2024 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2024): Machine Learning and Data Driven Modeling
Program Organizers: Joseph Beaman, University of Texas at Austin
Monday 1:30 PM
August 12, 2024
Room: 615 AB
Location: Hilton Austin
Session Chair: Srikar Rairao, University of Tennessee Knoxville
1:30 PM
Data Driven and High Fidelity Modeling Approaches to Advance Understanding and TRL Level of 3D Printing: Saad Khairallah1; Amit Kumar1; Justin Patridge1; Gabe Guss1; Eric Chin1; Youngsoo Choi1; Joseph Mckeown1; 1Lawrence Livermore National Laboratory
A multi-scale ALE3D high fidelity model is developed to simulate directed energy deposition. The model captures the powder transport from the coaxial nozzle to the work piece as well as the effect of the carrier gas and laser ray tracing heating and reveals a new kind of air-cushioned. The high cost of modeling is brought down by using deep learning and data driven reduced order modeling at different scales. The end goal is to combine modeling with a data driven approach for “first time right” also referred to here as intelligent feedforward (IFF). We showcase how IFF is used to optimize laser power and scan speed to print complex large parts with overhang geometries and obtain high geometric accuracy.Work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract DE- AC52-07NA27344. Lawrence Livermore National Security, LLC. LLNL-ABS-854766
1:50 PM
Microstructure Prediction in Laser Powder Bed Fusion using Heterogeneous Sensing, Process Modeling, and Machine Learning: Prahalada Rao1; Benjamin Bevans1; Antonio Carrington1; Alex Riensche1; Christopher Barrett2; Harold (Scott) Halliday3; Adrianne Tenequer3; Raghu Srinivasan4; 1Virginia Tech; 2Laser Fusion Solutions; 3Navajo Tech University; 4Wright State University
In this work we predicted microstructural properties in laser powder bed fusion (LPBF) additive manufacturing of Inconel 718 alloy by combining real-time in-situ sensor data and thermal history estimates from a physics-based model within machine learning. This digital twin approach was used to predict multi-scale aspects of microstructure ranging from porosity, solidified meltpool, primary dendritic arm spacing, and microhardness. A physics-based thermal simulation model was used predict the cooling rates in the process. These cooling rate estimates along with real-time in-situ sensor data are then used as features into simple machine learning models to predict the onset of lack-of-fusion porosity, meltpool depth, grain size, and microhardness with a prediction score in excess of 90% (R-square). These models were validated across multiple part geometries and processing conditions.
2:10 PM
Predicting Microstructure Properties Using Transfer Learning: Farzana Tasnim1; Joshua Grose1; Nathan F Sheu1; Remi Dingreville2; Michael Cullinan1; 1University of Texas at Austin; 2Sandia National Laboratories
This study explores the application of transfer learning using pre-trained convolutional neural networks (CNNs) to investigate the relationship between process parameters and microstructure evolution in scientific applications, aiming to predict their correlation with the resistivity ratio between the laser-sintered samples and fully-sintered samples. This method is demonstrated using a dataset comprising microstructures fabricated via micro-Selective Laser Sintering (µSLS). The proposed approach utilizes pre-trained CNNs to extract informative features from Scanning Electron Microscope (SEM) images based on their associated process parameters. These extracted features are utilized to train a fully connected neural network to predict the resistivity ratio. The model achieves high accuracy in predicting the resistivity ratio directly from SEM images, even when dealing with noisy and varied datasets. This approach offers computational efficiency via transfer learning, robust noise handling, and the ability to generalize to unseen data. It has potential for scientific fields needing microstructure analysis and process-property understanding.
2:30 PM
Deep Learning Based Prediction of Thermal History During the Laser-based Powder Bed Fusion (PBF-LB) Additive Manufacturing Process: Philipp Schuessler1; Volker Schulze1; Stefan Dietrich1; 1Karlsruhe Institute of Technology
Understanding the thermal history during the laser-based powder bed fusion (PBF-LB) additive manufacturing process is crucial for minimizing failed printjobs and optimizing part properties. Finite Element Method (FEM) simulations are commonly used but pose challenges for materials with phase transformations due to high computational costs (requirements for temporal and spatial accuracy). In this study, we propose a novel approach utilizing deep learning, specifically the Recurrent Neural Network (RNN) Long Short-Term Memory (LSTM) models, to predict thermal histories based on validated FEM simulations for quench and tempering steel AISI 4140. Our sequential model leverages the thermal history data from multiple simulations as input for training, offering a computationally efficient solution. By accurately predicting thermal histories, our method facilitates the optimization of printing parameters and material properties, thereby enhancing the reliability and performance of additive manufacturing processes for AISI 4140 steel components.
2:50 PM
A Novel Surrogate Model for Scanwise Detailed Process Simulations in Large-scale LPBF Processes: Berkay Bostan1; Shawn Hinnebusch1; David Anderson1; Albert To1; 1University of Pittsburgh
In Laser Powder Bed Fusion (LPBF), part quality significantly depends on thermal conditions related to local geometry, impacting defect formation and microstructure. Due to the substantial scale difference between the laser beam (µm) and the part (cm), conducting centimeter-scale scanwise simulations becomes nearly impossible, often requiring weeks or even months because of the intense computational demands. This study introduces a surrogate model for efficient time-dependent, scanwise thermal simulations in LPBF. The model comprises image processing, deep neural networks, and LSTM (Long Short-Term Memory) units. Its input vector includes a range of features, indicating both scanning strategy and local conductance of various geometrical features, alongside time-dependent heat source locations and preheat temperatures. Achieving more than a 200x computational speedup, this model significantly enhances the feasibility of simulating thermal history, defect, and microstructure formation in centimeter-scale LPBF parts.
3:10 PM Break
3:40 PM
A Surrogate Model for Capturing the Relationship between Physics-based Process Model Parameters and Interpass Temperature History in Laser Powder Bed Fusion Parts: Shawn Hinnebusch1; Alaa Olleak1; Praveen Vulimiri1; Florian Dugast1; Albert To1; 1University of Pittsburgh
Layerwise thermal process simulations are critical for predicting heat buildup in Laser Powder Bed Fusion (LPBF) parts. However, calibrating and validating process simulation models, such as absorptivity and convection coefficients, through optimization methods typically requires hundreds to thousands of simulations. Each simulation can take hours to run, resulting in optimization processes lasting weeks. This study proposes a surrogate model to accurately capture the relationship between simulated temperature history and the model parameters. This surrogate model, constructed via polynomial fitting in a low-dimensional model space obtained through Principal Component Analysis (PCA), enables rapid calibration of model parameters while providing details about the most important calibration parameters. Demonstrations show the surrogate model effectively calibrates multiple geometries with less than 3.3% mean absolute percentage error compared to infrared camera experiments. This approach offers a fast and accurate method for calibrating layerwise process models, facilitating precise temperature field predictions for large LPBF parts.
4:00 PM
Enhancing Predictive Accuracy in Thermal Modeling of Refractory Alloys: A Bayesian Approach Integrating Analytical Models and Experimental Data: Brent Vela1; Peter Morcos1; Cafer Acemi1; Alaa Elwany1; Ibrahim Karaman1; Raymundo Arróyave1; 1Texas A&M University
We aim to identify processing parameters to reduce porosity in additive manufacturing (AM) of refractory alloys. Printability maps based on melt pool dimensions are effective in this regard but are data-hungry. We propose enhancing printability predictions from the analytical Eagar-Tsai (ET) with experimental data using Bayesian methods. To create a thermal model that is both fast-acting and accurate we propose correcting analytical models with experimental data in a Bayesian manner. Specifically, we propose modifying both 1) the prior mean function 2) and the co-variance function of Gaussian Process Regressors (GPR, a Bayesian non-parametric regressor that is defined by a prior mean and co-variance) with the physics captured in the ET model. This essentially creates a data-corrected ET model. Adhering to best practices, we benchmark the effect of the prior mean and the physics-constrained co-variance function using a 2-fold cross validation scheme, demonstrating our method is effective under data-sparse conditions.
4:20 PM
Synergizing Machine Learning and Multiphysics Simulation for Spatter Process Map Generation in LPBF Processes: Olabode Ajenifujah1; Odinakachukwu Ogoke1; Florian Wirth2; Jack Beuth1; Amir Barati-Farimani1; 1Carnegie Mellon University; 2Exentis Group AG
Laser powder bed fusion (LPBF) is the most common metal additive manufacturing (AM) process. However, its full utilization for part production across different industries is inhibited by defects, which limit parts' mechanical properties. Spatters are offshoots that occur due to the complex melt pool dynamics during laser-material interactions in LPBF. Spatter generation is known to promote the formation of defects such as porosity and roughness. In this talk, we will present the characterization of the spatter and the meltpool, using datasets that was procured from the result of modelling LPBF processes via an open-source software known as OpenFOAM. Furthermore, we will discuss the mechanistic insights we derived from evaluating our dataset for classification tasks via machine learning models. Our study culminates in the development of a comprehensive process map, leveraging machine learning to optimize process parameters and mitigate defects, thereby enhancing AM's viability in industrial applications.