۱- Introduction

۲- Methods for generating data sets

۳- Artificial neural network (ANN) model

۴- Results and discussion

۵- Summary and conclusions

References

Abstract

Machine learning techniques are widely used to understand and predict data trends and therefore can provide a huge computational advantage over conventional numerical techniques. In this work, an artificial neural network (ANN) model is coupled with a rate-dependant crystal plasticity finite element method (CPFEM) formulation to predict the stress-strain behavior and texture evolution in AA6063-T6 under uniaxial tension and simple shear. Firstly, stress-strain and texture evolution results from the crystal plasticity simulations were verified with experimental observations for AA6063-T6 under simple shear and tension. Next, results from crystal plasticity simulations were used to train, validate and test the ANN model. The proposed ANN framework, was successfully applied on single crystal simulation results to predict stress-strain and texture data. Then, the proposed ANN framework was applied to predict the stress-strain curves and texture evolution of AA6063-T6 during uniaxial tension and simple shear. The flexibility of the proposed ANN model was also tested, for simple shear, with a completely new data set and the predicted results showed excellent agreement with corresponding crystal plasticity simulations. Finally, the predictive capability of the proposed model was further demonstrated by successfully validating the ANN model for non-proportional loading paths such as uniaxial tension followed by simple shear and simple shear followed by tension. The results presented in this research clearly demonstrate that the proposed ANN model provided significant computational time improvements without any major sacrifice in accuracy.

Introduction

Interest in Machine learning (ML) models to understand and predict the material behavior and properties has grown rapidly in recent years. Artificial neural network (ANN) is one group of algorithms used for Machine learning (ML). ANN frameworks model the data using mathematical models that try to simulate a neuron in the brain. In the brain, the process of knowledge acquiring, or learning, for a particular task occurs through experience and does so till we reach the required objective. Such approaches attempt to establish relations from simulated or experimental data using computing systems with learning capabilities. Similar to conventional simulation tools such as FE etc, computing systems such as ANN provide an alternative way to accurately assess the material properties as well as to design new materials in an accelerated manner. Some applications of ANN modelling in the context of material science have been evaluated by Bhadeshia (1999) and Raabe (2002). Specifically, ANN models could be used with complex problems with nonlinear correlations and interactions between inputs and outputs.