Abstract: This study focused on the intelligent optimization design of small modular lead-cooled fast reactors (SM-LFRs). Addressing the complex issue of mutual coupling among multiple variables, multiple constraints, and multiple objectives in reactor design, it proposed an optimization framework integrating the BP neural network and NSGA-Ⅲ algorithm, with the initial scheme of QJMF-S as the research object for exploration. This reactor model was intended to provide energy support for scenarios such as digital intelligence centers and remote areas. Its initial design adhered to principles including full natural circulation, transportability, and the use of low-enriched uranium. However, there was room for optimization in terms of size adaptability and scheme diversity, requiring intelligent methods to achieve efficient optimization. To improve optimization efficiency and accuracy, a BP neural network was used to construct an enrichment prediction model and innovatively proposes an input-output exchange training method. This method invertd the process originally used for calculating critical eigenvalues, ensuring that the generated schemes always meet criticality requirements while significantly reducing the time consumed by a single calculation, thus laying the foundation for large-scale scheme comparison and selection. During model training, measures such as data preprocessing and early stopping mechanism were adopted to control errors and ensure prediction reliability. In the multi-objective optimization phase, the NSGA-Ⅲ algorithm was selected, and optimization was carried out focusing on three core objectives of the reactor including total height, total diameter, and total power. The first two objectives needed to be as small as possible to meet transportation requirements, while the last one needed to be as large as possible to enhance energy supply capacity. In the process, key efforts were made to optimize hyperparameters such as the initial value range and population size. The hypervolume indicator was used to evaluate the convergence and diversity of the population, ensuring uniform coverage of the solution space. At the same time, constraints such as coolant flow rate, fuel enrichment, and linear power peak were strictly followed to screen schemes that meet safety and performance requirements. The optimization results show that this intelligent method breaks through the limitations of manual optimization and obtains a variety of quasi-Pareto frontier schemes. Some schemes achieve comprehensive improvements in the three core objectives and exhibit core configuration characteristics different from the initial scheme, which are more in line with the needs of natural circulation. This study not only verifies the feasibility of the optimization framework and provides key references for the scheme demonstration of related series of reactors, but also confirms its versatility, which can be extended to other multi-objective optimization scenarios in the field of nuclear engineering.