2023 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2023): Data Analytics: Laser Powder Bed Fusion
Program Organizers: Joseph Beaman, University of Texas at Austin
Wednesday 8:00 AM
August 16, 2023
Room: 412
Location: Hilton Austin
Session Chair: Paul Hooper, Imperial College London
8:00 AM
Self-repair of Defects: The Achilles Heel for In-process Detection of Small Pores?: Richard Williams1; Sebastian Larsen1; Harry de Winton1; Paul Hooper1; 1Imperial College London
Pores created in laser powder bed fusion are frequently repaired as a natural part of the manufacturing process. This self-repair occurs because the melt pool spans several layers in depth, remelting material that contains pores created in previous layers. This self-repair presents a problem for in-process defect detection. Some pores that are detected will not be present in the final part, leading to high false positive rates even for a perfect system. In this work we conduct layer-wise ex-situ micro-CT scanning, where the build sample is removed from the machine after each layer, to observe pore locations and self-repair over a number of layers. In-process monitoring using off-axis illuminated imaging and co-axial melt-pool imaging is performed during the processing of each layer to look for signatures of pore formation and healing. Results showing links between the previously melted layer, the monitoring signals and the subsequent layer will be presented.
8:20 AM
Unveiling Melt Pool Defect Signatures with Interpretable Machine Learning: Sebastian Larsen1; Paul Hooper1; 1Imperial College London
As machine learning (ML) becomes an essential part of in-situ monitoring, model interpretation is needed to ensure robustness and alignment. In this study, we use interpretable ML methods to better understand melt pool defect predictors captured from co-axial high-speed imaging of the laser powder bed fusion process. The methods explored include a variety of explainable AI tools, such as saliency maps and feature importance metrics, to visualise and quantify melt pool features. Results from three datasets are presented: single tracks, laser defocus, and localised defects. The analysis enables a comparison between high-speed imaging and an equivalent photodiode, with each source of improvement measured. Through this analysis, key descriptors can be better understood, providing a deeper understanding of the melt pool behaviour.
8:40 AM
Real-time Melt Pool Characterization in Laser Powder Bed Fusion Using Acoustic and Photodiode Monitoring: Haolin Liu1; Christian Gobert1; Brandon Abranovic1; Hongrui Chen1; Kevin Ferguson1; Jack Beuth1; Anthony Rollett1; Levent Burak Kara1; 1Carnegie Mellon University
As the demand for high-quality fabrication grows, the need for real-time monitoring of the laser powder bed fusion (LPBF) process has also grown, leading to the incorporation of a range of online sensing methods such as acoustic sensing and high-speed imaging. In practice, real-time high-resolution melt pool image capture remains computationally demanding. In this work, we propose a cost-effective monitoring approach that can replace the need for direct visual sensing in LPBF. In particular, we utilize acoustics and photodiode-based data to predict highly time-resolved visual melt pool characteristics in a nearly real-time fashion. Our approach enables a data-driven mapping within a time window of 2.0 ms. Our work demonstrates the feasibility of using a cost-effective method to achieve online visual melt pool characterization and contributes to the advances in quality control for LPBF.
9:00 AM
Process Parameter Optimization Using Photodiode Intensity in Laser Powder Bed Fusion: Amit Verma1; Gabe Guss1; Saad Khairallah1; Ava Ashby1; CE Kim1; Ibo Matthews1; 1LLNL
One limiting factor towards the wider adoption of laser powder bed fusion (LPBF) process, in regard to complex geometries, is the time required to optimize the machine parameters. Currently, most commercial LPBF printers include an in-situ photodiode sensor to monitor the melt-pool, which can be leveraged for this optimization in a timely manner. In this talk, we will present a 1D-CNN framework which leverages photodiode data to optimize process parameters. Since this framework leverages experimental data, we will comment on training prints & their limitations, experimental validation of the model on new prints, feature engineering to capture 2d & 3D geometric constraints in a 1D model, and how it relates to improving dimensional accuracy for features such as overhangs, thin walls, down-skins, etc. Lawrence Livermore National Laboratory is operated by Lawrence Livermore National Security, LLC, for the U.S. Department of Energy, National Nuclear Security Administration under Contract DE-AC52-07NA27344.
9:20 AM
Monitoring of Single-track Quality in Laser Powder Bed Fusion using In-situ Thermionic Sensing: Benjamin Bevans1; Philip DePond2; Aiden Martin2; Nick Calta2; J-B Forien2; Gabe Guss2; Brian Giera2; Prahalada Rao1; 1Virginia Tech; 2Lawrence Livermore National Laboratory
This work concerns the laser powder bed fusion (LPBF) additive manufacturing process. In this work track quality was monitored in-situ using a novel thermionic sensing approach. It is important to monitor the quality of the track in LPBF as it is the basic building block of the part. In this work, track quality is defined as track width and the percent continuity of the track. The objective of this work is to detect the onset of track deviations using signatures extracted from a novel thermionic sensor. This thermionic sensor is attached to the substrate and measures the voltage response of the electrons released when the laser interacts with the build plate. The signals from the thermionic sensor are analyzed using empirical mode decomposition, and the derived signatures are used subsequently within elementary machine learning models to predict the quality of track.
9:40 AM Break
10:10 AM
Mining Complex Feedstock-Geometry-Process-Quality Relationships for Powder Bed Fusion Thin Features using Graphical Model: Naresh Koju1; Xiaoyu Chen1; Li Yang1; 1University of Louisville
Thin features (thin walls and thin struts) in the strut-based and planar lattice structure show a powder feedstock-geometry-process-quality (PGPQ) characteristic. However, the relationship between the input variables (powder feedstock, laser Power, scan speed, feature type, and feature dimension) and their properties are complex to understand. Besides input variables, the intermediate variables such as dimension (dimensional variability within a sample), porosity and pore size distribution, and even the grain sizes are expected to correlate to the quality/mechanical properties of these thin features. Therefore, the main objective of this study is to explore the complex graphical relationship between the input variables, intermediate variables, and the final flexural properties of thin features by utilizing a graphical model-based machine learning (ML) model. The ML model depicts the influence of some of the intermediate variables (e.g. porosity) on the flexural properties of the thin features, which helps revealing the complex PGPQ characteristics.
10:30 AM
Distilling Thermal Signatures from Reduced Order Physics Models for Electron Beam Powder Bed Fusion Processing: Patxi Fernandez-Zelai1; Sebastien Dryepondt1; Amir Ziabari1; Michael Kirka1; 1Oak Ridge National Laboratory
Extracting meaningful process-structure trends from spatiotemporal thermal data simulation can be extremely challenging. Model parameter uncertainty, simplified physics, and idealized conditions, introduced model bias which further makes difficult the quantification of salient thermal features. In this work we propose an approach to encode the spatiotemporal outputs from a reduced order thermal model into a compact latent space representation. This is achieved using a self-supervised video-transformer machine learning framework. The identified latent representation summarizes the relevant physics and, given sufficient data, allows for calculation of similarity measures. As an example the approach is used to establish a data-driven process-structure model for an additively manufactured Ni-based superalloy. This methodology is well suited to be used towards in-situ process monitoring, scan pattern design, and component qualification.
10:50 AM
Detecting Failures in Laser Powder Bed Fusion Additive Manufacturing of Complex Lattice Structures using Multi-sensor Data and Machine Learning: Anis Asad1; Benjamin Bevans1; J-B Forien2; Aiden Martin2; Nick Calta2; Philip DePond2; Gabe Guss2; Brian Giera2; Prahalada Rao1; 1Virginia Tech; 2Lawrence Livermore National Laboratory
The goal of this work is to detect the probability of strut failures in complex lattice structures built using laser powder bed fusion (LPBF). In pursuit of this goal, the objective is to predict the probability of strut failure as a function of heterogeneous sensor data from a pyrometer and a photodiode placed coaxially (in-line) via supervised machine learning models. The result is a 3D digital twin of the lattice created to demarcate areas of failure. This model was trained on a single lattice structure with artificially generated broken struts and tested on an additional lattice with smaller broken struts than the training lattice. In this work we show that the developed approach is capable of accurately detecting broken lattice struts with a statistical fidelity exceeding 80% (F-score) even when transferred to a different lattice geometry with finer resolution of breakage.
11:10 AM
Data-driven Local Porosity Prediction in Laser Powder Bed Fusion via In-situ Monitoring: Berkay Bostan1; Shawn Hinnebusch1; David Anderson1; Albert To1; 1University of Pittsburgh
In this study, the geometry-agnostic deep learning scheme has been developed for defect detection during the laser powder bed fusion (LPBF) process. DNNs model has been trained that gives +90% accuracy with a relatively smaller dataset. Inputs to DNNs include various thermal signatures (interpass temperatures, heat intensities, and cooling rates) and spatter locations. At the same time, when making predictions, the DNNs architecture considers the features of not only the relevant pixel, but also neighboring pixels in all directions (desired order of neighbors in the current, upper, and lower layers). The potential outcomes of this study are simultaneous defect prediction during manufacturing and repairing the defects by rescanning the concerned region. Furthermore, defect formation mechanisms have been investigated by SHAP (SHapley Additive exPlanations) feature importance analysis method, and it is observed that spattering is the most dominant factor for defect formation until the melt pool reaches a certain size.
11:30 AM
Investigating Correlation Between Melt Pool and Overhang Surface: Zhuo Yang1; Jaehyuk Kim2; Yan Lu2; Brandon Lane2; Yande Ndiaye2; 1Georgetown University; 2NIST
Real-time control through the use of sensors and controllers is being applied in powder bed fusion AM systems to manage in-situ process features like melt pool size and temperature. Despite this technology, geometry deviation and surface roughness defects still persist due to difficulty of obtaining accurate geometric information during the build. Although in-situ monitoring captures several in-process data, precise geometric information is challenging to acquire. To address this problem, a statistical approach is proposed to predict material formation in 3D, particularly the geometry of overhang surface using process parameters and coaxial melt pool images. The preliminary results from an experiment involving four overhang parts, with 100,000 melt pool images generated, indicate that the proposed method provides a solution for real-time control of part geometry and surface roughness. This presentation would present the findings of overhang dross depth, related melt pool features, and their statistical correlations.