Prediction of Resilient Modulus Value of Cohesive and Non-Cohesive Soils Using Artificial Neural Network

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence
• Pavement Engineering

Background:

• The resilient modulus (M) is crucial for pavement design, but traditional lab testing is time-consuming.
• Accurate M prediction is vital for optimizing pavement performance and durability.

Purpose of the Study:

• To apply Artificial Neural Networks (ANNs) for predicting the resilient modulus (M) of subgrade and subbase soils.
• To evaluate the accuracy and interpretability of ANN models in M prediction.
• To explore ANNs as a practical alternative to conventional laboratory testing.

Main Methods:

• A dataset of 1683 M observations was used with an ANN model incorporating eight input variables.
• Model optimization was performed using a quasi-Newton method.
• SHAP (SHapley Additive exPlanations) analysis was employed for model interpretability.

Main Results:

• The ANN model achieved a high coefficient of determination (R^2) of 0.9613.
• The model demonstrated low error rates on both selection and testing datasets.
• SHAP analysis identified saturation ratio, plasticity index, and soil gradation as key influential parameters.

Conclusions:

• ANNs provide a practical and accurate tool for estimating resilient modulus.
• This AI-driven approach can enhance pavement design processes and improve road structure efficiency.
• The study validates ANNs as a viable alternative to laboratory testing for M determination.