LOAD-DISPLACEMENT RESPONSE OF LATERALLY LOADED CONCRETE PILES IN STIFF UNSATURATED CLAY USING MACHINE LEARNING ALGORITHMS

Abstract

This dissertation focuses on the utilization and comparison of machine learning (ML) algorithms to construct predictive models capable of predicting the horizontal failure force and displacement of horizontally loaded reinforced concrete (RC) piles embedded in stiff unsaturated clay. The proposed formulae are based on numerical models that were verified through on-site physical experiments and further validated using an out-of-sample dataset. This validation process demonstrated the model’s ability to accurately predict the mechanical response for different geometrical configurations of RC pile-soil systems while including various soil and concrete material properties. The significance of understanding the horizontal failure force and displacement of piles is an important parameter, particularly for structures such as wind turbine structures supported by piled foundations and buildings or bridges that are exposed to large lateral loads. As a result, it is important to develop accurate models that can predict the piles’ capacities, ensuring the design of safe, sustainable, and cost-effective structures. In order to highlight the importance of this work, even though there has been several attempts to develop methods for determining these aspects, there has been no prior attempt to validate the sophisticated 3D finite element (FE) techniques through experimental data. While comprehensive static finite element analysis (FEA) can be time-consuming and computationally demanding compared to the utilization of simplified design code formulae, it is useful to create models capable of accurately predicting and capturing the nonlinear force-displacement behaviour for various pile-soil configurations that include different concrete and soil material properties. In a recent pilot study conducted for the needs of this research work, an in-depth investigation of the nonlinear mechanical behaviour of a full-scale pile subjected to lateral loading was performed. This pile was embedded in stiff unsaturated clay, and advanced 3D numerical techniques were employed to replicate the physical experiment. Through extensive parametric investigations and addressing uncertainties beyond the scope of the physical experiment, a calibrated and optimized numerical model was created (Braun et al., 2023). Not only did this model accurately replicate the experimental findings, but it also provided insight into the respective failure mechanisms within both the soil and RC domains. This research provided valuable insights into the soil-structure interaction (SSI) phenomenon exhibited by the structure and enhanced the understanding of its behaviour under specific loading conditions. The primary objective of this research work is to construct a dataset by creating multiple SSI models that encompass a range of geometrical features along with a variety of concrete and soil material properties. A total of 6,561 numerical models were generated and analyzed, from which specific outcomes were extracted and stored in an Excel spreadsheet. This dataset served as input for an open-source ML software (nbml), wherein 85% of the dataset was used to train ML algorithms, and the remaining 15% was used to test the developed ML predictive models. The ML algorithms investigated in this study include linear regression (LR), polynomial regression with hyperparameter tuning (POLYREG-HYT), extreme gradient boosting with hyperparameter tuning and cross-validation (XGBoost-HYT-CV), random forests with hyperparameter tuning (RF-HYT), artificial neural networks by neighbourhoods (ANNbN), and deep learning artificial neural networks with hyperparameter tuning with message-passing interface (MPI) and horovod (DANN-MPIH-HYT). The performance of these ML algorithms is compared, and the accuracy of each predictive model is validated using results from out-of-sample data that were developed by assuming new geometrical features and varying concrete and soil material properties. For each of the ML algorithms, three different datasets were used to develop the proposed predictive models. All three datasets contained the same set of seven input features, including the pile diameter, embedded pile depth, reinforcement ratio, as well as the Young’s modulus and compressive strength for both the soil and concrete material. Furthermore, these datasets were designed to consist of three distinct target values: the horizontal failure force (Fx Max), horizontal failure displacement (Disp Max), and the horizontal displacement occurring at half of the horizontal failure force (Disp Fx2). The results indicated that RF-HYT outperformed all other ML models for Fx Max, achieving a mean absolute percentage error (MAPE) of 9.81% for the training set and 13.02% for the testing set. In contrast, for Disp Max, RF-HYT performed best in the training set, producing a MAPE of 19.83%, while XGBoost-HYT-CV produced the best results for the testing set with a MAPE of 24.20%. Similarly, for Disp Fx2, results mirrored those of Disp Max, with RF-HYT achieving a MAPE of 17.03% for the training set and XGBoost-HYT-CV providing a MAPE of 21.59% when testing the proposed predictive model. The models underwent additional validation using datasets that included out-of-sample results for each of the three target values. These datasets involved the creation of numerical models with different geometric configurations and material properties than those used during the training and testing phases, aiming to explore the numerical limitation of the proposed predictive models on out-of-sample data. It was found that POLYREG-HYT-3 stood out by achieving the lowest MAPE of 17.38% for Fx Max, the highest correlation of 53.49% for Disp Max, and the best correlation of 54.61% for Disp Fx2. DANN-MPIH-HYT, on the other hand, demonstrated exceptional performance with a 97.32% correlation for Fx Max. Meanwhile, ANNbN excelled in terms of MAPE, deriving a MAPE of 23.86% for Disp Max and 22.56% for Disp Fx2. The findings of this research indicate that the proposed formulae and the developed predictive ML models demonstrate good predictive capabilities when it comes to predicting the horizontal failure force and displacement of laterally loaded RC piles embedded in stiff unsaturated clay. This is the first research work of its kind in the international literature, where it successfully manages to address this type of structural problem that also deals with the phenomenon of SSI.