The use of artificial neural networks for modelling buildability in preliminary structural design

The Construction Industry has long been criticised for the manner in which parties involved in a construction project communicate. Since the early 1960s a number of reports, commissioned by the U.K. governmental agencies, have highlighted a fundamental malaise in the industry: the Jack of integration between design and construction processes. This chronic enigma has manifested itself in cost overruns, prolonged durations, poor quality and complex designs. Buildability and Design for Construction have emerged as key drivers for improving project objectives. Despite considerable progress in identifying the generic concepts ofbuildability, similar progress in its implementation, particularly during the preliminary structural design stage is still in its infancy. This implementation requires a framework for knowledge acquisition of construction information for use by designers. However, there is currently minimal documented experience in capturing technical information, construction expertise and knowledge implicit in previously completed projects for the benefits of new ones. The focus of this research is to develop computerised models for acquiring construction knowledge from past projects to integrate buildability considerations into the preliminary structural design process. A novel artificial intelligence approach has been adopted in this study. Five Artificial Neural Network models have been developed. These allow the generation of an expeditious solution for given sets of design and buildability constraints. Once information is entered into the developed models, a recommendation of which structural scheme to choose is generated instantaneously. Thus, valuable design time is released allowing designers the opportunity to invest this in performing other equally important design tasks. The input information to the models consists of site-related information including site access; availability of working space; and speed of erection, and conceptual design information including type of building; and number of storeys. Four of the five models achieved a high level of accuracy in the range of81.25% to 94.74%. Preliminary structural design is a complex process which relies heavily upon past experience and intuition. These characteristics cannot be represented by the use of conventional computational techniques and only those that are capable of generalising the knowledge implicit in past projects can be of real benefit. In this research, it has been demonstrated that the aforementioned characteristics of structural design fall naturally into the Artificial Neural Networks' problem domain.