Abstract: The graph convolutional network (GCN) algorithm is applied to geological profile prediction, and an improved graph convolutional network algorithm is proposed through modifications. This algorithm can flexibly establish connection graphs based on unequal and sparse boreholes within the site, enabling the prediction of the two-dimensional geological profile of the site. A boundary accuracy rate is proposed to better judge the accuracy of the predicted geological profile. This method is applied to actual geological profile cases and compared with the existing IC-XGBoost method. The results show that this method improves the global accuracy and the boundary accuracy compared with other methods which indicates its accuracy in predicting two-dimensional geological profiles. Finally, the influence of the number of boreholes and unequal spacing on the prediction results is discussed. The results indicate that the more boreholes there are, the more accurate the predicted geological profile boundary will be. However, the increase in the number of boreholes and the improvement in prediction accuracy are not a simple linear relationship. The improvement in prediction accuracy depends on the soil layer spatial information provided by the additional boreholes. When the number of boreholes is the same, the influence of unequal spacing of boreholes on the results depends on the soil layer spatial information provided by the locations of the boreholes. The more soil layer spatial information provided by the boreholes, the more accurate the prediction will be. Finally, this improved GCN method applied in Hangzhou real case to predict the geological profile. This method works well in real case that all the measurement accuracy of validated boreholes above 0.8.

Key words: Graph convolutional neural network, Sparse boreholes, Unregular, Stratigraphic interface, Boundary accuracy