2023 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2023): Modeling: Machine Learning, Data Driven, Digital Twins
Program Organizers: Joseph Beaman, University of Texas at Austin
Monday 1:30 PM
August 14, 2023
Room: 415 AB
Location: Hilton Austin
Session Chair: David Rosen, Agency for Science, Technology and Research
1:30 PM Cancelled
Validation of Simulation Based Predictions of Recoater Interference in Laser Powder Bed Fusion: Michael Gouge1; Chao Li1; Jeff Irwin1; 1Autodesk Inc.
Recoater interference is one of the most common and most expensive laser powder bed fusion (LPBF) failure modes. It is critical to investigate the mechanisms and the cause of recoater interference. Furthermore, accurate simulation-based predictions of recoater interference prior to manufacturing could help improve the printability, part quality, and economical feasibility. In this work, twenty-two overhang structures with various overhang angles, lengths, and build orientations were built to trigger recoater impacts. In situ imaging was used to identify the abnormality of powder surface. Nine of the parts penetrated the powder evidenced by recoater scraping, while three of parts caused recoater jams. The results show that the overhang angle and length were the most important indicators for recoater interference. Netfabb Local Simulation was used to create a digital twin of the experiment and was shown to accurately predict when parts were likely to cause recoater scraping and jams.
1:50 PM
High Fidelity Model of Directed Energy Deposition: Laser-Powder-Melt Pool Interaction and Effect of Laser Beam Profile on Solidification Microstructure: Saad Khairallah1; Eric Chin1; Michael Juhasz1; Aurelien Perron1; Scott Mccall1; Joseph Mckeown1; 1Lawrence Livermore National Laboratory
A multiphysics model was developed to reproduce the laser directed energy deposition process to a high-fidelity. This includes resolving the laser-powder-melt pool interactions (powder impingement and incorporation into melt pool, hydrodynamics flow condition and laser absorption inefficiencies) as well as the resulting solidification microstructure. This micrometer scale digital twin captured the details of a grain refinement mechanism caused by high powder flow rate. Furthermore, it was used to explore how laser beam shaping could impact the microstructure. Using a ring laser beam profile instead of the standard Gaussian laser profile, the model predicted a decrease in the thermal gradient which in turn can increase propensity for more desirable equiaxed grains. This work was performed under the auspices of the U.S. Department of Energy (DOE), by Lawrence Livermore National Laboratory (LLNL) under Contract No. DE-AC52-07NA27344. Funding: Laboratory directed research and development project 22-SI-007. IM release LLNL-JRNL-844475.
2:10 PM
Thermal Monitoring and Cure Process Modelling of Dual-wavelength VPP Printing: Heyang Zhang1; Yue Zhang1; Xiayun Zhao1; 1University of Pittsburgh
Dual-wavelength vat photopolymerization (VPP) is an emerging multi-material additive manufacturing (AM) technology in which liquid photopolymers are selectively cured by two individually controlled light beams that have disparate wavelengths. In this work, ultraviolet and visible light beams are delivered through digital light processing (DLP) using digital micromirror devices. A model multi-material system that comprises epoxy and acrylate curing components is used. As the heat generated differs between epoxy and acrylic curing reactions, a thermal model with binary heat sources is created using in-situ thermal monitoring, ex-situ photo differential scanning calorimetry, and inverse heat conduction (IHCP) optimization methods to estimate the exothermic rate, time-resolved thermal profile, and degree of curing during the dual-wavelength VPP. This work will develop a framework of integrating in-situ thermal monitoring, physics simulation, and data-driven modeling to identify and quantify different reactions occurring during multi-wavelength multi-material VPP processes.
2:30 PM
Layer-wise Prediction of Microstructural Evolution in Laser Powder Bed Fusion Additive Manufacturing using Physics-based Machine Learning: Alexander Riensche1; Ajay Krishnan2; Benjamin Bevans1; Grant King3; Kevin Cole3; Prahalada Rao1; 1Virginia Tech; 2Edison Welding Institute; 3University of Nebraska-Lincoln
In this work we developed a framework to predict microstructure formation in the laser powder bed fusion (LPBF) of Inconel 718 parts. The microstructure is predicted as a function of sub-surface cooling rate estimated from a rapid part-level computational thermal model within elementary machine learning models. In this work, the microstructure evolved is quantified layer-by-layer in terms of three aspects: meltpool depth, grain size (primary dendritic arm spacing), and microhardness. The approach predicts the microstructure evolved with statistical fidelity exceeding 85% (R2). This is substantial improvement over existing microstructure prediction which are only able to predict the microstructure of a small region (~1 mm3) and not of the entire part.
2:50 PM
Toward Post-superficial Temperature Monitoring During Additive Manufacturing through Data-driven Inpainting: Jiangce Chen1; Mikhail Khrenov1; Jiayi Jin2; Sneha Narra1; Chris McComb1; 1Carnegie Mellon University; 2Tsinghua University
Understanding the temperature history of a built part during additive manufacturing (AM) is critical for studying the relationship between process parameters and product quality as temperature plays determinant role in melt pool dimensions, defect formation, and microstructure evolution. Unfortunately, the current thermal sensors used to monitor the AM process cannot provide a complete temperature distribution, which restricts the ability to study this relationship. In this paper, we propose a data-driven inpainting machine learning (ML) model that restores the temperature of the entire built part from incomplete temperature data captured by thermal sensors. We generate a dataset of temperature histories for parts with various geometries using a finite element model calibrated using experimental data. Our experiments demonstrate that the inpainting ML model accurately predicts both simulation and experimental data. This ML model has the potential to establish digital twins for AM-built parts, enabling efficient process optimization.
3:10 PM Break
3:40 PM
Modeling and Simulation of Vat Photopolymerization Additive Manufacturing: A Review
: Yousra Bensouda1; Heyang Zhang1; Yue Zhang1; Xiayun Zhao1; 1University of Pittsburgh
Vat photopolymerization (VPP) additive manufacturing (AM) has gained a lot of popularity across many industries, with various applications in biomedical engineering and electronics. This work aims to present a comprehensive review of existing work on model and simulation of conventional VPP processes and identify the critical gaps to improve the model prediction accuracy. Further, we propose a framework of developing comprehensive VPP processes via combining multiscale multi-physics modeling and simulation with machine learning to achieve surrogate modeling of VPP for compute efficiency and thus real-time process control. Moreover, we discuss the unique challenges in modeling and simulation of some emerging VPP processes such as two-wavelength VPP and possible physics-based data-driven methods by extending the methods that are aimed for improving traditional single-wavelength VPP.Keywords: Photopolymerization, additive manufacturing, modeling and simulation, multimaterial, multiphysics, machine learning
4:00 PM Cancelled
Multi-scale Modeling of Thermal/reisidual Stress in Additive Manufacturing Across Grain- , track- and Part-scales: Wentao Yan1; 1National University of Singapore
To obtain comprehensive understanding and accurate prediction of thermal/residual stress, we have developed multi-scale models. In the grain-scale model using the crystal plasticity finite element method, the grain structure evolutions are implemented from phase field simulation to resolve the interactions of different grains. With the temperature profiles from meso-scale thermal-fluid flow model, both the thermal deformation during heating and redistribution of the plastic deformation during cooling are simulated. In the track-scale model, besides the temperature profiles, the realistic geometry including rough surfaces and internal voids is implemented from the meso-scale thermal-fluid flow model, to reproduce the thermal stress concentrations and explain the cracking phenomenon. The part-scale model incorporates the track-scale thermal stress results to ensure acceptable computation burden and good accuracy. Moreover, to reduce the computational cost, we develop a physically-informed data-driven prognostic model of temperature, with a training database of only ~40 high-fidelity simulation cases under different manufacturing parameters.
4:20 PM
Powder Bed Fusion Surrogate Models via Convolutional Neural Networks: David Rosen1; John Ong2; U-Xuan Tan2; Qing Tan2; Umesh Kizhakkinan2; Huy Do2; Clive Ford1; 1Agency for Science, Technology and Research; 2Singapore University of Technology and Design
To support the design of structural metal parts for demanding applications, accurate predictions of part properties are needed, which high fidelity simulations can provide. However, metal powder bed fusion (PBF) simulations are far too computationally demanding, due to their very complex physical phenomena, for use in design optimization that can require dozens or hundreds of iterations. Rather, we are developing surrogate models of PBF process simulation results based on 3D convolutional neural network (CNN) technology. These CNN surrogate models compute part properties at high resolution in much less than one second. In this presentation, we summarize PBF process simulations and detail the CNN surrogate models developed for residual stress, deformation, and mechanical property distribution predictions of part designs. Examples of metal part fabrication results are compared with simulation and surrogate model predictions. Application of the surrogate models in part design optimization is illustrated.
4:40 PM
Predicting Temperature Field for Metal Additive Manufacturing using PINN: Bohan Peng1; Ajit Panesar1; 1Imperial College London
Performing thermomechanical simulation for selective laser melting is a non-trivial and critical task for printability simulation. In addition to the numerical methods, attempts of using a physics-informed neural network (PINN) have shown promise in predicting the temperature fields. In this work, a PINN is constructed with the physics of only homogeneous heat transfer but augmented with data points from a heterogeneous condition with phase change occurring (i.e. from metal power to solid metal). It demonstrates the capability of adopting a PINN (even based on a simple and imperfect physical model) to account for real and more complex phenomena, paving the way for more complex and faster printability simulation for SLM as supplemented by PINN.