Deep Learning for Prediction of Unconfined Compressive Strength Depending on Various Geotechnical Characteristics of some Soils of Boroughs of Sulaymaniyah City, Iraq Using H2O-ANN

Geotechnical engineering faces significant challenges when dealing with fine-grained soils due to their high plasticity and, in some cases, swelling behavior. To address these challenges, this study aims to predict the Unconfined Compressive Strength (UCS) of various soil types collected from different districts of Sulaymaniyah City, Iraq, by developing an Artificial Neural Network (ANN) model using H2O’s deep learning framework. A comprehensive set of geotechnical properties including Atterberg limits, moisture content, density, specific gravity, and shear strength parameters was analyzed using 60 intact soil samples. The ANN model, developed using Python and the H2O-ANN library, successfully captured the complex nonlinear relationships between these input variables and UCS. The model achieved high predictive accuracy, with an R² value of 0.90 and low error metrics (RMSE ≈ 33.42 kPa and MAE ≈ 14.2 kPa). Model performance was further validated through residual analysis and partial dependence plots, which highlighted the significant influence of density, moisture content, and plasticity characteristics on UCS. Compared with conventional regression models and similar ANN-based studies on stabilized expansive soils, the proposed model demonstrates greater flexibility and improved prediction accuracy for natural, non-stabilized soils. This approach provides an efficient and practical alternative to time-consuming UCS laboratory testing and offers substantial benefits for geotechnical investigations in complex urban environments. The findings are particularly valuable for geotechnical analysis and design in construction projects, where ANN models can effectively represent critical soil behavior.