Machine Learning to Optimize Additive Manufacturing Parameters

PROJECT NUMBER: 42

AUTHOR: Damien Churchwell, Computer Science | MENTOR: Hua Wang, Computer Science

ABSTRACT

—

One fundamental theory underpins materials science: process dictates structure and structure dictates properties. However, predicting this process—structure—process or PSP relationship is elusive, because it requires knowledge of the orientation of every grain, the location and distribution of defects, and every processing condition that lead to this structure, down to the smallest fluctuations in temperature, pressure and composition. From an informatics perspective, this creates an extremely large input space. Weak and redundant variables and variables that cannot be measured convolute important, measurable values. However, machine learning provides a number of robust techniques to extract PSP relationships from these convoluted data streams. Approximately 6000 samples have been printed to characterize the build parameters for 3D printed metal samples. The tested samples connect the microstructure and mechanical properties to laser power, speed, spot size, powder size, shape and part orientation. These data serve as the basis for development of machine learning (ML) algorithms — including decision trees, scalable vector regression, and random forest networks — that focus on two-way modeling of process-property and process-structure relationships. Our results show how these parameters effect mechanical performance through microstructure, particularly keyhole and lack-of-fusion porosity defects. This project will focus on the development of a data collection, processing, validation and distribution framework; on ML performance, accuracy and validation procedures.

VISUAL PRESENTATION

—

AUTHOR BIOGRAPHY

—

Damien Churchwell is a Computer Science – Data Science student in his junior year at Mines. He performs a variety of research with the MInDS@Mines Machine Learning research team. His current focus lies in image segmentation methods for additive manufacturing applications and natural language processing for gene-disease relationship predictions. He intends to continue his degree with a Statistics minor, and continue research into ML and neural network architectures.

Sponsored by Phillips 66