First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022): Poster Reception
Program Organizers: Taylor Sparks, University of Utah; Michael Dawson-Haggerty, Kerfed, Inc.; Elizabeth Holm, University of Michigan; Jin Kocsis, Purdue University; Adam Kopper, Mercury Marine; Benji Maruyama; James Warren, National Institute of Standards and Technology
Monday 5:20 PM
April 4, 2022
Room: Sternwheeler
Location: Omni William Penn Hotel
Deep Learning Based Clustering Technique to Identify and Generation of RVE Models in Duplex Structure Stainless Steels: Israr Ibrahim1; Ramana Pidaparti1; 1University of Georgia
Duplex structural stainless steels (DSS) are used in chemical, petrochemical, marine, power generation, and nuclear industries due to their unique properties. The DSS nanophases (ferrite and austenite) are made of numerous topological properties, such as perimeter length, area, orientation, and centroid. The objective of this study is to characterize each nanophase units based on its topological properties. The nanophase units can be represented as Representative Volume Elements (RVE) and are extracted from the pixel image by computing approaching the contour through marching square algorithm. Topological properties then are calculated. A dataset linking each nanophase unit shape with topological properties are generated using clustering algorithms The results of RVE obtained from clustering techniques to demonstrate the deep learning approach in generating variety of RVEs. Currently, more results are being obtained in relating RVEs and their mechanical behavior and will be presented at the conference.
Neuromorphic Utilization of NVM Devices Using GeTe: Chaeho Lim1; 1SUNGKYUNKWAN UNIV.
This study relates to the possibility of neuromorphic utilization of devices with metal-insulator transition materials. Chalcogenide compounds are the most important substance using in RRAM, due to their unique behavior(metal-insulator transition). Various efforts are being made to discover more improved materials that satisfy switching characteristics (on/off ratio) and memory device behavior (microstructure change). In this study, Germanium–Tellurium compounds are used as Metal-Insulator Transition materials. It is also constructed in MIM structures for DC/AC electrical measurements. Volatile / Non-volatile Memory I-V characteristics (unipolar/bipolar) were controlled by multilayer structure. (GeTe+ Other). Impedance measurements were performed to correlate structure and characteristics. The results of this study demonstrate the availability of volatile, non-volatile memory devices from GeTe materials. Furthermore, it is expected to provide insight into the possibility of neuromorphic utilization of NVM devices using GeTe.
Prediction of Stress-Strain Curve of High-Entropy Alloys and 3D Printed Steels Using Machine Learning: Shalini Priya1; Nitish Bibhanshu1; 1Indian Institute of Technology Patna
Predicting stress-strain curves is a challenging problem in material science due to non-linear properties of materials. Recently, artificial intelligence based models are found efficient for modelling such relationships and represent important material characteristics. In this work, we present a deep neural network (DNN) approach for strain-strain curve prediction. The stress-strain curves of high entropy alloys as well as of the steel were selected for the different range of temperatures. However, it’s been observed that any specific model is not able to sufficiently capture the properties across the materials hence we present an ensemble based technique of different DNN’s with its variations in parameters for fine-tuning and get the best prediction curve with minimal error. We compare our proposed model performance with standard machine learning and neural network predicted curves as well as with standard state-of-the-art techniques. We further compare our ensemble model with experimental data generated by the machine.
Modeling and Simulation of Additively Manufactured Lattice Structures to Support Component Qualification: Andrew Swanson1; 1Sandia National Laboratories
Additively manufactured (AM) lattices and cellular structures give designers unique capabilities such as tunable material properties, custom surface texturing, and the ability to optimize a component’s strength and/or weight. Designers are faced with challenges for the qualification of these structures caused by high costs and the need to quantify AM material variability and reliability. To address this barrier a qualification approach for lattices that doesn’t use extensive hardware testing is needed. This can be accomplished by coupling high-fidelity finite element simulations validated through physical testing with machine learning algorithms. By conducting designed computer experiments with the validated finite element model gaussian process metamodels can be developed. These metamodels are then used for Monte Carlo simulations that assess the margins and uncertainty of key performance indicators. This data can be used as qualification evidence to show a component can reliably meet performance requirements for the material uncertainty from additive manufacturing.
Computer Vision and Machine Learning Methods to Characterize Recycled Powders for Additive Manufacturing: Nathan Love1; Srujana Yarasi1; Andrew Kitahara1; Elizabeth Holm1; 1Carnegie Mellon University
The properties of additive manufacturing feedstock powders change with recycling, but the characteristics of the powder that are responsible for these changes are not known. Three different approaches were developed to represent SEM images of metal powders in order to associate their visual appearance with the degree of recycling. The performance and shortcomings of the models are compared and examined. One of these three techniques, a novel method that aggregates computer vision features, was able to predict the degree of recycling of a powder with high accuracy and also to suggest some of the changes in individual particles that may be responsible for altered properties. This high-performing method was also applied to another powder characterization task based on material type and again succeeded with high accuracy, which suggests high potential as a general powder characterization method.
Optimization of the Additive Manufacturing Process for Refractory Metals Using Numerical Simulations and Machine Learning: Adrian Dalagan1; Damilola Lawal1; Kyle Snyder2; Prasanna Balachandran1; Richard Martukanitz1; 1University of Virginia; 2Commonwealth Center for Advanced Manufacturing
Refractory metals, including alloys of niobium, tungsten, and tantalum, offer significant opportunity to increase operational temperatures for a wide range of applications, and additive manufacturing is considered a critical enabling technology for wider utilization of these materials. However, the implementation of this process is challenged by inherent properties such as high melting temperatures and oxidation rates, which can form compromising defects such as lack-of-fusion. Additionally, the resultant microstructures influenced by these conditions must also be considered. Thus, this research investigated parameters for adequate energy input and then developed optimal processing maps by factoring these constraints. Numerical simulations predicted melt pool geometry and thermal response over a range of process variables. The thermal data also informed the phase field model for predicting solidification morphology and concentration redistribution. Finally, the compilation of the simulation results and experimental validations allowed for the construction of a machine learning model that defines the optimal parameters.
Automating Solid-Liquid Interface Identification in Additive Manufacturing Simulator Experiments: Gus Becker1; 1Colorado School of Mines
Additive manufacturing (AM) simulator experiments performed at the Advanced Photon Source (APS) at Argonne National Laboratory (ANL) explore the effects of rapid solidification phenomena on resulting microstructures. High-speed x-radiography captures the laser melting of metallic samples through time. Measuring the solidification velocity in these experiments is a difficult and time-consuming task consisting of manually locating the solid-liquid (S-L) interface and tracking the change in location of the interface from one frame to the next. This work automates this interface identification through traditional image processing with Python libraries including NumPy and scikit-image.
Productivity Enhancement of Photolithography by Big Data Learning: Juyoung Jung1; Hee Joon Jung2; ByoungDeog Choi1; 1Sungkyunkwan University; 2Northwestern University
Photolithography, consuming the longest time and the highest cost among manufacturing processes, has developed to achieve maximum productivity over the past few decades. To maximize productivity, the usage time of the equipment is categorized through a time-based productivity model, which is based on big data comprised total 237 billion data in Samsung electronics for six months. These enormous data sets are designed to find causes of productivity degradation faster and more accurately than ever before. On the other hand, productivity improvements often lead to quality degradation. To prevent this trade-off, the overall equipment efficiency is designed to include the quality element of products. This paper could categorize productivity improvement into four sectors: continuity, synchronization, elimination, and reduction based on the time-based productivity model. Our study would pave the road to utilize a systematic approach to big data for the lithography process.
Modeling a Monte Carlo Potts Solidification Model Using a Generative Adversarial Network: Gregory Wong1; Anthony Rollett; Gregory Rohrer; 1Carnegie Mellon University
Machine learning can be used to reduce the order of microstructure and grain orientation modeling. These reduced order models can be used to produce many synthetic microstructures at greatly reduced computational expense. The generative adversarial network (GAN) class of neural networks is used in this work to simulate a simple Monte Carlo Potts solidification model with an end goal of modeling solidification in AM. The GAN model consists of a generator network that creates artificial images and a discriminator network that labels the produced images as belonging to a training set of images or not. Additionally, a layer of conditionality can be added to the GAN to generate images that belong to a specific labeled class. Such a class could be the material’s processing conditions, allowing for artificial microstructures to be generated based on specific processing conditions. Training and output images along with model structure will be presented.
Effect of Interlayer Delay Time on the Melt Pool Dimensions in Direct Energy Deposition Process using Machine Learning Techniques: Rajib Halder1; Anthony Rollett; Amit Verma; Zhening Yang; Ali Guzel1; Anthony Rollett; 1Carnegie Mellon University
Ti-6Al-4V is a high-performance two-phase (a + B) alloy, where properties strongly depend on the microstructure of the as-printed parts. Since microstructure is defined by the cooling rate, which also correlates with the melt pool dimensions, it becomes important to study the correlation between the melt pool dimensions and the cooling rate as a function of process parameters for the desired microstructure and geometry. In this study, we used the Random Forest (RF) algorithm to predict the final bead dimensions and established a correlation between the heat accumulation with interlayer delay time during the build. The process-microstructure-property relationships were investigated using both, Canonical Correlation Analysis and RF algorithm.