A New View of Microstructure: Extracting Knowledge from Microstructural Images Using Computer Vision and Machine Learning

Hosted by Dr. Burton Patterson

Department of Materials Science and Engineering Seminar Series

Tuesday, January 10, 2017  – 4:05 – 4:55 p.m. – Florida Gym 270



Dr. Elizabeth A. Holm

Department of Materials Science and Engineering

Carnegie Mellon University



A New View of Microstructure: Extracting Knowledge from Microstructural Images Using Computer Vision and Machine Learning


Materials science is, at its core, the science and engineering of microstructure, and microstructural images are the foundational data of materials science. Indeed, materials scientists have developed numerous tools for acquiring, analyzing, and comparing microstructural images over the past century. When the catalog of interesting microstructural features is known – that is, when we know what we’re looking for – these techniques can segment, quantify, and compare microstructures with high precision. However, when we don’t know a priori which microstructural attributes are most important, these methods cannot help us.


Our goal is to develop a general method to find useful characteristics and relationships within and between micrographs without any assumptions about what features may be present and without significant human intervention. To achieve this, we capitalize on machine vision concepts to construct “microstructural fingerprints” that can be used to automatically find relationships in large and diverse microstructural image data sets. For example, the microstructural fingerprint can form the basis for a visual search engine in a database of micrographs. Likewise, it can be used with machine learning to train a computer to classify microstructures into groups by material system (e.g. ductile vs. brittle cast iron) or structural similarity (e.g. twinned vs. martensitic). Finally, the microstructural fingerprint can be correlated with quantitative microstructural metrics (e.g. particle size distribution), thus providing highly accurate image analysis without the need for segmentation or measurement. The results offer a new way to extract knowledge from 100 years of microstructural images in order to design new materials, optimize material processes, and tailor material properties.



Prior to joining CMU in 2012, Professor Holm spent 20 years as a computational materials scientist at Sandia National Laboratories, working on simulations to improve processes for lighting manufacture, microcircuit aging and reliability, and the processing and welding of advanced materials. Prof. Holm obtained her B.S.E in Materials Science and Engineering from the University of Michigan, S.M in Ceramics from MIT, and dual Ph.D. in Materials Science and Engineering and Scientific Computing from the University of Michigan. Active in professional societies, Prof. Holm has received several honors and awards, is a Fellow of ASM International, 2013 President of The Minerals, Metals, and Materials Society, an organizer of several international conferences, and has been a member of the National Materials Advisory Board. Prof. Holm has authored or co- authored over 110 publications.