2024 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2024): Special Session: Big Data Anayltics I
Program Organizers: Joseph Beaman, University of Texas at Austin
Monday 1:30 PM
August 12, 2024
Room: 616 AB
Location: Hilton Austin
Session Chair: Jia Liu, Auburn University
1:30 PM
Unsupervised Multi-Modal Machine Learning for In-Process Monitoring of Additive Manufacturing: Anthony Garland1; Matthew McKinney1; Benjamin White1; Brad Boyce1; Michael Heiden1; Jesse Adamczyk1; Dan Bolintineanu1; 1Sandia National Laboratories
This work introduces a machine learning framework for analyzing multi-modal data in additive manufacturing, specifically laser powder bed fusion (LPBF). It uses the CLIP model to learn joint embeddings from in-process monitoring data (images and audio) and post-build data (optical images, height maps, force-displacement curves) without labeled data. The framework effectively predicts changes in LPBF process parameters. The embeddings are physically meaningful, as verified through clustering and 2D projections. This approach has potential for process monitoring and quality assessment in additive manufacturing and shows an effective approach for using large amounts of unlabeled data.
1:50 PM
Accurate In-situ Detection of Localized Porosity in LPBF Using Infrared Camera: Berkay Bostan1; Shawn Hinnebusch1; David Anderson1; Albert To1; 1University of Pittsburgh
Porosity in Laser Powder Bed Fusion (LPBF) parts significantly impacts durability and strength that are crucial in safety-critical applications. Addressing this challenge, precise porosity control is essential for the reliability of LPBF-manufactured components. Previous studies on LPBF porosity prediction often used inappropriate process parameters to simplify detection, typically focusing on larger, more noticeable pores. This study diverges from such methods by developing a machine learning framework that precisely predicts localized porosity. Utilizing various features from in-situ infrared camera monitoring, the framework was tested on entirely unseen parts and achieved over 90% accuracy while maintaining a false positive ratio of less than 3%, even for pores ten times smaller than the sensor resolution. Furthermore, SHAP (SHapley Additive exPlanations) analysis was employed to investigate pore formation mechanisms, revealing complex interplays in different regimes. This research not only advances in-situ porosity detection in LPBF but also deepens the understanding of pore formation mechanisms.
2:10 PM
On the use of Event-based Change Detection Neuromorphic Imagers for In-Process Monitoring of Additive Manufacturing: David Mascarenas1; Andre Green1; Allison Davis1; John Bernardin1; 1Los Alamos National Laboratory
Imagers are attractive sensors for in-process monitoring of additive manufacturing (AM) processes on account of their ability to capture high spatial-resolution data. However, this advantage comes at the price of large memory generation. Additionally, some metallic AM processes have high light intensity and high speed dynamics which further complicates data collection. Emerging change detection, event-driven, neuromorphic imagers represent an alternate imaging modality. Change detection event imagers only report data for pixels whose log intensity value has changed beyond a specified threshold. These imagers tend to generate less data than conventional imagers, respond to sub-millisecond phenomena in the environment, and have a high dynamic range. These properties make these imagers attractive for a variety of AM processes. We demonstrate the performance of change detection event-based neuromorphic imagers in the context of AM/welding processes such as glass laser AM and arc/laser welding.
2:30 PM
Nondestructive Fatigue Life Prediction for Additively Manufactured Parts through a Multimodal Transfer Learning Framework: Anyi Li1; Arun Poudel1; Shuai Shao1; Nima Shamsaei1; Jia Liu1; 1Auburn University
Understanding the fatigue behavior and accurately predicting the fatigue life of laser powder bed fusion (L-PBF) parts remain a pressing challenge due to complex failure mechanisms, time-consuming tests, and limited fatigue data. This study proposes a physics-informed data-driven framework, namely, a multimodal transfer learning (MMTL) framework, to understand process-defect-fatigue relationships in L-PBF by integrating various modalities of fatigue performance, including process parameters, XCT-inspected defects, and fatigue test conditions. It aims to leverage a pre-trained model with abundant process and defect data in the source task to predict fatigue life nondestructively with limited fatigue test data in the target task. MMTL employs a hierarchical graph convolutional network (HGCN) to classify defects in the source task. The synergies learned from HGCN are then transferred to fatigue life modeling in neural network layers. MMTL validation through numerical simulations and real-case studies demonstrates its effectiveness in fatigue life prediction of L-PBF parts.
2:50 PM
Multi-Scale Defect Monitoring in Laser Powder Bed Fusion Using Acoustic Emission Sensing: Prahalada Rao1; Benjamin Bevans1; Alex Riensche1; Antonio Carrington1; Mihir Darji1; Yuri Plotnikov2; John Sions2; Kyle Snyder2; Derek Hass2; 1Virginia Tech; 2Commonwealth Center for Advanced Manufacturing
In this work, we used in-situ acoustic emission sensors for online monitoring of part quality in laser powder bed fusion (LPBF) additive manufacturing process. Currently, sensors such as thermo-optical imaging cameras and photodiodes are used to observe the laser-material interactions on the top surface of the powder bed. However, these existing sensing modalities are unable to penetrate beyond the top surface of the powder bed. Herein, four passive acoustic emission sensors were installed in the build plate of an EOS M290 LPBF system. Acoustic emission data was acquired during processing of stainless steel 316L samples. The spatially localized acoustic emission signatures were statistically correlated (R2 > 85%) to multi-scale aspects of part quality, such as thermal-induced part failures, surface roughness, and solidified microstructure (primary dendritic arm spacing).
3:10 PM Break
3:40 PM
Fully Registered Overhang X4 Data from Additive Manufacturing Metrology Testbed (AMMT): Multi-Sensor Datasets Integration: Zhuo Yang1; Yan Lu2; Ho Yeung2; Brandon Lane2; 1Georgetown University; 2NIST
This presentation details a comprehensive dataset integrating multiple raw datasets from the National Institute of Standards and Technology (NIST) Additive Manufacturing Metrology Testbed. The data originates from an experiment involving four identical overhang parts. It encompasses five raw data types: digital commands, real position and laser power, melt pool monitoring (MPM) images, layerwise images, and X-ray computed tomography. The raw data underwent processing to remove noise and extract features. For instance, MPM images were used to extract melt pool geometric features using various methods. Subsequently, extracted features from each dataset were spatially and temporally registered to the machine's coordinate system. Fully registered data comprises millions of data points with dozens features. The document describes the data processing, feature extraction, and data registration techniques employed. It details existing uncertainties encountered during the integration process. The data is presently undergoing NIST's data publishing process. Sample data will be released during the conference.
4:00 PM
Detection of Humping and Porosity Flaws in Wire Arc Directed Energy Deposition Using In-situ Meltpool Imaging: Prahalada Rao1; Andre Ramalho2; Anis Asad3; Benjamin Bevans1; Joao Oliveira2; 1Virginia Tech; 2NOVA University, Lisbon; 3Universidade Federal do Paraná
In this work we detected humping and porosity flaws in wire arc directed energy deposition additive manufacturing (WA-DED) processes using data acquired from an in-situ meltpool imaging sensor. As a first-step toward closed-loop process control, there is a burgeoning need to monitor and detect incipient process drifts. We instrumented a WA-DED system with a high-speed meltpool imaging camera to detect two common anomalies, namely, humping and porosity. Physically intuitive signatures encompassing meltpool morphology and intensity features were extracted from the acquired images. These physically intuitive features were subsequently used as inputs to a hierarchical machine learning classification model. Through experiments conducted over multi-layer parts, we show that the model achieves an overall classification fidelity of ~90% (statistical F1-score). The approach was further benchmarked against black-box deep learning models, trained directly with meltpool images resulting in F1-score of 85%.
4:20 PM
Multimodal Federated Product-of-Experts for Collaborative Fatigue Life Prediction in Additive Manufacturing: Anyi Li1; Jia Liu1; 1Auburn University
Accurately predicting the fatigue life of additive manufactured (AM) parts from inspection and testing data is crucial for their adoption in critical applications like aerospace. However, not every manufacturer possesses nondestructive inspection or fatigue testing capability to enable their efforts in fatigue life prediction. Also, inspection and testing specimens are too limited to build an accurate predictive model for individual manufacturers. To tackle these challenges, we propose a multimodal federated product-of-experts (MMFedPoE) framework, which can aggregate information from the manufacturers for collaborative fatigue life prediction without sharing their proprietary data. Manufacturers can train fatigue life prediction models based on their fabrication conditions and inspection or testing capabilities (e.g., X-ray CT scans, fatigue testing conditions). Then MMFedPoE trains joint distributions of different inspection and testing data by employing a product-of-experts approach and shares their distributions among the manufacturers. MMFedPoE can benefit manufacturers with different capabilities in accurate fatigue life prediction.
4:40 PM
Correlating Layer-wise Laser Powder Bed Fusion Process Data: Srikar Rairao1; Caleb Campbell1; Brett Brady1; Kevin Shay2; Kevin Lamb2; Bradley Jared1; 1University of Tennessee Knoxville; 2Y-12 National Security Complex
Optimizing the process of layer-wise laser powder bed fusion (L-PBF) is crucial for enhancing part quality. Our research focuses on the integration of image and computed tomography (CT) data to develop efficient datasets as inputs for models that can predict and analyze defects across various layers of fabricated parts. Utilizing a Farsoon FS271M L-PBF machine with low-cost cameras, our team seeks to correlate data streams to improve real-time monitoring and post-process validation of stainless steel. This project aims to refine defect detection through advanced image analysis techniques and also focuses on using neural networks like Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to ensure part reliability. The goal is to create a predictive model that enhances the detection of defects, thereby reducing manufacturing time and costs while maintaining high-quality standards in additive manufacturing processes.
5:00 PM
Melt Pool Depth Contour Prediction from Surface Thermal Images with Transformer Models: Odinakachukwu Ogoke1; Peter Pak1; Alexander Myers1; Guadalupe Quirarte1; Jack Beuth1; Jonathan Malen1; Amir Barati Farimani1; 1Carnegie Mellon University
The use of Laser Powder Bed Fusion in certain applications is limited due to defect-induced variation in mechanical and fatigue performance. However, most methods for defect characterization take place post-build, requiring significant material and time expenses to fully characterize changes in build conditions. The three-dimensional geometry of the melt pool can serve as a real-time indicator of defect formation, as non-overlapping melt pools lead to lack-of-fusion porosity, and deep, narrow melt pools indicate keyhole porosity formation. However, the sub-surface appearance of the melt pool is not visible through in-situ monitoring methods feasible to implement during the build process. Therefore, we create a deep learning framework for predicting the melt pool sub-surface morphology from in-situ high-speed surface thermal images. We evaluate model performance by comparing the geometric properties of the predicted melt pool boundary with optical micrographs of the corresponding physical melt pool cross-section in both single-track and multi-track melting scenarios.