Characterization and reconstruction of 3D stochastic microstructures via supervised learning

R. Bostanabad, W. Chen, D. W. Apley*

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

22 Scopus citations

Abstract

The need for computational characterization and reconstruction of volumetric maps of stochastic microstructures for understanding the role of material structure in the processing–structure–property chain has been highlighted in the literature. Recently, a promising characterization and reconstruction approach has been developed where the essential idea is to convert the digitized microstructure image into an appropriate training dataset to learn the stochastic nature of the morphology by fitting a supervised learning model to the dataset. This compact model can subsequently be used to efficiently reconstruct as many statistically equivalent microstructure samples as desired. The goal of this paper is to build upon the developed approach in three major directions by: (1) extending the approach to characterize 3D stochastic microstructures and efficiently reconstruct 3D samples, (2) improving the performance of the approach by incorporating user-defined predictors into the supervised learning model, and (3) addressing potential computational issues by introducing a reduced model which can perform as effectively as the full model. We test the extended approach on three examples and show that the spatial dependencies, as evaluated via various measures, are well preserved in the reconstructed samples.

Original languageEnglish (US)
Pages (from-to)282-297
Number of pages16
JournalJournal of Microscopy
Volume264
Issue number3
DOIs
StatePublished - Dec 1 2016

Keywords

  • 3D
  • Characterization and reconstruction
  • statistical equivalency
  • stochastic microstructure
  • supervised learning

ASJC Scopus subject areas

  • Pathology and Forensic Medicine
  • Histology

Fingerprint Dive into the research topics of 'Characterization and reconstruction of 3D stochastic microstructures via supervised learning'. Together they form a unique fingerprint.

Cite this