To address the limited accuracy of empirical formulas, the high computational cost of CFD simulations, and the poor small-sample generalization and limited uncertainty quantification of conventional machine learning methods in predicting erosion wear in the build-up section of horizontal-well fracturing strings, this study proposes a CFD-driven adaptive Gaussian process regression (GPR) method for maximum erosion rate prediction. A liquid-solid two-phase erosion model of the build-up section was developed using the DDPM model to investigate the effects of pumping rate, sand concentration, particle density, and particle diameter on erosion wear. A small-sample dataset was then constructed from the numerical simulation results. A greedy strategy was used to adaptively search for composite kernel functions, and the prediction performance was compared with MLR, SVR, XGBoost, BPNN, and Kolmogorov-Arnold network models. SHAP interpretability analysis and response surface analysis were further used to quantify the influence of each factor on the maximum erosion rate. The results show that the maximum erosion rate increases exponentially with pumping rate, increases approximately linearly with sand concentration and particle density, and decreases with increasing particle diameter. The optimized adaptive GPR model achieves the best prediction accuracy and generalization performance under small-sample conditions, with a test-set R2 of 0.946 and the ability to output 95% confidence intervals. SHAP and response surface analyses indicate that the influence of the factors on the maximum erosion rate follows the order of pumping rate > sand concentration > particle density > particle diameter. By constructing an adaptive composite-kernel GPR prediction model, this study enables rapid prediction and uncertainty quantification of erosion wear under small-sample conditions, providing a reference for erosion risk assessment and fracturing parameter optimization in the build-up section of horizontal-well fracturing strings.