First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022): On-Demand Oral Presentations
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 8:30 AM
April 11, 2022
Room: On-Demand Session Room
Location: Omni William Penn Hotel
A Walk Through Material Microstructures: Using Deep Learning and Geometry to Better Visualize Large Collections of Material Microstructure Images: Jordan Weaver1; Henry Kvinge2; 1University of Washington; 2Pacific Northwest National Laboratory
The increasing rate at which material microstructure imagery is being collected has led to a need for better tools to explore and understand these large datasets. This means understanding not only individual images, but also relationships between collections of images from multiple samples. A common approach to achieving this understanding is to leverage data-driven visualizations that represent relationships spatially. Unfortunately, image data is generally very high-dimensional, and thus an image dataset can have complicated structure that cannot be captured in the 2 or 3-dimensions that the human visual system is limited to. We describe a method called DeepStroll that we have developed to overcome this challenge. Our method combines state-of-the-art computer vision models and principles from differential geometry to construct a 'tour' through an image dataset. We use the temporal component of the resulting video to capture structure in the dataset that cannot be summarized in a single visualization.
Development of Recipe Optimization Method for Additive Manufacturing Process Parameter Determination: Steven Osma1; Jue Wang2; Kousuke Kuwabara2; Hyakka Nakada3; Shinji Matsushita1; Hirotsugu Kawanaka1; Minseok Park1; Yusuke Yasuda1; 1Hitachi Limited; 2Hitachi Metals Limited; 3Recruit Company Limited
Additive Manufacturing (AM) is a process that forms a three-dimensional structure directly from a CAD file. AM processes typically demand fine tuning of a large subset of process parameters (recipes, hereafter). Due to the many parameters, a trial-and-error approach has proven insufficient for determining recipes for new materials. In order to enable rapid recipe determination, we developed a recipe optimization method which is a set of machine learning algorithms including a kernel-ridge regression, multi-dimensional optimization, and feature selection algorithm. The developed method was applied to determine parameters for Selective Laser Melting process of ADMUSTER-C21P ®, a Ni-base alloy, ADMUSTER-C00P ®, a high-entropy alloy, and ADMUSTER-W285P ®, a maraging steel. Utilizing this method along with a thermal deformation measuring technique, the recipe determination process was expedited by 50%. Obtained recipes enable parts of superior quality compared with parts built using traditional recipes.
Machine Learning Guided Design of Aluminium Alloys: Ninad Bhat1; Amanda Barnard1; Nick Birbilis1; 1The Australian National University
Aluminium (Al) alloys continue to remain critical in the aerospace and transportation industries due to their high specific strength and low density. The discovery of new Al-alloys is still primarily guided through the empirical and "trial and error" process. This approach creates a bottleneck in alloy development, as it is time-consuming and expensive. In the present work, machine learning techniques have been used to screen and predict new Al-alloys targeting Yield Strength, Tensile Strength and Elongation. A dataset of aluminium alloys, including the processing condition and alloy concentration, was curated. A Random Forest Regressor was used to predict mechanical properties and analyse the features contributing to the mechanical properties. Further, these regression models can be combined with Evolutionary Algorithms to predict new aluminium alloys.
Additive Manufacturing of Aluminium: Alloy Design and Machine Learning Assisted Process Optimisation: Xiaopeng Li1; Qian Liu1; Jay Kruzic1; 1University of New South Wales
Laser powder bed fusion technique (LPBF) has been widely used to fabricate various aluminium alloys in the past decade. However, due to intrinsic materials characteristics, e.g., solidification cracking, not many aluminium alloys have satisfactory LPBF processability. Therefore, it is in urgent need to design and develop more suitable aluminium alloys and their composites for LPBF. Meanwhile, once new aluminium alloys are designed, it is also of great importance to optimise the LPBF process to achieve high quality components without any apparent processing defects such as cracks or porosity. In this presentation, a novel in-situ alloy design process was first introduced to develop nanoparticle decorated aluminium alloys for LPBF and the resultant microstructure along with mechanical properties were investigated. Following this, a machine-learning assisted LPBF process optimisation process for aluminium alloys is described in detail to provide new insights into the microstructure control and properties manipulation of LPBF fabricated aluminium alloys.
A Semi-Supervised Approach to Characterizing Multiple Morphological Features in Microstructure Images: Arun Sathanur1; William Frazier1; Jing Wang1; Ram Devanathan1; 1Pacific Northwest National Laboratory
Complex microstructures comprise of multiple structural features such as grains, phases, voids, and precipitates. Automatic statistical characterization of individual volume fractions is challenging. Recent advances in deep learning using supervised machine learning rely on large, labeled datasets. Labeling large datasets of complex microstructure images at the pixel-level is an intractable task. Further these models also suffer from lack of interpretability. In this work, we present an interpretable semi-supervised machine learning approach to solve this problem. This approach leverages unsupervised computer vision approaches to first segment the image into feature regions. Next, it builds interpretable descriptions of the segmented regions using Hu and Zernike image moments. Finally, with user inputs on a small number of labels, it uses a similarity graph between the segmented regions and a graph convolutional network to predict the labels on all of the segmented areas. We demonstrate our approach on synthetic and real-world microstructure datasets.
Methods of Surface Inspection for Plane Metal with the Use of CV: Maxim Shamshin1; 1United Metallurgical Company (OMK)
The paper analyzes the application of methods and computer vision algorithms based on the determination of texture features for the problem of flat rolled products surface defects detecting from the image. It is shown that the efficiency of deterministic methods is not inferior to the efficiency of methods based on deep learning algorithms and artificial neural networks, which currently have difficulties for their practical application in industry. At the same time, deterministic algorithms make it easy to interpret intermediate and final analysis results. The experience of industrial application of deep leаrning and deterministic algorithms in metallurgical production is also considered.
Automated Microstructure Property Multi-Classification of Ni-Based Superalloys Using Deep Learning: Irina Roslyakova1; Uchechukwu Nwachukwu1; Abdulmonem Obaied1; Oliver Horst1; David Bürger1; Muhammad Adil Ali1; Ingo Steinbach1; 1ICAMS, Ruhr-University Bochum
This study presents a solution for automated multi-classification of SEM/TEM images of Ni-based superalloys with respect to creep strain values, heat treatment conditions and chemical compositions using deep learning methods by training convolutional neural network (CNN) in two steps. First, phase-field simulations that displayed similar results to experiments were utilized to build a model with pre-trained CNN architectures (e.g. AlexNet, ResNet34). Then, the optimized hyper-parameters were refined by re-training the CNN with experimental SEM-images of Ni-based superalloys. This fine-tuning process was applied to compensate for the lack of “big data” of available experimental images while training the model. The proposed model was tested on micrographs from other existing publications and showed a promising performance in identifying and predicting strain levels, heat treatment conditions and chemical compositions of SEM/TEM micrographs. Moreover, the refined model is proved to be independent of image scale size.
Bayesian Deep Learning Methods for Microstructural Feature Characterization of LiAlO2 Pellets: Karl Pazdernik1; Alexander Hagen1; Marjolein Oostrom1; Nicole LaHaye1; 1Pacific Northwest National Laboratory
LiAlO2 is an important material that is used as a tritium producer for the Tritium Sustainment Program. To better understand the tritium release from the material during irradiation, comprehensive microstructural analysis of unirradiated and irradiated LiAlO2 is required. Recently, deep learning has been employed as a fast approach to classifying various microstructural features in LiAlO2 pellets that are visualized by scanning electron microscopy (SEM), including grains, grain boundaries, voids, precipitates, and zirconia impurities. While these methods produce high overall accuracy, the boundaries between microstructural features are predicted with higher error. Since aggregate characteristics, such as defect area and relative proportion, are highly dependent on these boundaries, these errors should not be ignored. To address this concern, we present a Bayesian deep learning approach to uncertainty quantification in the semantic segmentation of SEM images. We highlight areas where prediction is less certain and discuss the impact on physics-based modeling.
Machine Learning from Large and Sparse Data for Novel Materials Discovery: Fadwa El-Mellouhi1; 1Qatar Environment and Energy Research Institute, Hamad Bin Khalifa University
Materials discovery has prospects of significant acceleration in the upcoming years thanks to the adaptation of data-science and machine learning techniques. In this talk, I will give an overview of the latest advances in this field with focus on materials for energy and environmental applications. I will show how large and sparce datasets constructed from density functional theory (DFT) calculations or experiments can be used to perform a systematic analysis of structure-to-property relationships. Focusing on the correlations between the structural deformations and the thermodynamic stability of compounds, various machine learning algorithms were trained then tested. I will also highlight how our approach offers an interesting guideline on how to engineer novel materials compositions enabling to reduce the huge space of experimental trial and error.
AI/ML for Intelligent Design and Processing of Metal Castings: Jiten Shah1; 1PDA LLC
Metal casting design and manufacturing process has many variables and there is uncertainty involved. Author will present meta models developed using AI/ML for predicting properties and porosity utilizing historical production data and controlled experiments. The meta models can be used for a near real time intelligent processing of the castings in a production environment. The author will present the methodology and the framework utilizing a hybrid approach of physics based ICME simulation and the machine learning techniques with uncertainty quantification. Meta model driven design guidance tool for predicting the local properties in various section thickness will be presented.
DeepTemp: Predicting Material Processing Conditions with Artificial Intelligence: Colby Wight1; Sarah Akers1; WoongJo Choi1; Tegan Emerson1; Luke Gosink1; Elizabeth Jurrus1; Keerti Kappagantula1; Xiaolong Ma1; Scott Whalen1; Reza Rabby1; Tianhao Wang1; Timothy Roosendaal1; Nicole Overman1; Henry Kvinge1; 1Pacific Northwest National Lab
Material development using advanced manufacturing processes where only nascent physics-based understanding is available can be resource-intensive. Because of this, there is an opportunity to accelerate R&D within the field by leveraging the predictive power of deep learning to aid in experimental design. We describe how this approach was applied to a PNNL developed forming technology called Shear Assisted Processing and Extrusion (ShAPE), using recurrent neural networks to predict the temperature that a material experiences during processing. Temperature is an important factor in the ultimate material microstructure and properties of a sample produced by ShAPE, but which the experimenter does not directly control. We show how recurrent neural networks can be trained to reliably predict the temperatures a material experiences throughout the extrusion process making experiment planning more efficient.
xT SAAM – An Industrial Small Data AI Platform(Design-Expert is Trying to Simplify Classical DoE. Making it a Bit Easier to Use): Varun Gopi1; Matthias Kaiser; 1Exponential Technologies LTD.
Global optimization based on AI or active learning. That is, the user of our software does not have to select a suitable design since the software generates its own agile design based on the answer. Because of this, our software does not scale the number of tests with the granularity of the search field. The number of experiments required does not grow exponentially with the number of parameters/dimensions. That implies we require less genuine trials in general than any other DoE, especially when you go above four parameters/dimensions.Most importantly, the user of our software does not need any statistical expertise to use it, and in theory, if automated feedback is given, the optimist may be eliminated. Because no user input is necessary between iterations, the optimization process may be completely automated.
Data-Driven Learning of Constitutive Laws and Material Parameter: from Molecular Dynamics to Continuum Models: Marta D'Elia1; 1Sandia National Laboratories
In this talk I will present a machine learning technique to obtain accurate surrogates that reproduce molecular dynamics (MD) simulations at coarser scales. Starting from MD data of graphene at different temperatures, we first apply a coarse-graining technique to project the data onto a much coarser grid and then use coarse-grained data to train a nonlocal, continuum model that accurately reproduces MD data from a validation set. Our results for a perfect crystal and in presence of thermal noise illustrate the ability to recover material parameters and show excellent generalization properties of our learning algorithm, enabling transfer learning.
Artificial Materials Intelligence (AMI) of Creep Indicator Model (CIM) in Single Crystal Super Alloys: Irina Roslyakova1; Yuxun Jiang1; Abdulmonem Obaied1; Uchechukwu Nwachukwu1; Muhammad Adil Ali1; Ingo Steinbach1; 1ICAMS, Ruhr-University Bochum
A novel modeling strategy, which combines artificial intelligence (AI) with artificial materials (AM) will be proposed and called Artificial Materials Intelligence (AMI). We define AMI as a hybrid physically-based data-driven modeling strategy that uses well-established AI-components, such as mathematical algorithms, statistics, machine and deep learning, computer vision, robotics, and others to analyze heterogeneous simulated and experimental data on different modeling scales. To build the so-called creep indicator model (CIM) using AMI modeling strategy, physically-based and data-driven models are combined to identify statistically sound correlations between materials chemistry, thermodynamics, microstructures and mechanical data of single crystal Ni- and Co-based super alloys. The proposed CIM allow us to identify the influence of individual physical effects from the considered contributions on selected material properties. Moreover, they are frequently required to accelerate the computer-assisted design of new materials and alloys, for example, the development of rhenium-free Ni-based.
Predicting Multicomponent Alloy Properties with Neural Network Surrogate Models: Jong Youl Choi1; Massimiliano Lupo Pasini1; Ying Yang1; Jian Peng1; Dongwon Shin; Sam Reeve1; Paul Laiu1; 1Oak Ridge National Laboratory
CALPHAD is a crucial tool in materials science for predicting phase stability and thermophysical properties in multicomponent systems. However, development of a CALPHAD database is time consuming, especially for highly concentrated alloys that require the unary, binary, and ternary systems to be modeled for an n-component database. Judicious use of machine learning can significantly speed time to prediction, as well as improve accuracy; in this work we develop surrogate models using CALPHAD data to train neural networks (NN). Using the unary and binary data for various thermodynamic and thermophysical quantities of interest (QoI), we train NN to predict the same QoI for ternary systems, as well as higher component systems. We will discuss material systems including high entropy alloys, issues of data scarcity and coverage, and potential paths to improve the approach.
Assessing the Robustness of an EBSD-Data-Based U-Net Model to Classify Phase Transformation Products in Steels: Tomas Martinez Ostormujof1; Simon Breumier2; Nathalie Gey1; Mathieu Salib3; Lionel Germain1; 1Université de Lorraine, CNRS, Arts et Métiers ParisTech, LEM3, F-57000 Metz, France; 2Institut de Recherche Tecnologique Matériaux, Métallurgie et Procédés (IRT M2P) 4 rue Augustin Fresnel 57070 METZ; 3ArcelorMittal Maizieres, Research and Development, Voie Romaine, BP30320, F-57283 Maizieres-les-Metz, France
Quantitative characterization of complex steels microstructures requires a considerable amount of time, effort and expertise. To address this issue, we propose to couple Artificial Intelligence techniques with Electron Backscattering Diffraction (EBSD) to improve and simplify the task. We have trained a supervised Deep Learning model capable of classifying and quantifying Dual Phase (DP) steel microstructures based on a semantic segmentation strategy using Convolutional Neural Networks and the UNET architecture. DP steel microstructures were classified in seconds with an accuracy of ~98%.In this contribution, we assess the robustness of our model against the variability of the input (influence of sample preparation and EBSD acquisition set-up). Additionally, we discuss the amount of data needed to train an accurate model, including the contribution of simulated EBSD microstructures and data augmentation. Finally, we demonstrate the applicability of our model to analyze more complex microstructures including Widmanstätten ferrite and Bainite.
Nanoindentation Load-Displacement Analysis Using a Genetic Algorithm: Abe Burleigh1; Andy Lau2; Jeff Terry1; 1Illinois Institute of Technology; 2Boise State University
An automated tool for analysis of nanoindentation load-displacement curves using a Genetic Algorithm (GA) with the Oliver Pharr method is presented. Some materials, such as polycrystalline isotropic graphites, are difficult to fit using least squares methods. At the indentation depths required for reproducible results in these graphites the material cannot recover significantly during unloading. This results in hard to fit sharply-peaked unload curves that result in overestimation of the parameter describing indenter tip geometry. GA, a robust metaheuristic method, automatically processes batches of nanoindentation data with minimal user input while producing physically meaningful parameters. The GA begins with a population of temporary solutions; from these we find a fitness value for each solution, and select the best. These are mixed with random solutions to crossover producing the next generation. Following this a mutation operator is applied to existing solutions by random perturbations, and finally the optimal solution is selected.
Correlation Between Additive Manufacturing Process Parameters and Microstructural Descriptors Via Automatic Feature Engineering: Mohamed Heddar1; Mehdi Brahim2; Nedjoua Matougui1; 1ENSMM - Annaba; 2USTHB
This paper explore the use of automatic feature engineering coupled with machine learning for Additive Manufacturing (AM) processing of metals, to derive empirical relationships between processing parameters and microstructural descriptors. A kinetic Monte Carlo algorithm is used to generate the microstructures. The resulting artificial microstructures were then analyzed using image processing to extract key morphological properties. Additional non-linear transformations and combination using arithmetic operators were then applied on to the features dataset. This transformation increases the performance of the linear model for predicting all the microstructural descriptors, for grain size prediction the R2 score improves from 0.4 to 0.902, similar improvements were also recorded for other descriptors. To further simplify the models, the number of added features was reduced by analyzing feature importance and coefficients magnitude while retaining a reasonable prediction accuracy.