Applying Machine Learning for SPT-N Prediction from Raw CPT Data-Paper

Paper

Pdf Summary

This study evaluates machine-learning models for predicting corrected Standard Penetration Test (SPT-N) values from raw Cone Penetration Test (CPT) data. The dataset came from the I-405 freeway expansion investigation in Orange County, California. Researchers matched CPT soundings with nearby SPT boreholes (within 50 ft), aggregated CPT measurements across the depth intervals represented by SPT-N values, and retained 43,520 records from 90 CPT–borehole pairs. Model inputs were depth, cone-tip resistance (qc), sleeve friction (fs), and pore pressure (u2). Models were assessed for the complete dataset and separate sandy- and clayey-soil subsets.<br /><br />Random Forest (RF) and XGBoost models were trained using an 80/20 train–test split. RF performed better across all datasets. Its R² scores were 0.917 for the complete dataset, 0.942 for sandy soils, and 0.835 for clayey soils, compared with 0.787, 0.857, and 0.685, respectively, for XGBoost. Sandy soils were most predictable, while clayey soils showed greater complexity and lower predictive accuracy. RF feature-importance analysis identified depth as the strongest predictor for the soil-specific datasets and qc as the strongest predictor for the complete dataset; fs was least influential overall.<br /><br />The findings show the potential of machine learning to estimate corrected SPT-N values when CPT data are available. However, performance was evaluated using data from one project and a random train–test split, so the models’ transferability to other sites remains uncertain. The authors recommend expanding the dataset, exploring hybrid models, and incorporating additional geotechnical parameters to improve reliability across diverse soil conditions.

Keywords

Standard Penetration Test

SPT-N prediction

Cone Penetration Test

CPT data

machine learning

Random Forest

XGBoost

geotechnical engineering

soil classification

feature importance