APPLICATION OF AN AI MODEL: DEEP NEURAL NETWORKS WITH BAYESIAN OPTIMIZATION FOR PREDICTION OF SUBGRADE RESILIENT MODULUS

Abstract

The resilient modulus (Mr) of subgrade is an important parameter in pavement structural design because it characterizes the subgrade under repeated vehicular traffic loading. It impacts pavement design life, layer thickness, economy, distress, and rehabilitation. Determination of Mr in the laboratory through repeated load triaxial test (RLTT) is considered the most reliable method; however, it is time-consuming, expensive, equipment-intensive, and requires expert personnel. Additionally, to overcome the RLTT limitations, widely used California Bearing Ratio (CBR)-based Mr–CBR correlations are often unreliable. Recently, machine learning (ML) techniques, like Artificial Neural Networks (ANNs), have been used to predict subgrade Mr using basic soil properties and stress state variables. However, manual trial-and-error tuning of ANN hyperparameters can be time-consuming and may overlook optimal configurations. This study used a dataset from literature containing 995 processed data points of A-4, A-6, and A-7-6 soils, with their Atterberg limits, dry density (DD), moisture content (MC), percent passing No. 200 sieve (P#200), and stress state variables, as model inputs to predict subgrade Mr using a Deep Neural Network (DNN) with hyperparameters optimized via Bayesian Optimization (BO). The BO framework included six (6) tunable hyperparameters. The optimized DNN showed strong performance, achieving R² values greater than 0.90 for training and validation sets, and 0.85 for the testing set, with corresponding MAE and MAPE values below 2.6 MPa and 10%, respectively. Sensitivity analysis showed that plasticity index (PI), deviator stress (σd ) and MC were the most influential, while DD showed minimal influence on subgrade Mr prediction.