The application of Finite Element Analysis (FEA) is an efficient approach for investigating engineering problems. In this method, the stiffness matrix of the structure, along with several time-consuming an‎d multi- step relationships, must be determined. This process is computationally intensive an‎d entails a significant computational cost. To overcome this challenge, machine learning techniques, particularly Artificial Neural Networks (ANNs), can be employed as an effective alternative. In this research, as an initial step for a specified number of elements, a Convolutional Neural Network (CNN) has been utilized. In this approach, an artificial neural network designed to predict the stiffness matrix replaces its direct calculation an‎d achieves high prediction accuracy. The network, trained on datasets derived from analyses based on the Finite Element Method (FEM), is capable of estimating the response of a finite element model with high speed an‎d acceptable accuracy. Since this study represents the initial stage of developing stiffness matrix prediction, a limitation on the number of elements an‎d nodes within a defined length range has been imposed to first demonstrate the feasibility of the proposed method. Integrating the Finite Element Method with the capabilities of neural networks not only reduces the computational burden an‎d solution time but also lays the groundwork for developing intelligent engineering tools applicable to industrial problems. Accordingly, in this thesis, the proposed method has been applied to the analysis of two-dimensional structures to preserve accuracy while facilitating the modeling an‎d analysis process. The primary objective of this research is to predict the stiffness matrix an‎d subsequently calculate local damage using an uncoupled approach for quasi-brittle materials, specifically concrete. In this framework, the stiffness matrix is predicted by the neural network, an‎d the damage calculation is incorporated into an in-house computational code. Quantitatively, a relative comparison index was defined as the ratio of each predicted entry to its corresponding entry in the calculated stiffness matrix, demonstrating strong agreement. A comparison between the damage calculation results obtained from the in-house code an‎d those from the Abaqus software indicated a small percentage of error, dependent on the damage magnitude an‎d its evolution stage. This error evidently arose from the constraints imposed on the in-house code.

محمد مشايخي

محمد سيلاني , عليرضا شهيدي ريزي