AI Models Predict Undrained Shear Strength with Incomplete Data

Researchers have developed advanced probabilistic indirect models to accurately predict undrained shear strength (su), a critical parameter in geotechnical design. This is particularly challenging due to the high rates of missing data and inherent variability in traditional empirical methods. The study leverages a global database (CLAY/10/7490) and focuses on predicting su from Atterberg limits and piezocone cone penetration (CPTU) measurements. The methodology involves a two-step process. First, three imputation methods—multivariate normal (MN), multiple imputation by chained equations (MICE), and miss forest (MF)—were tested to handle missing values, with their effectiveness validated using a Probabilistic Extreme Gradient Boosting (PXGB) model. Second, the indirect prediction model, named MHA-PNN, integrates a multi-head attention (MHA) mechanism into an artificial neural network (ANN) to improve information extraction from limited data. The MN-enhanced MHA-PNN model significantly outperformed other models in both prediction accuracy and uncertainty quantification, demonstrating its potential for robust geotechnical applications with sparse datasets.