Reducing the solution processes in structural optimization problems using neural networks

In the past twenty years, artificial intelligence (AI) research has significantly entered our scientific and daily lives, although its origins date back to the 1950s. The rapid resolution of hardware limitations from its discovery period has enabled the development and widespread use of AI solutions in many fields. These advancements have allowed algorithms, which are mathematically intensive and repetitive, to be solved more quickly. In particular, the computational intensity of traditional methods in structural optimization complicates the analysis and optimization of complex structures. Structural optimization involves determining the design variables of structural systems to maximize certain objective functions or meet specific constraints. This plays a crucial role in enhancing the mechanical performance, energy efficiency, and cost-effectiveness of structures. Advancing optimization algorithms provide the capability to design lighter and stronger structures, making them applicable to more complex problems. Metaheuristic methods are widely used for structural optimization and are effective in improving the performance and efficiency of structures. However, when used in conjunction with structural analysis methods such as the finite element method (FEM), the solution time can increase. This is due to the intensive computation required by FEM to accurately model the behavior of structures. This thesis aims to reduce the processing times of structural optimization algorithms. To achieve this, an artificial neural network (ANN) has been developed to improve the "Evaluation Phase," which is the common and most time-consuming step of optimization algorithms. The focus is on reducing the time generated during the displacement calculations, which is the main step of the structural analysis process. While displacement calculations performed by traditional methods can be time-consuming and complex, the combination of ANN and optimization algorithms significantly reduces these times. The ANN-based approach developed in this study aims to offer faster and more efficient solutions by overcoming the limitations of traditional methods. The ANN-FEM solution is independent of the type of structural problem, providing an applicable solution for different types of structures. The ANN-based matrix-free method developed focuses on predicting the displacement vector by bypassing the Gauss elimination processes, which consume the most time and memory in FEM analysis. To integrate ANN predictions into the optimization process, three new hybrid Biogeography-Based Optimization (BBO) search methods (YT-1, YT-2, and YT-K) have been developed. These methods strategically use ANN predictions to guide the optimization process and aim for faster convergence. Tests conducted within the scope of the study showed that the ANN-based method significantly reduced computation time compared to FEM for different types of structures. Although the accuracy level of ANN does not reach FEM, its speed advantage becomes a significant factor, especially in scenarios requiring real-time applications or rapid design iterations. These findings demonstrate that ANN-based methods provide a viable and efficient alternative for structural optimization and open a promising path for further research.