Additive Manufacturing Modeling, Simulation, and Machine Learning: Microstructure, Mechanics, and Process: Machine Learning and Artificial Intelligence
Program Organizers: Jing Zhang, Indiana University – Purdue University Indianapolis; Brandon McWilliams, US Army Research Laboratory; Li Ma, Johns Hopkins University Applied Physics Laboratory; Yeon-Gil Jung, Korea Institute of Ceramic Engineering & Technology
Tuesday 8:00 AM
October 11, 2022
Room: 303
Location: David L. Lawrence Convention Center
Session Chair: Jing Zhang, Indiana University - Purdue University Indianapolis; Li Ma, Johns Hopkins University Applied Physics Laboratory; Yeon-Gil Jung, Changwon National University; Brandon McWilliams, CCDC Army Research Laboratory
8:00 AM
Embracing Variability: Machine Learning-based Sequential Optimization of Additive Manufacturing Processes: Maher Alghalayini1; Surya Kalidindi2; Christiaan Paredis1; Fadi Abdeljawad1; 1Clemson University; 2Georgia Institute of Technology
The large variability of Additively Manufactured (AM) materials properties remains a challenge to these emerging techniques. Sequential design is a technique used to explore higher dimensional spaces such as the AM processing parameter space. We develop a sequential design method that implements machine learning to integrate the variability in the properties of AM materials. Our approach learns from previous experiments and adaptively proposes new sites that result in maximum expected information gain. The novelty of our approach lies in the use of Utility Theory to define the optimization criteria and the flexibility in identifying the number of new sites and AM samples. The proposed method is tested on synthetic data to showcase its performance. More specifically, a response surface with multiple peaks in terms of AM laser power and speed is used to benchmark our optimization approach. This study is expected to result in an efficient, variability-embracing adaptive optimization method.
8:20 AM
Generating Novel Porosity Distributions Produced by Metal Additive Manufacturing via Deep Learning: Odinakachukwu Ogoke1; Chris Laursen2; Kyle Johnson2; Michael Glinsky2; Sharlotte Kramer2; Amir Barati Farimani1; 1Carnegie Mellon University; 2Sandia National Laboratories
AM-produced parts can be subject to undesirable porosity, negatively influencing the properties of printed components. Therefore, a precise understanding of the porosity distribution is crucial for accurately simulating potential fatigue and failure zones. In this work, a method for generating new, synthetic, samples of porous parts with novel porosity distributions from limited amounts of training data is presented. To do so, the generation problem is deconstructed into its constitutive parts. First, new examples of the individual pore geometries and surface roughness are created using Generative Adversarial Networks and Scattering Transformations, then, these components are sampled to construct new examples of a porous printed part. The generated parts are compared to the existing experimental porosity distributions based on statistical and dimensional metrics, such as nearest neighbor distances, pore volumes, pore anisotropies and scattering transform based auto-correlations.
8:40 AM
Machine Learning Segmentation Methods for Fatigue Fracture Surface Defect Analyses: Austin Ngo1; David Scannapieco1; Shuheng Zhang1; Shuyue Bian1; Collin Sharpe1; John Lewandowski1; 1Case Western Reserve University
Machine learning algorithms for feature segmentation were utilized on fracture surfaces of LPBF fatigue samples. Process-induced defects were identified and quantified by training an image classification system using SEM images of fatigue fracture surfaces. Process defects were found to range in size, shape, and population due to variations in LPBF build process parameters. Different types of process defects (i.e. lack of fusion, keyhole) were found to be more prevalent based on particular process parameter sets. All process-induced defects across each fracture surface were quantified, with ‘killer’ fatigue crack initiating defects being identified. The fracture surface defect characteristics are compared to the corresponding S-N fatigue data for defect-based fatigue life modeling in a Kitagawa-Murakami-type approach. In addition, this computer vision method is compared to manual identification and quantification of fracture surface defects. The advantages of implementing ML algorithms to streamline fracture surface defect quantification will be discussed.
9:00 AM
Prediction of Microstructure Formation Under Rapid Solidification Using a Deep Learning Approach: Anindya Bhaduri1; Chen Shen1; Alex Kitt2; Lee Kerwin2; Siyeong Ju1; Luke Mohr2; Yang Jiao1; Marissa Brennan1; Shenyan Huang1; Sreekar Karnati1; Monica Soare1; Arushi Dhakad2; Hamedreza Seyyedhosseinzadeh3; Liping Wang1; Changjie Sun1; Lang Yuan3; 1GE Research; 2EWI; 3University of South Carolina
In this study, a cellular automaton-based solidification model was considered to predict the grain structures in single tracks during laser scanning under additive manufacturing conditions. The model directly utilized the thermal history calculated via Rosenthal solution under a wide range of process parameters. The challenge here is that the simulation involves a complex spatiotemporal stochastic solution and thus very expensive to solve. To make the predictions fast with sufficient accuracy, the goal is to develop a machine learning framework to efficiently map the process parameters to the final grain structure. Specifically, a probabilistic deep learning model is developed that can successfully tackle the issues of efficiency, accuracy, and stochasticity.
9:20 AM
Using Generative Adversarial Networks for the Design of Metamaterials to Reach New Property Spaces: Chandra Veer Singh1; Sahar Choukir1; 1University of Toronto
Recent advances in additive manufacturing and machine learning are ushering in a new age of data-driven material design. Guided by bio-inspiration, experimentation and systematic optimization, a number of metamaterials have been synthesized with mechanical properties reaching theoretical stiffness limits. However, experimental design of such materials remains challenging. Here, we present an approach for the design of metamaterials with optimal stiffness using machine learning approaches via generative adversial networks. Finite elements models were carried on millions of randomly generated 3D architectures based on different crystallographic symmetries to extract young’s moduli, shear moduli and bulk moduli. The data served to train the networks and identify hundreds of new metamaterials designs at the unit-cell level with our target mechanical property: optimal stiffness. The significance of the approach lies in the development of a ML-based platform that allows computers to design novel 3D isotropic metamaterials achieving the Hashin-Shtrikman upper bound with no prior knowledge.
9:40 AM
Using Machine Learning to Characterize Powder Behavior and Surface Roughness in Powder Bed Fusion AM: Srujana Rao Yarasi1; Elizabeth Holm1; Anthony Rollett1; 1Carnegie Mellon University
In powder bed fusion additive manufacturing (AM), characteristics of the powder feedstock such as particle size distribution, sphericity, and morphology affect the flowability of the powder and the layer density distribution of the powder bed. The use of computer vision and machine learning tools in the AM domain have enabled the quantitative investigation of qualitative factors like powder morphology. The use of Convolutional Neural Networks (CNN) to generate feature descriptors is proposed as part of a framework to generate powder morphology distributions that describe morphological characteristics of the powder. Similarly, the measurement of surface roughness is an image data-rich problem that can benefit from characterization with machine learning techniques. There are multiple factors that affect surface roughness including powder size and process parameters. ML techniques are used to understand the correlation between these different factors and surface roughness metrics.
10:00 AM Break
10:20 AM
Additive Manufacturing Moment Measure: A Reduced Order Model of the Laser Powder Bed Fusion Process: Samuel Hocker1; Brodan Richter1; Joseph Zalameda1; Peter Spaeth1; Erik Frankforter1; Andrew Kitahara2; 1NASA; 2National Institute of Aerospace
The multi-scale and complex process of printing additively manufactured (AM) parts can have unexpected, but predictable, build conditions that result in material microstructure variability. In this work, we describe a fully parallel reduced order modeling approach that has been developed to evaluate the evolution of AM processes, termed the AM moment measure method. This method couples the known sequence of the AM process with a physically informed nearest neighbors’ calculation to map the conditions of a part-scale build. The result is a map of the build that is derived directly from build files or in-situ process monitoring sensors. The methodology and terminology of the approach will be described, and computed build maps will be calculated and compared for various laser powder bed fusion (LPBF) builds of Ti-6Al-4V. Such comparative results develop understanding of how the sequential process actions can affect the LPBF-AM build quality and microstructure variability.