The present study addresses the gap in geotechnical prediction by applying a deep learning model trained on 721 samples combining laboratory and field data (CPT and SPT) from Al-Furat Al-Awsat Technical University and in-situ projects in Najaf, Iraq. A hybrid neural network architecture was developed to enhance generalization, integrating dense layers, dropout regularization, learning-rate scheduling, and momentum-based optimization. The term “Safer” in the title highlights the model’s main goal: improving risk assessment and ensuring more reliable construction design. Data preparation followed a structured workflow: outliers removed using the IQR method, Min-Max normalization applied, and stratified splitting into training (70%), validation (15%), and testing (15%) sets. The network consisted of three dense layers (256, 128, 64 neurons) with ReLU activation. Adam optimization (learning rate 0.001) and early stopping were used to prevent overfitting. The Results showed a 24.7% reduction in RMSE, achieving 44.2 kPa compared with the Random Forest model’s 58.7 kPa, with R² = 0.89. The model performed particularly well for clay (RMSE = 38.4 kPa) and silt (RMSE = 41.2 kPa), while organic soils remained challenging (RMSE = 53.6 kPa) due to data inconsistency and sampling bias. Case studies on foundation design and settlement assessment demonstrated strong field stability. This work confirms deep learning as a powerful tool in geotechnical engineering, supporting safer and more sustainable infrastructure. Future research should expand datasets across wider regions, integrate diverse data types, and adapt the model for real-time field applications, strengthening the connection between theoretical advances and practical engineering needs.
