A Computationally Efficient Framework for Optimizing the Designs of Reinforced Concrete Special Moment Resisting Frames

Abstract

Structural design of reinforced concrete special moment resisting frames requires analyzing multiple design alternatives with the goal of obtaining the most economic, safe, and constructible solution. The purpose of this research is to reduce the computational burden of structural analysis and to automate the iterative design process of reinforced concrete special moment resisting frames.

To reduce the computational burden of structural analysis Artificial Neural Networks (ANNs) are first investigated, and a novel algorithm known as the Weight Matrix method is then explored. Artificial neural networks are computationally efficient mathematical models that can be trained to simulate computationally costly processes. For this research, ANNs are used as complete\partial replacements for the structural analysis of reinforced concrete 2D frames. The results show that using ANNs as surrogates for structural analysis is impractical since the training time grows exponentially with the size of the frame. On the other hand, the Weight Matrix method relies on linear algebra, vector operations, and sparse matrix algorithms to produce a framework that allows for efficiently updating the global stiffness matrices of plane frame structures when needed. Results show that the Weight Matrix method is more computationally efficient than the conventional method of constructing the global stiffness matrices of 2D frames.

Furthermore, the genetic algorithm (GA) was used to simulate the iterative design process. GAs are population based optimization methods founded on the concept of natural selection. In each iteration, the structural responses and costs are evaluated for a population of design alternatives, and the designs with the most favorable costs and structural responses are selected and combined to produce a new population for the next iteration. However, like many population-based optimization algorithms, GAs suffer from the issue of premature convergence. During premature convergence, the population collapses to copies of a single individual, and the algorithm becomes trapped. To improve the convergence of the GA, this research explored population diversity maintaining mechanisms found in the GA literature. Finally, this research will investigate the constructability of the solutions obtained from GAs by considering the effect of buildability factors (e.g. formwork surface area, quantity of reinforcement, concrete volume …) on the total cost of the design alternatives.