Design of Reinforced Concrete Isolated Footings Under Axial Loading with Artificial Neural Networks

Abstract

In engineering practice, the design of structural elements is a repetitive task that has proven to be difficult to fully automate. This is mainly because of the complex relations of the design variables and the multiple strength and other requirements that must be fulfilled based on code provisions to ensure safety and endurance, usually under extreme loading conditions or harsh environments. An optimal design can be defined as a set of values for the design variables that correspond to the optimal performance of the structural element in terms of a given criterion, usually related to the minimization of cost, while also satisfying all constraints related to strength, serviceability, functionality and safety. Such a design problem can be formally written as a function that maps a structural element, under given loading conditions, into a unique optimal design. In recent years, Artificial Neural Networks (ANN) have been adopted as a powerful strategy to solve complicated regression and classification problems where the underlying mapping function is generally unknown and difficult to formulate analytically. The ANN learns patterns contained in large databases through an automated process called training and uses that information to make highly accurate predictions. In the present study, a methodology that uses ANNs for the optimal design of structural elements is developed and applied to the design of reinforced concrete (RC) isolated footings under axial loading. First, a Genetic Algorithm is employed for the generation of the training dataset for the ANN, which includes RC footing designs that are optimized in terms of the material cost. Then, the ANN is trained and finally asked to produce new optimal designs for new sets of input parameters. Parametric tests are performed to determine the required size of the dataset and the most suitable network architecture. The results show that the accuracy of the prediction is very good, especially when larger datasets are used. It is shown that training an ANN to design structural elements is a viable option that gives acceptable solutions quickly, requiring extremely low computational cost. Furthermore, it is highlighted that good results can be obtained using a simple ANN architecture and a relatively small training dataset.