Process metallurgists (like me) use math, physics and chemistry to describe what happens during metal manufacturing. Other researchers use math, statistics and computer science to study and optimize manufacturing at a larger scale.
While there has always been some overlap, it seems to me the micro and the macro views of manufacturing are more rapidly starting to unify as part of the “digital transformation.” Many terms associated with this idea: advanced manufacturing, Industry 4.0, data mining, big data, AI, machine learning, optimization, Internet of Things (IoT), etc. The general goals are to increase process knowledge, process automation and product quality while decreasing costs.
As the macro and micro views of manufacturing continue to unify, we all need to keep learning. For example, I do not have a strong background in statistics or programming, so it was a challenge to try to understand fields like machine learning. My opportunity to start learning came a few years ago when a new master’s student joined my group. We started applying methods in machine learning to analyzing nonmetallic inclusions in steels.
Inclusions are small (~ 1-100 µm) oxide, sulfide or nitride particles. Hundreds to thousands can be observed on a 1 inch2 area of steel. Although small, inclusions can have major impacts on the processing and properties of steel. One of the challenges in dealing with inclusions is their complexity. They are often comprised of multiple elements. They can be solid, liquid or both. Their shapes can be geometrically complicated.
The complexity of inclusions means that it is challenging to describe them. We can visualize some variables, but we are limited because we visualize on two-dimensional screens. Machine-learning methods are useful here because they perform tasks like regression, classification and clustering on datasets with both large numbers of variables and large numbers of observations.
I have begun integrating these new techniques into my own research on steel inclusions in two different projects. The first is with clustering, where we have tried to group inclusions with similar chemical compositions. This is currently done with user-defined rules, but these must be developed by an expert and are static.
Clustering algorithms can automatically identify groupings in each sample. We also used clustering algorithms to find physical clusters of inclusions (i.e., groups of smaller inclusion particles that have joined together). These clusters have irregular shapes and can sometimes be hundreds of micrometers in size. They are very detrimental to steel processing and product quality.
The figure shows an example of automatically identified clusters and an image of one particularly ugly one. Eventually, we want these methods to be tools that industry engineers can use to automate the analysis of nonmetallic inclusion data.
The second area of research has been a collaboration with Professor Elizabeth Holm to use computer vision to classify inclusions with images from scanning electron microscopy (SEM). We take SEM images of inclusions that contain size, shape and chemical composition information. Holm’s group uses convolutional neural networks to extract features from images (basically turning an image into a long list of numbers) and machine-learning classification methods. Thus far, we have been able to distinguish inclusions from things like pores and surface contamination, and we are working on grouping inclusions by chemical composition.
As this is my final Academic Pulse column, I wanted to say thank you to Industrial Heating for the opportunity and to everyone who has read this column. I’ve enjoyed writing about my group’s work, and I hope you’ve found it interesting.