Optimized deep learning framework for reliable prediction of pavement subgrade CBR

California Bearing Ratio (CBR) is a fundamental parameter determining the load-bearing capacity essential for pavement subgrade design. This study introduces a deep learning (DL) based predictive framework for estimating CBR of pavement subgrade soil. Utilizing data from Sudan and South Sudan, the study develops and optimizes multiple DL architectures, including Feedforward Neural Networks (FNN), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRU). These models were systematically optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to improve predictive performance. The analysis revealed that the LSTM model achieved better performance with a Mean Absolute Error (MAE) of 5.21, Root Mean Squared Error (RMSE) of 7.31, and an R 2 value of 93.47%. Furthermore, permutation-based feature importance analysis identified Plastic Limit (PL) and Plasticity Index (PI) as the most influential factors affecting CBR, emphasizing the role of soil plasticity in subgrade stability. In addition, Partial Dependency Analysis (PDA) revealed significant interdependencies, particularly between Maximum Dry Density (MDD) and Optimum Moisture Content (OMC), reinforcing the impact of soil compaction characteristics on CBR. The study highlights the potential of DL frameworks as reliable and cost-effective alternatives to conventional CBR testing, providing a rapid and data-driven solution for pavement engineering applications.