PSI - Issue 71

A Shivnag Sharma et al. / Procedia Structural Integrity 71 (2025) 469–476

473

and plastic deformation gradient F P ij tensors were used to train and test the model. The training was performed with no. of cells (where a single cell represents a single ConvLSTM layer) varying from ij and 2 nd Piola-Kirchhoff stress S

Fig. 1: Training vs validation loss with no. of epochs for different ConvLSTM cells

2 to 10 and the best suitable architecture with minimal training and validation loss was chosen to predict the results. Figure 1 depicts the semi-log plot of the training and validation loss for the component of the 2 nd Piola-Kirchhoff stress tensor in the loading direction (S 22 ). A total of 500 increments were trained for each of the variables separately and then the next 50 increments were predicted. The results obtained for the prediction of the S 22 component of the 2 nd Piola-Kirchhoff stress tensor and g α 1 , the first out of the 12 slip system resistances and the plastic deformation gradient component (F P 22 ) are shown in figures [2], [3] and [4] respectively. Figure 2(a) displays the comparative analysis of the results at 511 th increment and 2(b) shows for the 560 th increment for S 22 , similarly 3(a) and 3(b) shows for the first component of g α and 4(a) and 4(b) corresponds to F P 22 . The image in the extreme left corresponds to the original contour obtained from the simulations, the center one shows the predicted contour from ConvLSTM and the right image shows the heat map (error map). The mean pointwise error (MPE) for the shown data propagates for every variable and with every increment. However, it can be seen that the MPE is within acceptable limits for the shown variables. This paper highlights the use of ConvLSTM model for the full field prediction of the crystal plasticity simulations, hence accelerating the overall process. 4. Summary The study shown in this paper leverages the potential of LSTM networks used in conjunction with CNN as a spatio temporal prediction tool which can capture the future responses of a polycrystalline microstructure based on initial inputs from CPFEM for few time increments. In this paper, the crystal plasticity variables such as g α , S ij and F P ij were predicted successfully based on the previous time steps. These variables evolve with time and application of load, and describe the nature of the polycrystalline metals hence were chosen to evaluate the accuracy of this method. This study