Prediction of compression coefficient of Nanjing floodplain soft soil based on explainable artificial intelligence

Abstract

The low bearing capacity and high compressibility of soft soils significantly influence the design of building foundations. Consequently, accurate prediction of the compression coefficient is essential for ensuring the stability and safety of structures. This study established a database consisting of 699 samples of Nanjing floodplain soft soil and developed a hybrid machine learning model, CNN (Convolutional Neural Network) − CatBoost (a gradient algorithm utilizing symmetric decision trees), which utilizes deep feature extraction through CNN to accurately predict the compression coefficient of Nanjing floodplain soft soil. The coefficients of determination (R²) for the training and testing sets were 0.965 and 0.933, respectively. In comparison to traditional models, the hybrid model demonstrated significant advantages in prediction accuracy and error management, exhibiting improved fitting and generalization capabilities. Furthermore, SHAP and PDP analyses were conducted to evaluate the influence of five input features—wet density, plastic limit, plasticity index, liquidity index, and depth—on the output results, indicating that the plasticity index had the most substantial effect on the compression coefficient estimated by the hybrid model. This model offers a promising tool for advancing geotechnical engineering applications, enhancing prediction accuracy and decision-making in foundation design.