AI for Big Data Problems in Advanced Imaging, Materials Modeling and Automated Synthesis: AI and ML for Imaging and Characterization
Sponsored by: TMS Advanced Characterization, Testing, and Simulation Committee
Program Organizers: Mathew Cherukara, Argonne National Laboratory; Subramanian Sankaranarayanan, University of Illinois-Chicago; Badri Narayanan, University of Louisville
Tuesday 8:00 AM
October 11, 2022
Room: 311
Location: David L. Lawrence Convention Center
Session Chair: Mathew Cherukara, Argonne National Laboratory; Subramanian Sankaranarayanan, University of Illinois Chicago
8:00 AM
Deep Learning-based Algorithms for X-ray Microtomography Analysis: Unravelling Challenges for 4D Experiments: Hamid Torbatisarraf1; Nikhilesh Chawla1; 1Purdue University
X-ray microtomography is an important 3D characterization technique due to its non-destructive nature which provides time-dependent (4D) information. However, image processing and segmentation of 4D tomographic data is extremely time intensive. Moreover, factors such as phase transformation or defect propagation during a time-evolved tomography experiment limits the scan time and/or number of scan iterations.Thus, there is a need to establish a robust algorithm that can render time-dependent x-ray datasets accurately and efficiently. In this talk, we describe the application of Deep Convolutional Neural Network (DCNN) algorithms for X-ray image quality enhancement and segmentation. Using a modified Generative Adversarial Network (GAN) algorithm we provide a workflow to transform low quality x-ray tomography acquired by a fast scan to a high-quality dataset. Our results point to the ability to drastically reduce x-ray data acquisition times, thereby opening a window for efficient 4D experiments.
8:20 AM
Deep Neural Networks for Laser Absorptivity Prediction from Synchrotron X-ray Images: Runbo Jiang1; Joseph Aroh1; Brian Simonds2; Tao Sun3; Anthony Rollett1; 1Carnegie Mellon University; 2National Institute of Standards and Technology; 3University of Virginia
The quantification of the amount of absorbed light is essential for understanding laser-material interactions and melt pool dynamics in order to minimize defects in additively manufactured metal components. The geometry of a vapor depression, also known as a keyhole, in melt pools formed during laser melting is closely related to laser absorptivity. This relationship has been observed by the state-of-the-art in situ high speed synchrotron x-ray visualization and integrating sphere radiometry. These two techniques create a temporally resolved dataset consisting of keyhole images and the corresponding laser absorptivity. In this work, convolutional neural networks (ResNet50 and ConvNeXt) and vision transformer, are adapted to interpret an unprocessed x-ray image of a keyhole and predict the amount of light absorbed. CAM is used to highlight class-specific regions of images.The high-dimensional features extracted by the CNNs are visualized using PCA to identity the behavior of the relationship between the input keyhole geometry and output laser absorptivity.
8:40 AM
Estimation of Sub-micron Carbide Sizes and Morphologies in Dual-Phase Steels from Light Optical Micrographs Using Generative Adversarial Networks: Bo Lei1; Martin Müller2; Dominik Britz2; Frank Mücklich2; Elizabeth Holm1; 1Carnegie Mellon University; 2Saarland University
Mechanical properties of dual-phase steels are highly related to the sizes and morphologies of carbide precipitates. Quantitative measurements rely on image processing of high-resolution SEM micrographs. However, due to the time and cost limitations of SEM imaging, it cannot be used on a large scale. LOM provides fast imaging, but it cannot capture sub-micron characteristics of the carbide precipitates, hence impractical for measurements. Here, we developed an LOM-to-SEM transformation strategy using deep learning and a correlative dataset. We demonstrate that Generative Adversarial Networks (GAN) can be applied to generate high-quality correlative SEM images from LOM images. The approach is further validated by comparing the statistics of the carbide characteristics derived from the synthetic images against real images. The LOM-to-SEM image generation scheme provides a novel route for high-resolution microstructure characterization based only on LOM micrographs.
9:00 AM
Machine Learning Enabled Stacking Fault Energy Prediction in Concentrated Alloys: Dilpuneet Aidhy1; Gaurav Arora1; 1University of Wyoming
High entropy alloys (HEAs) present a paradigm shift in materials design. While these materials present opportunities to unravel novel properties due to a large compositional phase space, they also present an equally large challenge to survey the phase space thereby presenting a data-science challenge. We present a machine learning framework coupled with electronic structure methods whereby properties in complex alloys could be predicted by learning from simpler alloys. A database of charge density is used to predict stacking fault energies in HEAs using regression and neural network models; the latter opens a way to bypass search of descriptors that is a key bottleneck in machine learning methods applied to materials. As the database of simpler materials grows, the self-learning algorithm gradually sharpens its predictive capability and continues to expand into newer material compositions thereby overcoming the challenge of the phase space enormity.
9:20 AM
Phase Identification by Neural Networks Trained from Experimental and Theoretical Structure Data: Nam Le1; Michael Pekala1; Alexander New1; Eddie Gienger1; Janna Domenico1; Christine Piatko1; Elizabeth Pogue1; Tyrel McQueen2; Christopher Stiles1; 1Johns Hopkins University Applied Physics Laboratory; 2Johns Hopkins University
X-ray diffraction (XRD) is a critical tool in high-throughput materials screening, whether to confirm synthesis of target phases or to identify new phases in unexplored systems. However, inverting XRD patterns to estimate which phases are present requires significant human interpretation. Fortunately, machine learning models have shown promise in breaking this bottleneck by automating phase identification. Previous examples have been restricted to phase identification within composition spaces of 4 to 5 elements and relied on databases of experimentally determined structures such as the Inorganic Crystal Structure Database. We will present the applicability to broader discovery settings in unexplored spaces by expanding to phases among 21 elements and synthesizing patterns using computationally determined structures from the Materials Project. These findings will be useful across high-throughput materials discovery settings where properties are strongly tied to particular phases.
9:40 AM
Physics-informed Machine Learning for Selected Area Electron Diffraction Data Analysis: Yu Lin1; Nestor Zaluzec2; Xiaoting Zhong3; Jiadong Gong1; 1QuesTek Innovations LLC; 2Argonne National Laboratory; 3Lawrence Livermore National Laboratory
With the advancement of various electron probe scanning and manipulation techniques such as precession-based electron diffraction, compressive imaging, and hyperspectral imaging – including 4D-STEM, grain structure/orientation mapping has been made possible at nanoscale spatial resolution. This presents exciting opportunities for accelerated multidimensional data real-time processing with high levels of automation. We propose to use a deep learning-based approach for automatic crystal structure identification from the diffraction pattern images. We have demonstrated this concept using a simple type of diffraction data – single phase selected area electron diffraction (SAED) patterns. Deep learning models for automatic crystal structure identification in simulation SAED patterns have been developed by training on simulated data. We also proposed to couple highly relevant physics into ML models. A workflow to improve diffraction pattern analysis ML model performance by incorporating CALPHAD (i.e., CALculation of Phase Diagram) has been developed to improve the performance for phase/structure identification.
10:00 AM Break
10:20 AM
Real-time and Large FOV Ptychography through AI@Edge: Anakha Babu1; Tao Zhou1; Saugat Kandel1; Yi Jiang1; Yudong Yao1; Sinisa Veselli1; Zhengchun Liu1; Tekin Bicer1; Martin Holt1; Antonino Miceli1; Mathew Cherukara1; 1Argonne National Laboratory
X-ray ptychography is a high-resolution imaging technique that relies on the oversampling of the real space information with scattering from a coherent x-ray beam. We present PtychoNN, a novel approach to solve the phase retrieval problem based on deep convolutional neural networks. Trained on experimentally reconstructed data, PtychoNN is both accurate and sample-beamline agnostic. It is capable of predicting phase at a speed up to 100 times faster compared to conventional iterative methods, achieving phase retrieval in real time. Moreover, PtychoNN infers on single shot diffraction patterns rather than an ensemble of data, thus removing the oversampling constraints in ptychography while enlarging its field of view in the process.We shall present results from a recent demonstration of PtychoNN at the edge, performed on the HXN beamline at the APS. We show live stitching of large field-of-view ptychography at 2.5 Hz and real time phase retrieval at 100 Hz.
10:40 AM
Neural Network Prediction of Dynamical Electron Back-Scattered Diffraction Patterns Based on Kinematical Patterns: Clement Lafond1; Marc De Graef2; 1CEA Saclay; 2Carnegie Mellon University
Dictionary indexing (DI) of Electron Back-Scattered Diffraction (EBSD) patterns has shown to overcome issues of commercial Hough Indexing (i.e., lack of robustness against pattern noise) using a comparison between experimental and simulated EBSD patterns. The simulated patterns are computed from dynamical simulations leading to high computation time (up to days or weeks) for complex crystal phases such as precipates in metal alloys or low symmetry geological structures. We propose a new approach to predict dynamical EBSD simulation based on a fast kinematical simulation using Image-to-Image Translation Generative Adversarial Networks (GAN). The accuracy of model, and computation time are evaluated for different crystal structures. The results suggest that such GANs can predict accurate EBSD patterns while reducing the computation time by several orders of magnitude.
11:00 AM Question and Answer Period