Session: 17-01-01: Research Posters

Paper Number: 150563

150563 - Machine Learning Approaches to Predict Surface Roughness in Inconel Samples Printed via Powder Bed Fusion Process 

Laser powder bed fusion (L-PBF) process is a metal Additive Manufacturing (AM) process that builds complex parts using a layer-by-layer method. During the printing process, parts are subjected to unique conditions which affect their surface roughness in the build direction (z-axis). Firstly, the material exhibits inherent waviness due to the layer-by-layer printing method. Secondly, surface roughness is significantly impacted by the sintering of satellite particles that partially melt into the surface as the printing surface solidifies.

Leveraging computational techniques and Machine learning (ML) for predicting surface properties and making design decisions for AM is advantageous, as it eliminates the need for numerous experiments. This is particularly beneficial given that setting up and conducting AM experiments is both time-consuming and costly for repetitive iterations. The AM design and print process can incorporate various Machine Learning (ML) techniques for functions such as design recommendations, topology and lattice optimization and enhancing the quality of the printed surface.

Controlling surface roughness is crucial when optimizing the powder bed fusion process for specific applications. Several factors and printing parameters contribute to the surface profiles, with laser power and scan speed being among the most influential. This study investigates the prediction capability of machine learning algorithms to estimate surface roughness profiles. An experiment is carried out using 15-53 microns Inconel powder particles. Cubic samples of 2 cm x 2 cm x 2 cm are printed by varying power and scan speed on three different levels using a design of experiments approach. The power levels are 100 W, 150 W, and 190 W, and the scan speeds are set to 400 mm/s, 600 mm/s, and 800 mm/s. Once the parts are printed, the samples are removed from the build plate and measured for their as-built surface roughness. Surface roughness data is acquired on the side surfaces of the samples using a laser confocal microscope.

The collected profile data is transformed using multiple feature extraction techniques and then utilized to train a Neural Network. This Neural Network classifies multiple line profiles based on their laser scanning speed and power, which generate labels for the input laser density. The trained Neural Network demonstrates more than 70 percent accuracy in classifying line profiles according to their associated laser parameters when tested with new data. The study aims to extend this research to encompass a broader range of process parameters in L-PBF, with the ultimate goal of training the Neural Network to predict surface roughness features based on a comprehensive set of input process parameters. In an ideal scenario, the ML algorithm will not only predict surface roughness but also provide insights into optimal parameter settings for achieving desired surface qualities.

Presenting Author: Santosh Rauniyar University of Houston

Presenting Author Biography: Dr. Rauniyar's current research focuses on the possibility of use of additive manufacturing techniques for the parts in concentrated solar power plants. His prior research was related to experiment and numerical studies of the transient nature in laser powder bed fusion process.

Authors:

Santosh Rauniyar University of Houston
Ben Xu University of Houston
Mathew Farias University of Houston