Predicting and Evaluating Load-Displacement Trends of Helical Piles Using Machine Learning–Based Algorithms

Abstract

Evaluating the load-displacement trend of helical piles is essential to meet strength and serviceability standards for pile analysis and design. While pile load tests are the most accurate method, their cost and time constraints often lead in practice to the routine use of load transfer methods. This study aims to explore the feasibility of employing various machine learning–based algorithms to assess the load-displacement trend of axially loaded helical piles, using an existing database. Multiple machine learning models, including multi linear regression, K-nearest neighbors, random forest (RF), extreme gradient boosting (XGB), and artificial neural networks, were developed using data from 98 helical piles with a depth of 2.4 m to 16 m from 14 separate sites and corresponding in situ test data. Statistical evaluation was conducted to assess the performance of these models, comparing their predictions to measured curves from pile load tests. Based on the predictive analysis results, the average area under the regression error characteristic curve for XGB and RF was 0.035 and 0.043, respectively. The performance errors for XGB and RF were 0.01 and 0.05, respectively. The root mean squared error (RMSE) values for XGB and RF on the training data set were 0.083 and 0.1, respectively, while the RMSE values on the testing data set are 0.08 and 0.12.