Machine Learning Predictions for Microscale Strain Accumulation in Polycrystalline Metals

Data-driven approaches based on machine learning (ML) have become increasingly popular for the study of the mechanical response of material behavior as both numerical and experimental novel methodologies have provided a wealth of big-data information for materials. On the experimental side, two dimensional (2D) and three dimensional (3D) full-field techniques such as digital image correlation (DIC) and digital volume correlation (DVC) provide datasets comparable in size and fidelity to numerical techniques such as finite elements (FEA) or molecular dynamics (MD). However, one criticism faced by data-driven approaches used to “fit” such big-data is that they often ignore or oversimplify the underlying physics and as such may not be suitable for predictively analyzing complex phenomena for which physics-based multiscale models are often developed. Failure event prediction in particular is especially challenging because failure is usually driven by local events though it is triggered by global loading, thus by its nature necessitating a multiscale approach. As a result, in this work we try to combine the ML “convenience” of bypassing explicitly modeling with the “physical accuracy” of a multiscale model by using ML within a physics-driven framework by combining experimental data as both training and constraints for the ML framework.<br/> <br/>In particular, we wish to use microstructural strain accumulation in metals as a predictor of ultimate failure whether through plasticity, creep, fracture or fatigue loading. Depending on material, grain boundary vs. grain interior (i.e., “mantle vs. core”, respectively) strain accumulation can serve as a driver for failure nucleation/initiation at the microscale. Experimental measurements based on high-resolution DIC (HiDIC) techniques which we have developed provide microscale strain fields with sub-grain level resolution that are then used to train artificial and/or convolution neural networks (ANNs and CNNs). Microstructural quantities such as grain misorientation, Schmid factor, and grain morphology, are used to identify strain hot-spot locations. Specifically, using a fitting neural network trained by experimental HiDIC we found that in an austenitic steel alloy the grain boundary inclination angle shows significant correlation with local stain accumulation in the mantle region. Considering the interior (core) strain build-up, we combined HiDIC measurements over large areas of a microstructure with crystal plasticity simulations to identify critical grains (or grain clusters–defined as neighboring grains similarly oriented) which would exhibit critical stain accumulation. Averaging the response, either within a grain or over several grains, allows for the development of an approach that can be more easily implemented at higher length scales to provide predictions of local failure using more global quantities (i.e., a reduced-order model of sorts) based on the knowledge gained through the multiscale ML efforts.