AbstractThe delineation of vertical geological cross-sections is an essential task in geotechnical site characterization and has a profound impact on subsequent geotechnical designs and analyses. It is a long-lasting challenge, particularly for complex geological settings, to properly produce a subsurface geological cross-section from limited boreholes that are usually encountered in engineering practice. Emerging machine learning methods, such as the convolutional neural network (CNN), provide a fresh perspective of this challenge and effective alternatives for exploiting the complex stratigraphic relationships between different soil deposits. In this study, a novel iterative convolution eXtreme Gradient Boosting model (IC-XGBoost) is proposed. This model interpolates a subsurface geological cross-section from limited site-specific boreholes and a training geological cross-section obtained from previous projects with similar geological settings. This direct application of previous geological cross-sections for training is based on the assumption of similar local spatial connectivity or stratigraphic relationships between soils in areas with similar geological settings. The proposed method can learn stratigraphic patterns from a training image in an automatic manner. In addition, the proposed method is purely data-driven and does not require the specification of any parametric function form. The model performance is illustrated using both a simulated example and real data from a tunnel project in Australia. The proposed method not only infers the most probable geological cross-section but also quantifies the associated interpolation uncertainty from multiple realizations. The effect of the borehole number on the interpolation performance is also explicitly investigated.