Abstract: To address the load non-uniformity phenomenon exhibited by composite material connection structures during loading, a data-driven load homogenization method for composite joint structure is proposed. This method ensures high prediction accuracy and generalization, while avoiding issues such as premature convergence seen in traditional optimization algorithms through an adaptive optimization process. As a result, better load homogenization is achieved. First, the load transfer mechanism under gap and no-gap conditions for a four-bolt single-lap connection structure is studied. To address the load non-uniformity of the connection structure, a load homogenization approach is proposed, using gap and preload as optimization parameters. Next, the optimal Latin Hypercube Sampling (OLHS) method is used to design the database, and a Bayesian optimization-based fully connected neural network (FCNN) prediction model is developed to establish the mapping relationship between the optimization parameters and load unevenness. On this basis, the artificial lemming algorithm (ALA) combined with an FCNN predictive model is employed to optimize the gap and preload with the minimization of load non-uniformity as the objective, thereby improving the load non-uniformity phenomenon in composite joint structures. The results indicate that the FCNN prediction model based on Bayesian optimization achieves R² values greater than 0.98 for both the training and test sets, with performance surpassing that of the RF, SVM, and RSM models. After ALA optimization, the load non-uniformity of the four-bolt single-lap joint structure is reduced from 0.193 to 0.019, significantly improving load distribution homogenization and outperforming traditional optimization algorithms.