Abstract: The fuel loading pattern optimization (LPO) for the HPR1000, a third-generation pressurized water reactor, represents a highly non-linear and large-scale combinatorial challenge. Traditional optimization methods primarily rely on high-fidelity neutron transport codes, which ensure physical accuracy but impose a prohibitive computational burden during large-scale heuristic searches. To address the inherent contradiction between computational precision and optimization efficiency, this study developed a synergistic physical-verification and surrogate-prediction co-optimization architecture based on machine learning. The primary objective is to enhance the thermal safety margin of the HPR1000 core by minimizing the power peaking factor (PPF) while maintaining criticality and fuel inventory constraints. In the methodological phase, a comprehensive optimization framework was established by coupling the Open Monte Carlo (OpenMC) code with an improved genetic algorithm (GA). A lightweight surrogate model, based on the random forest (RF) algorithm, was constructed to map the complex relationship between assembly spatial arrangements and core physics parameters. To ensure the reliability of the surrogate model in unexplored regions of the design space, an adaptive evolution strategy was implemented. During each optimization cycle, the GA-searched candidates were verified using OpenMC, and the high-fidelity results were fed back into the training database for self-calibration. A total of 80 initial samples were generated via Latin Hypercube Sampling (LHS), followed by approximately 350 iterative cycles of adaptive learning. The results show that the surrogate model achieves exceptional predictive stability after 250 cycles of self-evolution. The prediction residuals for the effective multiplication k_eff and PPF are reduced to below 0.05% and 1%, respectively. Crucially, the synergistic architecture significantly breaks the limitations of time-consuming transport calculations on the search scale. The total computational time for a single optimization task is reduced from the hundred-hour level to the minute level, representing a computational efficiency improvement of approximately two orders of magnitude compared to full-physics search modes. The optimized loading pattern obtained through this architecture demonstrates superior performance. Compared with the initial cycle reference loading pattern, the PPF of the optimized scheme decreases from 1.732 5 to 1.510 2, achieving a performance gain of approximately 12.8%. Physical mechanism analysis reveals that the optimized pattern adopts a “fragmented” microscopic interlaced layout, which effectively utilizes the spatial coupling effect between assemblies. By placing a cross-shaped arrangement of low-enrichment assemblies in the core center to create reactivity “traps” and isolating high-enrichment assemblies, the layout flattens the radial flux gradient and suppresses local power spikes at zone boundaries. Furthermore, the optimal solution maintains low-leakage characteristics at the core periphery, satisfying the requirements for pressure vessel protection. This study proves that data-driven methods possess significant potential for refined fuel management in complex reactor cores and provides a numerical reference for enhancing the thermal safety margins of next-generation pressurized water reactors.