Characterization of Materials through High Resolution Imaging: Algorithms for High Resolution Coherent Imaging of Materials
Sponsored by: TMS Structural Materials Division, TMS: Advanced Characterization, Testing, and Simulation Committee
Program Organizers: Richard Sandberg, Brigham Young University; Ross Harder, Argonne National Laboratory; Xianghui Xiao, Brookhaven National Laboratory; Brian Abbey, La Trobe University; Saryu Fensin, Los Alamos National Laboratory; Ana Diaz, Paul Scherrer Institute; Mathew Cherukara, Argonne National Laboratory
Thursday 2:00 PM
March 18, 2021
Room: RM 14
Location: TMS2021 Virtual
Session Chair: Mathew Cherukara, Argonne National Laboratory
2:00 PM Invited
Optimization Based Approach for 3D Alignment in X-ray Nano-tomography: Kanupriya Pande1; 1Lawrence Berkeley National Laboratory
The goal of tomography is to reconstruct an unknown 3D object from a series of 2D projections that are acquired as the object is rotated about one or more axes. As the resolution of tomography X-ray microscopes improves, calibration errors such as offsets of the X-rays and undesired wobble of the rotation stage result in a loss of spatial resolution and an enhancement of image artifacts. Here we present an algorithmic approach to correct misalignments in tomographic data that alternates between a reconstruction step based on regularized iterative algorithms, and an alignment step based on non-linear least squares optimization. We discuss strategies to avoid local minima, and demonstrate the performance of the algorithm on simulated and measured datasets.
2:30 PM
Adaptive Machine Learning for 3D Bragg Coherent Diffraction Imaging Reconstructions: Alexander Scheinker1; Reeju Pokharel1; 1Los Alamos National Laboratory
We present an adaptive machine learning (ML) framework for the iterative phase retrieval problem faced by 3D coherent diffraction imaging (CDI) experiments. We combine 3D convolutional neural networks with adaptive model independent feedback which allows us to handle time varying systems for which ML methods alone are insufficient because their predictions drift if the system for which they were trained changes with time. For example, this regularly happens at CDI experimental beam lines due to changes in instrument setups and fluctuations of the light source. The proposed method is a novel robust approach to 3D reconstruction based on diffraction intensity measurements. We demonstrate the effectiveness of the method with both synthetic 3D structures and with measured complex crystal grain shapes.
2:50 PM Invited
Exploiting Machine Learning Techniques in X-ray Ptychography: Pablo Enfedaque1; 1LBNL
Ptychography permits imaging macroscopic specimens at nanometer wavelength resolutions while retrieving chemical, magnetic or atomic information about the sample. It is a remarkably robust technique for the characterization of nano materials, being currently used in a variety of scientific fields. The main challenge of ptychography resides in solving a phase retrieval problem in order to retrieve a reconstruction of the imaged specimen. However, the end-to-end reconstruction normally also involves post-phase retrieval operations, e.g. segmentation, denoising, or super resolution. Currently, machine learning techniques are scarcely used on ptychographic reconstruction pipelines on DOE synchrotron facilities. This talk will present the challenges, state of the art, and preliminary results on the use of machine learning techniques for X-ray ptychography reconstruction. Exploiting such techniques has the potential to enable smart-, sparse-scanning and analysis, which would reduce the acquired data by orders of magnitude while heavily reducing the computational cost of an end-to-end ptychographic experiment.
3:20 PM
Ptychographic Inversion with Deep Learning Network and Automatic Differentiation: Tao Zhou1; Mathew Cherukara1; Saugat Kandel1; Stephan Hruszkewycz1; Alexander Hexemer1; Ross Harder1; Pablo Enfedaque1; Martin Holt1; 1Argonne National Laboratory
In this work, we demonstrate how machine learning and its related optimization tools can be used to replace conventional phase retrieval methods in X-ray transmission and Bragg ptychography. X-ray transmission ptychography has become a well-established technique for high resolution imaging and phase retrieval. We present PtychoNN, a novel approach to solve the ptychography problem based on deep convolutional neural networks. Once trained, PtychoNN is capable of generating high quality reconstructions up to hundreds of times faster compared to conventional iterative methods, essential for implementing on-the-fly phase retrieval. Moreover, by surpassing the numerical constraints of iterative methods, the sampling condition can also be significantly relaxed.The counterpart of transmission ptychography in diffraction condition is known as Bragg ptychography. The technique itself is less mature, as limited by the more complex diffraction geometry and data quality. Here we describe the forward propagation in Bragg ptychography using the Takagi-Taupin Equations (TTE). We show that, when combined with Automatic Differentiation (AD), TTE can be used as a general formalism for 3D phase retrieval, applicable to both Bragg ptychography and Bragg Coherent Diffraction Imaging. Compared to conventional Fourier Transform based methods, our approach accounts for additionally refraction, absorption, interference, dynamical effects, and is applicable to any kind of weakly strained material system.
3:40 PM
Image-based Simulation of Permeability and Image-to-Mesh Conversion of X-ray Tomographic Images of a Nickel Foam: S. Ali Shojaee1; Arsalan Zolfaghari1; 1Thermo Fisher Scientific
Nickel foams with high surface to volume ratio and lightweight structure have found applications in filtering, catalysis, and heat exchangers. Accurate measurement of flow in the porous structure is crucial for development and optimization of many applications. In this study, tomographic images of nickel foams (courtesy of Professor Patrick Perre, Barbara Malinowska, Cyril Breton, laboratory LGPM, Centralesupelec) were used to measure the pore-density and ligament thickness distribution. In addition, the permeability of the porous structure was directly estimated from the tomographic images using the Lattice Boltzmann method. A cluster environment was used to execute the high-performance calculations of the permeability simulations. The permeability was also estimated via the representative pore network model (PNM) and the results of the two techniques were compared with each other. Finally, the image was converted from 3D CT image to a tetrahedral mesh with different accuracies for the foam, open pores, and closed pore spaces.
4:00 PM Invited
Using Phase Field Simulations to Train Convolutional Neural Networks for Segmentation of Experimental Materials Imaging Datasets: Tiberiu Stan1; Jiwon Yeom2; Seungbum Hong2; Peter Voorhees1; 1Northwestern University; 2Korea Advanced Institute of Science and Technology
The ability to quickly analyze large imaging datasets is vital to the widespread adoption of modern materials characterization tools, and thus development of new materials. Image segmentation can be the most subjective and time-consuming step in the data analysis workflow. We show that it is possible to segment experimental x-ray computed tomography (XCT) data of dendritic solidification using a convolutional neural network (CNN) that was trained only using synthetic images. The phase field method is used to rapidly generate these synthetic training images and their associated ground truths without human annotation. The CNN trained on phase field images segmented the experimental data with 99.3% accuracy, comparable to CNNs trained on human-generated ground truths. The number of synthetic images needed for CNN training will be discussed, and the most important microstructural features required for CNNs to “understand” the contents of an image are ranked.