The technology of producing chemicals directly from crude oil involves the direct catalytic cracking of crude oil into chemical raw materials. This innovative process bypasses traditional atmospheric and vacuum distillation units and hydrogenation units, directly reducing both equipment investment and energy consumption. Consequently, this leads to lower production costs and brings significant economic benefits. The direct conversion method not only streamlines the production process but also minimizes the need for complex infrastructure, making it a more efficient alternative to conventional methods. With the continuous advancement of the national dual carbon goals-aiming to peak carbon dioxide emissions and achieve carbon neutrality-accelerating the development of technology for direct catalytic cracking of crude oil to produce chemicals is of paramount importance. This technology holds the potential to significantly reduce process energy consumption and contribute to carbon emission reduction efforts. By optimizing the cracking process, it is possible to achieve higher yields of desirable chemicals while minimizing the formation of by-products like coke, which are less valuable and contribute to increased emissions. As research on the modeling of technology for producing chemicals directly from crude oil deepens, establishing intelligent models to guide process production becomes increasingly crucial. These models can optimize process operations by fine-tuning parameters in real-time, thereby achieving a balance between economic benefits and environmental sustainability. Intelligent models leverage data-driven insights to predict outcomes and adjust variables dynamically, ensuring that the production process remains efficient and aligned with both economic and environmental targets. This study established a robust process simulation model in Aspen HYSYS based on industrial trial data from direct catalytic cracking of crude oil. Through detailed case analysis, single-factor analysis was conducted on four critical process parameters: preheating temperature, reaction temperature, regeneration temperature, and catalyst equilibrium activity. Each of these parameters played a vital role in determining the efficiency and yield of the cracking process. By systematically varyied these factors, researchers can identify optimal conditions that maximize the production of key chemicals such as ethylene and propylene while minimizing unwanted by-products like coke. To enhance the predictive capabilities of the model, a neural network was implemented using Python programming. This neural network model was trained on a comprehensive dataset derived from the process simulations. The model's ability to predict product distribution under different operating conditions was rigorously tested and validated. Furthermore, a multi-objective optimization algorithm, NSGA-II, was integrated into the deep learning framework. This algorithm focuses on maximizing the yield of low-carbon olefins while minimizing coke production, providing a balanced approach to optimizing the overall process. Compared to traditional optimization methods, the established surrogate model offers higher computational efficiency and faster optimization solution times. It enables the decoupling of multiple operational variables, allowing for more precise control over the process. This real-time optimization capability is particularly beneficial in dynamic production environments where conditions can change rapidly. The optimization results demonstrated notable improvements: coke yield decreased by 0.23%, while the yields of ethylene and propylene increased by 1%. In conclusion, the intelligent agent model developed in this study not only enhances solution efficiency and prediction accuracy but also provides valuable insights for guiding process production. Its application could lead to more sustainable and cost-effective chemical manufacturing processes, aligning with both economic and environmental objectives.