Conventional geological modeling methods are typically constrained by limited borehole data and are unable to effectively capture irregular stratigraphic variations within complex geological structures, resulting in modeling inaccuracies. To overcome these limitations, a novel geological modeling approach was proposed. In this method, a multi-layer perceptron model was employed to integrate borehole data with seismic interpretation data, enabling the learning of deviation patterns between seismic interpretation horizons and actual horizons. A multi-scale iterative prediction process based on horizon errors was then developed, in which horizon errors at different scales were progressively predicted and corrected to achieve accurate prediction of the spatial morphology of geological structures in unexplored regions. The proposed approach was validated using synthetic data from the North Sea Continental Shelf. The experimental results demonstrate that, in regular geological structures, an average prediction accuracy of 99.94% is achieved, with mean horizon boundary errors of 1.135 m, 1.266 m, and 1.127 m, corresponding to reductions of 98.3%, 97.65%, and 97.83%, respectively, compared with the IC-XGBoost method. In irregular geological structures, an average prediction accuracy of 99.95% is achieved, with mean horizon boundary errors of 1.021 m, 1.279 m, and 1.057 m, representing reductions of 98.82%, 97.37%, and 98.91%, respectively. These results verify that the proposed method can significantly enhance the accuracy and robustness of geological modeling in complex structural settings.It is obvious that the proved method plays an obvious role in improving the prediction accuracy of the layer and the stability of geological modeling.