2024 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2024): Special Session: Feedforward and Feedback Control in AM I
Program Organizers: Joseph Beaman, University of Texas at Austin
Tuesday 1:30 PM
August 13, 2024
Room: 400/402
Location: Hilton Austin
Session Chair: Chinedum Okwudire, University of Michigan
1:30 PM
Temperature Control of Digital Glass Forming Processes: Balark Tiwari1; Yujun Zhou1; Nishan Khadka1; Andre Bos2; Douglas Meredith2; John Bernardin2; Edward Kinzel1; Xiangliang Zhang1; Robert Landers1; 1University of Notre Dame; 2Los Alamos National Laboratory
Digital Glass Forming is a manufacturing process that uses a laser to heat glass, creating a workable zone that can be shaped with tools or by feeding glass filament or fiber into the zone to add material. This method allows for precise forming of glass into desired geometries. However, variations in scan speed, filament feed rate, part geometry, etc. affect the thermal process, complicating the selection of the laser power necessary to achieve uniform temperature distribution and consistent morphology. This paper introduces a feedback temperature controller that adjusts the laser power based on feedback from a thermal camera to maintain a constant maximum temperature. The feedback controller is then augmented with a data-driven machine learning model that is used to predict melt pool temperature, enabling feedforward control capabilities. Experimental studies are conducted to evaluate the effectiveness of these control methodologies when depositing straight tracks, square structures, and multi-layered chimney structures.
1:50 PM
Defect-free Ceramic Hybrid-AM using Intelligent Layer Reworking: Louis Masters1; Dan Davie1; Matthew Shuttleworth1; Tyler Green1; Mehmet Dogar1; Robert Kay1; 1University of Leeds
Hybrid additive manufacturing of advanced ceramics facilitates the production of highly dense and precise parts by combining additive and subtractive processes. However, extrusion-based processes are susceptible to stochastic defects such as voids, which can degrade material properties, leading to premature failure and lower yield. This research demonstrates deep learning informed selective layer reworking for a ceramic hybrid additive manufacturing platform. Each layer was evaluated in-situ using a vision-based monitoring system, consisting of a camera and laser profilometer. Through closed-loop control, a decision was made autonomously to remove defective layers via subtractive operations, prior to reprinting. The deep learning model detected voids with a precision of 90%, and parts have been demonstrated to be free of extrusion-related voids after corrective action. This unlocks new opportunities for regulated industries hoping to exploit quality-assured ceramic components that benefit from freeform fabrication.
2:10 PM
Toward Rapid Process Qualification of Laser Powder Bed Fusion Additive Manufacturing Using Physics-Based Model Predictive Control: Prahalada Rao1; Alex Riensche1; Benjamin Bevans1; Antonio Carrington1; Kaustubh Deshmukh1; Kamden Shephard1; John Sions2; Kyle Snyder2; Yuri Plotnikov2; Kevin Cole3; 1Virginia Tech; 2Commonwealth Center for Advanced Manufacturing; 3University of Nebraska-Lincoln
We developed and applied a physics-guided model predictive control approach to autonomously optimize the processing parameters for an LPBF part before it is printed. The control approach creates a customized layer-by-layer processing plan for each LPBF part shape, such that the model-predicted thermal history of a part matches a predetermined ideal or target thermal history. Currently, LPBF processing parameters are optimized through empirical studies based on simple test coupons. However, processing parameters optimized for test coupons seldom transfer to practical parts necessitating further testing. We demonstrate, with relatively complex stainless steel 316L parts processed on a commercial EOS M290 LPBF machine, the following advantageous outcomes from using the physics-based model predictive control approach: (i) elimination of anchoring supports in parts with prominent overhang features; (ii) improvement in geometric accuracy and surface integrity of hard to access internal features; and (iii) reduction in microstructure heterogeneity resulting in consistent part properties.
2:30 PM
Laser Powder Bed Fusion Process Feedback Control based on In-Situ Powder Layer Thickness: Jorge Neira1; Ho Yeung1; 1National Institute of Standards and Technology
Metal laser powder bed fusion (LPBF) is a promising additive manufacturing technique for producing complex metal parts. However, the quality of the final product can be affected by various factors such as laser power, scanning speed, powder spreading quality, and layer thickness. In this study, we propose a real-time feedback control method based on the powder spreading quality and layer thickness. We employ a laser 3D scanner to measure the surface profile of each layer before and after the powder has been spread. A powder re-spreading will be triggered if an abnormality is detected on the spread powder surface. The difference between the profile is taken as the actual powder layer thickness and used to adjust the scan strategy for the next layer. The feedback control system is integrated into the customized-built LPBF testbed, and the effectiveness of the proposed feedback control system is demonstrated through a series of experiments.
2:50 PM
Controlling Multi-track Melt Pool Size Variation Based on Local Thermal Environment for Inconel 718 Laser Powder Bed Fusion via Data-driven Approaches: Seth Strayer1; William Frieden Templeton2; Praveen Vulimiri1; Shawn Hinnebusch1; Sneha Narra2; Albert To1; 1University of Pittsburgh; 2Carnegie Mellon University
The melt pool size for Inconel 718 laser powder bed fusion (L-PBF) varies substantially throughout a part, causing uncertain changes in microstructure and properties that prevent the process’ broader adoption by the industry. This work presents an iterative, data-driven, feedback-feedforward control model for maintaining a consistent melt pool size based on the local thermal environment (LTE) to help address this issue. First, a finite element (FE) machine-learned heat source model is trained to predict the multi-track thermal fields from high-fidelity computational fluid dynamics (CFD). Different process parameters are iteratively tested using the trained model and provided to a data-driven algorithm to predict the appropriate parameters that maintain a consistent melt pool size for various LTE. Second, the same approach is implemented using a functional heat source model. The benefits and disadvantages of the two methods are presented, and their efficacy in controlling the multi-track melt pool size is compared.
3:10 PM
In-Situ Layer Temperature Monitoring Enabled by Machine Vision: Chris O'Brien1; Chad Duty1; Kris Villez2; 1University of Tennessee - Knoxville; 2Oak Ridge National Lab
Layer deposition time is crucial for the success of the additively manufactured (AM) parts as it directly affects inter-layer adhesion, structural integrity, and subsequent production efficiency. For large-format AM (LFAM), its optimization is essential due to print size and increased thermal capacity from high volumetric deposition rates. Setting the layer deposition time too short results in material collapse. However, if layer deposition time is too long it leads to debonding. Our goal is to develop a control method that optimizes the layer deposition time in LFAM systems using infrared cameras for real-time temperature measurement. In this study, we prototype our method with a small-scale 3D printer to generate data for a deep learning model that enables tracking of the printer head and estimation of the temperatures at the time of deposition. The work is the first step towards an image-based optimization loop to dynamically control layer time in LFAM processes.
3:30 PM
Pre-training Vision Encoders with Thermal Images for In-situ Process Monitoring in Laser Powder Bed Fusion: Peter Pak1; Francis Ogoke1; Andrew Polonsky2; Anthony Garland2; Dan Bolintineanu2; Dan Moser2; Brad Salzbrenner2; Mary Arnhart2; Jonathan Madison2; Thomas Ivanoff2; John Mitchell2; Bradley Jared3; Michael Heiden2; Amir Barati Farimani1; 1Carnegie Mellon University; 2Sandia National Laboratories; 3The University of Tennessee Knoxville
We investigate the use of melt pool thermal image datasets to pre-train vision encoders for real-time process monitoring in additive manufacturing. Feature learning is achieved through performing tasks such as inpainting temperature fields within masked melt pool regions. Different encoder mechanisms such as Vision Transformers or Convolutional Neural Networks are evaluated and compared against one another. By training our models on publicly available data from NIST and validating the models with experimental data, we enhance its prediction accuracy and enable its use in downstream tasks within in-situ monitoring systems. The features learned from this data can be correlated to those from optical imaging, acoustic signaling, or computed tomography. With this work we aim to improve quality control and process optimization in additive manufacturing environments.