Machine Learning Applied to the Design of Drilled Shafts: Prediction of the Nominal Axial Resistance in Cohesive-Dominant Soils using Artificial Neural Networks
Loading...
Authors
Agbemenou, Atsou Komla Herve
Issue Date
2023
Type
Thesis
Language
Keywords
artificial neural network , drilled shafts , machine learning , Nevada deep foundation load test database , nominal axial resistance
Alternative Title
Abstract
The use of drilled shafts as deep foundation systems in supporting large civil engineering structures such as highways, bridges, retaining structures and high-rise buildings is rapidly increasing. These foundations are typically designed to transfer mainly axial and lateral loads to the underlying soil through skin friction and end bearing. However, the conventional design methods were found to be not applicable to every soil type due to the high variability in the soil properties; and not taking into consideration certain sources of uncertainty such as the construction methods (considered in this study). These uncertainties can introduce nonlinearity to the analysis, leading to overestimation or underestimation of the resistance, which can be costly and/or pose safety concerns. The consequences (casualties) that can be associated with the failure of such structures are not negligible, therefore an accurate prediction is of great importance. To address this issue, Machine Learning techniques such as Artificial Neural Networks (ANNs) ‒ composed of interconnected layers of processing elements called neurons, using mathematical functions to process data mimicking the human brain ‒ have been increasingly used in civil engineering applications to solve nonlinear problems. The objective of this study is to make use of the ANN computational capabilities for predicting more accurately the nominal axial resistance of drilled shafts, in cohesive dominant soils. Two studies were conducted: the prediction of the total nominal axial resistance and the nominal side resistance. The research was based on field load tests collected from the extended version of the Nevada Deep Foundation Load Test Database. 45 total load tests were divided into 85% for training and 15% for validation using the root-mean-squared error, the mean absolute error, and the R-squared as evaluation metrics. Such models could serve as early-stage, supplemental tools for first-order estimation of the resistance and/or optimizing designs. ANNs can be utilized to investigate certain sources of uncertainty associated with the design (the construction methods for instance), mitigating the risks of over or underestimation, which can ultimately lead to more cost-effective and safer projects.