A Cone-Penetration-Test Inversion Model Trained by Machine Learning

Volume-averaging effects significantly impact the accuracy of cone penetration test (CPT) measurements in layered soil profiles. This study introduces a machine learning (ML)-based inversion model designed to correct these effects, trained on an extensive dataset that combines results from existing literature and new virtual calibration chamber experiments using the discrete element method. The inversion model achieves a mean absolute percentage error of 13% on unseen data, outperforming existing methods to a statistically significant degree. Its limitations include: (1) reduced accuracy in very thin layers and in conditions beyond the training set’s parameter space; (2) interpretability challenges typical of ML-based approaches; and (3) uncertainties inherent to virtual methods’ abilities to represent physical reality. Despite these limitations, this study demonstrates the feasibility of a novel approach to developing CPT inversion models and results in an easy-to-use model that warrants adoption and further scrutiny.