In this research, a computational method using artificial neural networks for predicting ductile damage based on the enhanced Lemaitre damage model is presented. Damage prediction is usually performed using the finite element method and includes various parts such as modeling the component geometry, different loads under different working conditions, selecting the appropriate damage model, creating a finite element model, and then implementing and extracting the results. This process must be repeated if any of the working conditions of the workpiece change, and it is necessary to continuously monitor the damage parameter, which can be a time-consuming and costly process. Today, with the increasing computational power of computer systems and the high speed of data-driven methods, there is a growing desire for data-driven mechanisms for damage prediction. In this research, using the data-driven artificial neural networks method, a different approach from the finite element method for predicting the stress tensor and damage was presented. Continuous damage models have received much attention in recent years, and much research has been conducted in this area. In this research, the improved Lemaitre damage model with Lode parameters was used to calculate the damage. The use of data-driven methods in complex phenomena requires access to a significant amount of data to properly train the network and achieve its optimal performance. The calculation of the stress tensor and damage in the material depends on the loading history of the material and is a function of the total strain and plastic strain values. Therefore, in order to reduce the number of simulations and extract more data, in each analysis, in addition to one point of the workpiece, the stress tensor data and the damage was used from different parts of the workpiece. The Lemaitre damage model, along with the Lode parameter, was implemented in a user-material subroutine (UMAT) in the Abaqus software. Subsequently, another subroutine was developed to automate the analysis in the Abaqus software. The analysis was performed for 400 samples of loading data, and the values of total strain, plastic strain, stress tensor, and damage amount were extracted. Next, the network structure, i.e., the number of layers, the number of neurons, and the corresponding model of the problem were introduced. Recurrent neural networks (gated recurrent units) have memory and can remember parts of the network's input parameter history. The parameters of neural network learning algorithm are determined through trial and error to achieve better performance. To assess the capability and accuracy of the neural network, a portion of the data that was not used for training the network was used to eva‎luate and benchmark the network. The error graphs obtained from the results of the artificial neural networks model and the results of the finite element model demonstrated the accurate performance and realistic prediction of the artificial neural networks.

محمد مشايخي

محمدرضا فروزان , عليرضا شهيدي ريزي