PIML Model to Predict the Behavior of Axially Loaded Large Piles During Conceptual Design Stage
Abstract
The reliable prediction of axial load-settlement behavior of large-diameter drilled shafts during the conceptual design stage remains a challenging task in geotechnical engineering, primarily due to the limited availability of detailed site-specific parameters. This study introduces a novel application of physics-informed machine learning (PIML) techniques to estimate the hyperbolic coefficients (a & b) governing the load-settlement response of drilled shafts under axial compression, utilizing only basic soil classification data and pile dimensions as inputs. The primary objective is to develop robust, yet practically parsimonious, predictive models suitable for early-stage project design, thereby reducing dependency on extensive and costly site investigations. To achieve this, three machine learning algorithms—evolutionary polynomial regression (EPR), artificial neural networks (ANN), and genetic programming (GP)—were implemented and rigorously trained and validated using a database of 82 vertical pile load tests derived from the Cairo monorail project. A conventional linear regression (LNR) model was employed as a benchmark for comparative performance evaluation. The results demonstrate that the EPR and ANN models substantially outperform both GP and LNR approaches, achieving prediction accuracies of up to 93% and 85% for coefficients (a & b), respectively. Notably, the EPR model exhibited excellent performance with an R² of 0.975 and RMSE of 1.15 for coefficient (a), while the ANN model achieved an R² of 0.915 and RMSE of 0.7 for coefficient (b), underscoring the efficacy of PIML in capturing complex load-transfer mechanisms from minimal input data.