Develop Machine Learning Models to Establish the Load-Settlement Curve of Piles from Cone Penetration Test Data-Paper

Paper

Pdf Summary

This study evaluates machine learning (ML) models for predicting the load–settlement behavior of axially loaded piles from cone penetration test (CPT) data. The researchers compiled 64 static load tests on square precast prestressed concrete piles from 34 Louisiana sites, alongside nearby CPT profiles. Digitized and normalized load–settlement curves provided 905 data points.<br /><br />Three models were developed: artificial neural networks (ANN), random forests (RF), and gradient boosted trees (GBT). Inputs included pile dimensions and embedded length, normalized settlement, and CPT measurements—corrected cone resistance and sleeve friction—summarized across five pile-shaft segments and the zone near the pile tip. Model predictions were assessed using training and testing data, then compared with measured curves and predictions from API and USACE load-transfer methods.<br /><br />All three ML models performed well on the testing data. ANN and GBT were the strongest overall: ANN achieved the lowest testing RMSE (0.022 ton/ft²) and a high R² of 0.98, while GBT had the lowest coefficient of variation (0.21) and predictions averaging about 2% below measured loads. ANN averaged about 6% underprediction. RF performed less well than the other models.<br /><br />For four test piles, all ML approaches captured measured load–settlement behavior more accurately than the traditional API and USACE methods, which substantially underestimated the curves. The authors conclude that ANN and GBT are promising tools for estimating pile load–settlement curves from CPT data, while noting that the study is based on a specific regional database and pile type.

Keywords

machine learning

load-settlement behavior

axially loaded piles

cone penetration test

artificial neural networks

random forests

gradient boosted trees

pile foundations

CPT data

static load tests