A novel GA-GBM based model to construct CPT-SPT relationship

Abstract

Numerous relationships between Cone Penetration Test (CPT) and Standard Penetration Test (SPT) have been developed which consist of empirical formulas, statistical models, and machine learning models. However, these models often suffer from low prediction accuracy and limited generalizability when converting CPT data to SPT N or SPT N₆₀. In this study, a GA-GBM based model was developed to construct the CPT-SPT relationship in order to address the aforementioned issues. The raw database was segmented by intervals of soil behavior type index (I_c) using the GA, and the BP method was then employed to identify and eliminate outliers within each I_c interval. The preprocessed dataset was analyzed using the GBM technique with the results returned to the GA to iteratively adjust boundaries of each I_c interval (i.e. Method A) or obtain an optimal number of I_c interval (i.e. Method B) that optimizes the predictive performance of the proposed model. As compared to other widely used models reported in literature, the accuracy of predicted SPT N₆₀ from CPT data was significantly improved with an R² of 0.80 and an RMSE of only 5.80. For the new dataset, R² of 0.73 was obtained by the proposed model which is much higher than other models. It is demonstrated the proposed GA-GBM model exhibits strong generalization capability and can predict SPT N₆₀ more accurately. Overall, the proposed model offers new insights into the relationship between CPT and SPT.