Machine learning–based estimation of subgrade resilient modulus using sensor signals from in-situ modulus detector

Evaluating the resilient modulus of subgrade soils is essential for pavement design. However, conventional field and laboratory testing methods remain time-consuming and subject to measurement variability. This study introduces a machine learning approach for estimating the in-situ resilient modulus of subgrades using sensor signals recorded from an in-situ modulus detector (IMD). Specifically, the acceleration and force signals generated during IMD penetration at an unpaved test site are used as the input features. The methodology involves assembling a database of 885 observations from 12 IMD tests and applying signal segmentation and downsampling to the recorded signals. Four machine learning models—random forest, gradient boosting, extreme gradient boosting, and K-nearest neighbors—are trained across three input-signal scenarios. The resilient modulus is predicted using acceleration signals, force signals, and a combination of both, and model performance is evaluated using standard regression metrics. The results show that the extreme gradient boosting model achieves the highest accuracy when both acceleration and force signals are used, exhibiting R² values of 0.90 with the raw segmented signals and 0.92 with the downsampled signals. The analysis indicates that appropriate segmentation and downsampling substantially enhance prediction accuracy while improving computational efficiency. The findings demonstrate that machine learning models trained on IMD sensor signals constitute a practical framework for rapid and automated evaluation of the in-situ resilient modulus of subgrade soils.