PSI - Issue 71

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

470

(2019) in their study. Many researchers have demonstrated the efficacy of CPFEM including Kalidindi et al. (1992), but its application still remains restrictive due to its complexity that includes the solution of PDEs which are highly non-linear and results in higher consumption of computational time especially for large number of Degrees of Freedom (DoF). Artificial Intelligence/Machine Learning (AI/ML), a fast-evolving technology, can be used as an aid to accelerate CPFEM simulations and hence reduce the computational cost. AI/ML is a data driven technique that takes valuable insights from data when appropriately trained and can predict the outcome almost accurately. Due to its strong predictive and classification capabilities, ML has drawn the attention of many material scientists/engineers for its use in the material science domain. Ghaboussi et al. (1991) proposed the use of Artificial Neural Networks (ANNs) as an alternative to traditional mathematical models of material behavior, marking the beginning of machine learning applications in computational materials science. Brahme et al. (2009) highlighted the ability of ANNs to forecast the cold-rolling fiber texture of steel by using initial texture data and rolling reduction parameters. Mozaffar et al. (2019) made significant strides by employing a sequence-learning model to predict path-dependent material responses, including stresses and plastic energy, for two-dimensional Representative Volume Elements (RVEs). Building on these advances, Li et al. (2019) introduced a ML model that efficiently predicted material hardening, offering a more effective alternative to traditional methods based on calibrated yield function. Further research by Gorji et al. (2020) demonstrated the impressive predictive capabilities of various ANNs, including feed-forward neural networks and gated-recurrent unit networks, in replicating the anisotropic Yld2000-2d model's predictions for different loading conditions, both monotonic and non-monotonic. Muhammad et al. (2020) presented a ML framework designed to predict local strain distribution, plastic anisotropy, and failure during the tensile deformation of 3D-printed aluminum alloy. Classification ML techniques were employed by Mangal and Holm (2018) to forecast the development of stress hotspots in polycrystalline materials subjected to uniaxial tension, with a focus on face-centered cubic (FCC) and hexagonal close-packed (HCP) structures. An ANN-based framework was introduced by Ali et al. (2019) to predict both the flow stress and texture evolution of polycrystals under multiaxial and non-proportional loading conditions. The limitations of the literature presented above was the lack of predictability outside the training model, since the ML models performs poorly on extrapolation tasks. Ibragimova et al. (2020) used ANNs and CPFEM to predict stress–strain behavior and texture evolution in face centered cubic (FCC) materials. The framework shows high accuracy and efficiency across diverse loading conditions, effectively predicting complex strain paths without extensive training on all scenarios. A combined Convolutional Neural Networks (CNNs) with CPFEM to predict stress and strain under uniaxial tension with high speed and accuracy was proposed by Ibragimova et al. (2020). The CNN model, trained on synthetic microstructures, showed excellent agreement with CPFEM and effectively predicted stress and strain in new microstructures like AA5754 and AA6061, demonstrating its accuracy and computational efficiency. Ahmad et al. (2023) combined autoencoder and Convolutional Long Short-Term Memory (ConvLSTM) models to accelerate phase-field simulations of microstructure evolution while maintaining high spatial resolution. Trained on microstructures from 10 compositions, the model predicts future microstructure frames based on past frames, significantly speeding up simulations and reducing computational costs. A similar technique involving the use of autoencoder ensembled with a Ridge regression model was proposed by Agrawal et al. (2024) to predict the plastic deformation gradient of a polycrystalline material based on CPFEM simulations. Zhou et al. (2024) used a Recurrent Neural Network (RNN) in conjunction with CPFE model to accurately predict the stress-strain response and orientation evolutions at both the single grain and aggregate scales. The present work attempts to extend the methodology proposed in Ahmad et al. (2023) to predict the full-field deformation behavior of a polycrystal with FCC crystal structure represented using crystal plasticity. The presence of sharp discontinuous boundaries, multiple state variables and strong anisotropy, makes the application of ConvLSTM