The macroscopic mechanical properties of gap-graded soils vary significantly due to gradation characteristics, and accurately establishing relationship expressions between gradation parameters and mechanical properties has long been a major challenge in engineering. To address this issue, stress-strain datasets of gap-graded soils were obtained using the discrete element method and an innovative constitutive model framework was proposed. This framework takes fine content, particle size ratio, and strain as input parameters, with deviatoric stress as the output parameter. Based on Bayesian Optimization (BO), the Random Forest (RF) model and the Back Propagation Neural Network (BPNN) were respectively adopted for model training and validation. Research results show that both RF and BPNN models demonstrate high coefficients of determination. When predicting the complete stress-strain curves of gap-graded soils, the RF model outperforms the BPNN model, with higher consistency between predicted and simulated values. This study successfully applies machine learning techniques to predict the mechanical properties of gap-graded soils, offering new insights into modeling the complex relationship between gradation parameters and mechanical properties.