2023 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2023): Data Analytics: Large Scale AM, Deposition Methods, Metals
Program Organizers: Joseph Beaman, University of Texas at Austin

Tuesday 1:40 PM
August 15, 2023
Room: 404
Location: Hilton Austin

Session Chair: Eric MacDonald, University of Texas at El Paso

1:40 PM
Automated Layer Identification in Large Area Additive Manufacturing (LAAM): A Comparison of Image Thresholding and Edge Detection Techniques: Aissata Wadidie¹; Gregory Studer¹; Kris Villez²; ¹Advanced Stuctures and Composites Center; ²Oak Ridge National Laboratory
    Our study aims to develop an automated method for identifying layers on images of 3D-printed walls from a LAAM printer, as manual identification can be time-consuming. We applied three different image processing methods to identify edges between layers. Otsu thresholding was found to be the most accurate and required minimal manual intervention. From our study, we propose a new composite algorithm combining multiple methods for even greater accuracy. This research demonstrates the feasibility of using computer-based methods to automatically identify layers in 3D printing, reducing manual time and effort and improving the strength and quality of 3D-printed parts.

2:00 PM
Coaxial Color Channel Focus Evaluation to Estimate Standoff Height in Directed Energy Deposition: Callan Herberger¹; Lauren Heinrich²; Erik LaNeave¹; Brian Post²; Blane Fillingim²; Eric MacDonald¹; Thomas Feldhausen²; James Haley²; ¹The University of Texas at El Paso; ²Oak Ridge National Laboratory
    Directed Energy Deposition is an additive manufacturing process that is being rapidly adopted by industry and is well suited for the fabrication of complex components in a variety of metal alloys. Inexpensive and minimally intrusive methods to find the best standoff are needed to supply real time control to maintain optimal standoff distance. The present work explores the quantification of the focus of the three-color channels of a coaxial camera to determine the standoff height. An experiment was performed in which a 254 mm wall is built and the standoff height, initially 5.0 mm below the optimal position, was then intentionally increased by 1.0 mm to a final position 7.0 mm above optimal. Computer vision is shown to monitor the focus in each color band and estimate standoff distance. A response can be calculated in under 40 ms using simple hardware and can work in most laser-based DED systems.

2:20 PM
Automated Fiber Length Measurement for 3D Printed Polymer Composites, Including Identification and Measurement of Non-trivially Placed Fibers: Chris O'Brien¹; Chad Duty¹; ¹University of Tennessee - Knoxville
    The mechanical stiffness and strength of a composite material is significantly influenced by the fiber length. Processing parameters for large-scale printing systems (e.g., extrusion screw speed) can directly impact the resulting fiber length distribution. There are limited standardized methods for reproducible and generalizable quantification of fiber length. Current measurement techniques are often expensive, laborious, and error prone due to dependence upon human involvement. Deep learning offers the potential to automate the laborious tasks such as segmentation, identification, and measurement of imaged fibers. The aim of the presented work is a first step in establishing a deep learning-based fiber measurement pipeline that may be generalized across various fiber imaging set-ups. This presentation focuses on the description of the measurement of both stand-alone fibers as well fibers those that are imaged in non-trivial placements (e.g., overlapping). The end goal is an open-source method for reliable, automated quantification of residual fiber length.

2:40 PM
Variational Autoencoders for Comprehensive Feature Identification in Fatigue Analysis : William Frieden Templeton¹; Tharun Kondareddy¹; Justin Miner¹; Sneha Narra¹; ¹Carnegie Mellon University
    Fatigue life is a key performance metric for metal AM parts, making it an active research area generating vast datasets of micrographs from fatigue test specimens. Using these image-based datasets, this study pursues data analytics to find correlations between fatigue life and porosity defects. To do so, the information in the images must be quantified for statistical analysis, thereby introducing a source of information loss. This work aims to preserve information by applying variational autoencoders (VAEs) to the image dataset. By encoding the micrographs to a low-dimensional latent space, valuable information in the dataset is captured and similar features are clustered. Encodings that correlate with poor fatigue life are identified using random forest models, and examples of the features they link to are generated. The results will demonstrate the potential application of VAEs to quantify micrographs and discern useful correlations between defects and fatigue life.

3:00 PM
In-situ Monitoring of Wire Arc Additive Manufacturing for Machine Learning Based Prediction of Shape Irregularites and Mechanical Defects: Eduardo Miramontes¹; Joshua Penney¹; Bennett Fowler¹; Ethan Rummel¹; Sean Caufield¹; Anahita Khojandi¹; Bradley Jared¹; ¹University of Tennessee
    Wire Arc Additive Manufacturing (WAAM) has made great strides in recent years however, there remain numerous challenges still hindering adoption by industry. Defects in the parts degrade their mechanical performance. Inconsistency in the geometry of the weld beads or undesirable anomalies such as waviness, or humps can lead to loss of geometric accuracy and in extreme cases, they can propagate to subsequent layers, causing build failure. Developing a controls framework for defect mitigation requires a model that maps undesirable outcomes to information about the process obtained in real time. The development of a multi-sensor framework for real time data acquisition and several approaches for arriving at defect prediction model, employing well known machine learning methodologies including Random Forests, and Neural Networks are explored. The models are trained first on data obtained on a single build layer, and subsequently on a multi-layer wall. Their merits and drawbacks are discussed.

3:20 PM
End-to-end AI Models for Error Detection and Correction in Extrusion AM: Douglas Brion¹; Sebastian Pattinson²; ¹Matta; ²University of Cambridge
    Material extrusion is the most widely used additive manufacturing method, but its use in many applications is limited by its vulnerability to diverse errors. Expert operators can detect errors but cannot provide continuous monitoring or real-time correction. This has led to significant research into automated methods for error detection. However, current approaches can often only detect limited error modalities across a narrow range of parts and materials. Additionally, errors remain particularly challenging to correct, primarily requiring manual intervention. This talk will discuss the application of recent advances in large AI models and deep learning to tackle these problems. End-to-end models show great promise for detecting errors autonomously and powering feedback systems that can correct process parameters in real time, opening the door to improved part quality and uptake in end-use applications.

3:40 PM
Bayesian Data-augmentation of Thermal Models for Design of Nb-Ta-W Alloys: Brent Vela¹; Cafer Acemi¹; Peter Morcos¹; Alaa Elwany¹; Ibrahim Karaman¹; Raymundo Arróyave¹; ¹Texas A&M University
    Refractory alloys are difficult to process via conventional means due to their high strength and brittle nature. Additive manufacturing (AM) is an emerging solution which can bypass these processing difficulties, enabling the fabrication of refractory alloys into complex shapes with reduced material usage. Despite this, there is limited experimental data on the printability of refractory alloys. Both AM simulation and experimentation are prohibitively expensive and are thus not appropriate for high-throughput alloy design. In this work we propose printable alloys can be designed within the Nb-Ta-W chemistry-process space using a combination low-fidelity (Eagar-Tsai model), high-fidelity (Thermo-Calc Additive Manufacturing Module) thermal models, and limited single-track experiments. Specifically, using the Eagar-Tsai model as our prior belief of printability within this chemistry-process space, we leverage Bayesian-updating and networks of Gaussian Process Regressors to 1) fuse information from both thermal models and 2) to correct this fused-model with experimental data.