline length by more than 20% and improved material transfer efficiency, albeit at the expense of reduced safety margins. The comprehensive optimal solution achieved a desirable balance between inherent safety and economic performance, demonstrating SAC-GEN's capability to navigate multi objective conflicts and produce practically adoptable layouts. Conclusions: The SAC-GEN framework provides a systematic and scalable method ology for intelligent layout generation in hazardous chemical plants. It successfully handles complex spatial constraints, nonlinear safety-economy relationships, and large continuous design spaces that challenge conventional methods. The approach significantly reduces dependence on expert knowledge, enhances inherent safety performance, and improves decision-making efficiency. Future work will focus on integrating digital twin technology, real-time risk assessment, and online optimization to support industrial deployment within smart chemical plant design and operation systems.