Interpretable Deep Learning Models for Predicting the Bearing Capacity of Strip Footings on Cohesionless Sloping Soil

Accurately predicting the bearing capacity of shallow foundations on sloping soil is a critical challenge in geotechnical engineering due to the complex interactions between soil properties and structural parameters. This study employs three deep learning architectures: Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and Residual Networks (ResNet) to enhance predictive accuracy and interpretability. A dataset comprising 792 finite element simulations from Plaxis 2D was used to train and validate these models. Performance was evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and R2 metrics. ResNet demonstrated the most accurate and stable predictions, attaining the lowest test error and the strongest cross-validation performance, outperforming both CNN and DNN architectures. A depth-wise ablation benchmark confirmed that skip connections prevent degradation and preserve performance at scale. CNN achieved competitive performance but was notably affected by input ordering, revealing its structural sensitivity to tabular data. To improve model transparency, SHapley Additive exPlanations (SHAP) were employed to analyze feature importance. Results revealed that the footing depth-to-width ratio (Df/B) had the most significant impact on bearing capacity predictions, followed by slope inclination (β) and internal friction angle. The findings demonstrate that ResNet is a reliable approach for geotechnical predictions and highlight the importance of integrating explainable AI techniques to enhance trust in data-driven decision-making. Future research should explore larger datasets and hybrid deep learning frameworks to further refine predictive performance.
The correct input format is Df/B, g,b/B, f,b,Ngq