Machine learning approaches for elastic localization linkages in high-contrast composite materials

Abstract

There has been a growing recognition of the opportunities afforded by advanced data science and informatics approaches in addressing the computational demands of modeling and simulation of multiscale materials science phenomena. More specifically, the mining of microstructure-property relationships by various methods in machine learning and data mining opens exciting new opportunities that can potentially result in a fast and efficient material design. This work explores and presents multiple viable approaches for computationally efficient predictions of the microscale elastic strain fields in a three-dimensional (3-D) voxel-based microstructure volume element (MVE). Advanced concepts in machine learning and data mining, including feature extraction, feature ranking and selection, and regression modeling, are explored as data experiments. Improvements are demonstrated in a gradually escalated fashion achieved by (1) feature descriptors introduced to represent voxel neighborhood characteristics, (2) a reduced set of descriptors with top importance, and (3) an ensemble-based regression technique.

Context Aware Machine Learning Approaches for Modeling Elastic Localization in Three-Dimensional Composite Microstructures

“…In our designed two-stage system, the first stage attempts to find the contexts in data, while the second stage builds context specific learning models. We compare the results from this new data model with benchmarks from previous work [8] and demonstrate that the two layer data modeling scheme provides a viable approach for capturing the elastic localization linkages in high contrast composites.…”

“…Most of these efforts were built on model forms suggested by Kroner's expansions obtained from classical statistical continuum theories [45,46]. In a recent study [8], we explored alternative approaches that were based completely on machine learning techniques that demonstrated significant promise for high contrast composites. This is particularly signficant because Kroner's expansions are known to be applicable to only low to moderate contrast composites [45,46].…”

“…The result of this approach was a single-agent model learned from all available MVEs. The target of this work is to improve the approach presented in [8] by increasing the efficiency in building the model and increasing the accuracy of predicting the linkages.…”

“…In other words, m s,0,0 is considered the focal voxel and p s is its corresponding local response. In our prior work [8], the main protocol we followed for predictive modeling was composed of two key processes: (i) feature extraction, and (ii) construction of a regression model where the local neighborhood information m s,l,t and the local response p s were used as input and output, respectively. The utility of each process was evaluated and demonstrated through an ablation test with various data experiments.…”

“…Historically, the multiscale materials modeling efforts have addressed either homogenization (communication of information from the lower length scale to the higher length scale) [30][31][32][33] or localization (communication of information from the higher length scale to the lower length scale) [8,13,14,[34][35][36]. Although both homogenization and localization have been studied extensively in literature using physically based approaches [31][32][33]37], recent work has identified the tremendous benefits of fusing these approaches with data-driven approaches [8,13,14,30,[34][35][36]. However, most of the prior effort has only addressed a limited number of the multiscale features.…”