Advancement of 3D Printed Aluminum Alloys for the Aerospace Industry Using Machine Learning

Bowles, Patrick; Desmonts, Braeden; Carlo, Charles; Cervone, Gabriella

Student Work

Advancement of 3D Printed Aluminum Alloys for the Aerospace Industry Using Machine Learning

公开

The usage, research, and volume of additive manufacturing (AM) have increased rapidly over the past decade. This is a highly attractive technology for several applications in the aerospace and medical industries. There are many advantages from both engineering and financial perspectives due to the multitude of materials and complex geometries that can be produced, all while minimizing fixed costs if production lots are low. Additive manufacturing, in comparison to conventional manufacturing methods, deposits the material layer-by-layer to create the desired part instead of removing material by subsequent machining. Focusing on metal AM, specifically Laser Power Bed Fusion (LPBF), the manufacturing process is controlled by process parameters like scan speed, laser power, hatch spacing, build environment, and many other factors. These parameters need to be optimized to produce AM parts with a low percentage of defects like porosity. Without understanding the way these parameters affect the as-built part quality, material properties that are less attractive than what can be achieved with conventionally manufactured parts can result. This usually leads to the need for significant post-processing, which increases manufacturing cost and time. Currently, the validation and qualification for the as-built AM parts are very time-intensive and expensive, and usually require manual inspection from trained professionals. Increasing the quality of these parts, while minimizing the time needed for validation inspections will decrease the time and cost to fabricate complex AM components. This project aims to develop a Machine Learning algorithm, via Deep Learning Convolutional Neural Network (CNN) to identify and classify several types of porosity defects in AM-LPBF Al-10Si-Mg samples. The CNN will learn from labeled micrograph images taken from samples that were built with varying laser power, scan speed, and hatch spacing. Matterport Inc.’s implementation of Mask R-CNN was used. The data were annotated by generating pixel-perfect masks corresponding to each defect, thresholding any defects considered too small to classify, separating classes into sub-classes to ensure a balance of data, and in some cases stitching entire samples and splitting along custom boundaries to ensure the geometries of extreme defects were preserved. After several iterations the model achieved an overall classification accuracy of 69.23%.

This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.

Creator

Subject

Publisher