Session: Government Agency Student Posters

Paper Number: 173392

Microscale Strain Field Predictions From Grain Microstructure of Polycrystalline Metals Using Fully Convolutional Networks 

The heterogeneous microstructure of polycrystalline metals generates complex strain variations at the microscale under deformation, which ultimately control failure mechanisms. Strain accumulation at grain boundaries or within grain interiors can also result in the development of stress concentration sites, which can be the precursors to fracture and/or fatigue crack initiation. Thus, elucidating the accumulated strain associated with the heterogeneous microstructural features, e.g., grain size, shape, and orientation, has fundamental implications on the engineering properties of a polycrystalline metal at the macroscale. For the current presented work, we train fully convolutional networks (FCNs) on numerical datasets generated by crystal plasticity finite element simulations (CPFEMs), to predict the two-dimensional (2D) patterns of strain field variations,  , , and , (output) from grain orientation information (input) at the microscale, across a large subset of grain morphologies.  Previously applied FCN architectures correctly predicted the general patterns of strain distributions, but with performance that saturates quickly with increasing size of the training dataset. We overcome this limitation by training a MBRes-SkipNet which augments the traditional convolution architecture with modern architectural enhancements such as skip connections, depthwise-separable convolutions, residual functions, and inverted bottleneck convolution modules (MBConvs), reducing the number of trainable parameters and required computations, specifically floating-point operations (FLOPs), by 88% and 77% respectively. We present an ablation study of models trained on a single strain state which investigates the individual contributions of the aforementioned architectural enhancements along with the full combination of the FCN building block components based on performance metrics such as parameters, efficiency, root mean squared error (RMSE), R², structural similarity index (SSIM), and peak signal to noise ratio (PSNR). Results indicate that skip connections are critical to improving prediction quality, while depthwise separable convolutions play a key role in computational efficiency, enabling equal or superior performance at significantly reduced cost. A final MBRes-SkipNet architecture, trained on predominantly equiaxed grains with a fixed (lognormal) distribution of grain sizes under a small subset of macroscopic strain states, is capable of interpolation and limited extrapolation to other strain states. In addition, the MBRes-SkipNet is capable of predicting the microscale strain patterns across a wide range of grain sizes, grain distributions, and grain shapes without retraining, suggesting its generalizability to different grain architectures. Finally, we discuss the utility of transfer learning (TL) to reduce the amount of training data required to retrain the MBRes-SkipNet to predict the strain response of materials with different stress-strain behavior.

Presenting Author: William Noh University of Illinois at Urbana-Champaign

Presenting Author Biography: William Noh is a Ph.D. candidate in Aerospace Engineering at the University of Illinois Urbana-Champaign. His research focuses on applying convolutional and fully convolutional neural networks to multi-scale solid mechanics problems. His work combines deep learning with computational mechanics to improve accuracy and efficiency in object detection and dense prediction tasks, illuminating novel physical insights on known mechanical and material relationships.

Authors:

William Noh University of Illinois at Urbana-Champaign
Renato Bichara Vieira Pontifícia Universidade Católica do Rio de Janeiro
John Lambros University of Illinois Urbana-Champaign
Huck Beng Chew University of Illinois Urbana-Champaign