Data-Driven Settlement Prediction for Pavements on Tunis Soft Clay Improved with Deep Soil Mixing: Artificial Intelligence and Response Surface Approaches

This study investigates the prediction of immediate settlement (Uz) in soft clay improved with Deep Soil Mixing (DSM) columns under heavy aircraft loading. Two key design parameters were considered: column spacing (2.25 m to 3.75 m) and column length (6 m to 20 m), with both rectangular and triangular arrangements analyzed. The datasets obtained from numerical simulations were modeled using Response Surface Methodology (RSM) and Artificial Neural Networks (ANN), with model calibration and validation performed through k-fold cross-validation. The statistical analysis revealed that both approaches achieved excellent predictive capability, with R2 values exceeding 0.999. For the rectangular arrangement, RSM yielded slightly lower errors (RMSE = 0.0636 cm, MAE = 0.0553 cm) compared to ANN (RMSE = 0.0828 cm, MAE = 0.0682 cm), suggesting that a second-order polynomial approximation can effectively describe the settlement response in this configuration. Conversely, for the triangular arrangement, ANN clearly outperformed RSM, reducing RMSE from 0.0725 cm to 0.0265 cm and MAE from 0.0615 cm to 0.0111 cm, thereby capturing the nonlinear stress redistribution associated with isotropic column layouts more effectively. Observed–predicted plots confirmed the high predictive accuracy of both methods, with ANN showing superior generalization in triangular grids. Overall, the findings highlight that RSM remains a robust and computationally efficient tool for rectangular layouts with relatively linear responses. In contrast, ANN provides enhanced accuracy for triangular configurations where nonlinear interactions dominate, making it particularly suitable for DSM design optimization in airport pavement foundations.