Abstract
The full-scale static pile loading test is without question the most reliable methodology for estimating the ultimate capacity of large diameter bored piles (LDBP). However, in most cases, the obtained load-settlement curves from LDBP loading tests tend to increase without reaching the failure point or an asymptote. Loading an LDBP until reaching apparent failure is seldom practical because of the significant amount of settlement usually required for the full shaft and base mobilizations. With that in mind, the supervised learning algorithm requires a huge labeled data set to train the machine properly, which makes it ideal for sensitivity analysis, forecasting, and predictions, among other unsupervised algorithms. However, providing such a huge dataset of LDBP loaded to failure tests might be very complicated. In this paper, a novel practice has been proposed to establish a labeled dataset needed to train supervised machine learning algorithms on accurately predicting the ultimate capacity of an LDBP. A comprehensive numerical parametric study was carried out to investigate the effect of both pile geometrical and soil geotechnical parameters on both the ultimate capacity and settlement of an LDBP. This study was based on field measurements of loaded to failure LDBP tests. Results of the 29 applied models were compared with the calibrated model results, and the variation in LDBP behavior due to change in any of the hyperparameters was discussed. Accordingly, three primary characteristics were identified to diagnose the failure of LDBPs. Those characteristics were utilized to establish a decision tree of a supervised machine learning algorithm that can be used to predict the ultimate capacity of an LDBP.