Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session S32: Deep Learning SpectroscopyFocus Recordings Available
|
Hide Abstracts |
Sponsoring Units: GDS Chair: William Ratcliff, GDS Room: McCormick Place W-192B |
Thursday, March 17, 2022 8:00AM - 8:36AM |
S32.00001: Prediction of materials properties from core-loss spectrum using neural network Invited Speaker: Teruyasu Mizoguchi Data driven approaches are now indispensable for modern materials characterization due to rapid increase of size and dimension of data observed in experiments and simulations. Based on this backgrounds, we are developing data-driven methods for the materials characterizations. |
Thursday, March 17, 2022 8:36AM - 9:12AM |
S32.00002: Review and Prospect: Deep Learning in Nuclear Magnetic Resonance Spectroscopy Invited Speaker: Xiaobo Qu
|
Thursday, March 17, 2022 9:12AM - 9:24AM |
S32.00003: Investigation of featurization approaches for supervised machine learning in X-ray spectroscopy Yiming Chen, Chi Chen, Chengjun Sun, Steve M Heald, Shyue Ping Ong, Maria K Chan Machine learning (ML) has been revolutionizing the analysis of materials characterization data. In X-ray absorption spectroscopy (XAS) and X-ray emission spectroscopy (XES), for instance, ML is used to extract critical structural information and accelerate the interpretation of spectra. While most ML models have utilized the raw spectra as the input, less emphasis has been placed on the preprocessing of spectra to derive alternative inputs that can potentially enhance model performance. In this talk, we will benchmark reduced dimension features, e.g., peaks and principal components, and overcomplete representations to identify the optimal representation of spectroscopy data for supervised machine learning approaches. The system of interest is a cathode material for Li-ion battery, LiNixMnyCozO2 (NMC), which is well-known for its high energy densit and long-term cyclability. The performance of these input transformations will be assessed for both XAS, which contains key information about oxidation states and local environment, as well as XES, which provides additional information about electronic states. We will demonstrate that such featurization can significantly enhance not only the prediction accuracy, but also the interpretability of ML models. |
Thursday, March 17, 2022 9:24AM - 9:36AM |
S32.00004: Combining machine learning and XANES spectra featurization to make chemical environment predictions of CdTe materials Justin Pothoof, Arun Kumar Mannodi Kanakkithodi, Srisuda Rojsatien, Xinyue Wang, Amy Stegmann, Yu-Hsuan Hsiao, Mariana Bertoni, Maria K Chan X-ray absorption spectroscopy is a popular method for unraveling the local atomistic and electronic structures of materials. This approach can be used for studying the bonding environment of various dopants in CdTe photovoltaics, which is important for understanding and improving their performance. In this work, we exploit the relationships between x-ray absorption near edge structure (XANES) features and local atomic structures to develop a machine learning (ML) framework for accurately predicting the coordination numbers of dopants in CdTe. Using FEFF9 simulations, we generated a large dataset of chemical environments that sample dozens of compounds focusing on Cu and As dopants, including bulk phases of Te- and Se-based structures, As and Cu defects in CdTe, etc. Simulated XANES are mapped to the coordination environment using random forest regression, neural network, and gaussian process regression models. The models are trained on the spectra in addition to other features such as its first and second derivative. The best performing ML models are deployed for predictions on measured XANES from a set of experimental samples, showing the utility and wide applicability of this approach. |
Thursday, March 17, 2022 9:36AM - 9:48AM |
S32.00005: Deep-learning-enabled optical ellipsometry for complex thin films and 2D materials ziyang wang, Yuxuan Lin, Shengxi Huang, Kunyan Zhang, Wenjing Wu Analysis of optical spectroscopy data often requires intensive model fitting. Reflectometry and ellipsometry are commonly used methods to measure the optical dielectric functions or the complex refractive indices of optical thin films such as 2D materials. However, the substrate structure needs to be simple (a single, thick, and transparent substrate is ideal) and perfectly defined. In addition, the available fitting models are extremely computational expensive, and very specific to the optical structures of the samples. In this study, we develop a deep learning method based on a encoder-decoder convolutional neural network that is capable of extracting refractive indices of thin-film materials (including 2D materials) on arbitrary complex multilayer substrates. Kramers-Kronig relations are incorporated into the model to reduce the dimensions of the training parameters. The model is trained using numerically generated data. Without any prior knowledge of stacked material structures, our model can predict the complex refractive indices of 2D materials from experimentally obtained optical reflectance data with high accuracies. This approach enables the in-situ optical characterization of functional materials and components in actual complex optoelectronic devices, a task previously not feasible with traditional reflectometry or ellipsometry methods. |
Thursday, March 17, 2022 9:48AM - 10:00AM |
S32.00006: Machine Learning-Accelerated Spectral Imaging Analysis for Nanomaterials Haili Jia, Canhui Wang, Chao Wang, Paulette Clancy Scanning transmission electron microscopy (STEM)-based electron energy loss spectroscopy (EELS) has been developed recently to probe the atomic-scale structure of nanostructures. Despite its potential, one critical challenge using STEM-EELS is the difficulty of extracting information from a large spectral image dataset with convoluted spatial, energy and spectroscopic data. The complexity arises largely from limited energy resolution and overlapping spatial and spectroscopic information from different materials and physical/chemical states, especially in a reactive environment. To address this challenge, we developed an "on the fly" machine learning-enabled real-time spectral imaging analysis system based on non-negative robust principal analysis and compressed sensing. This offers a new way to study structure-property relationships of complex nanomaterials to facilitate materials design that are otherwise challenging to obtain through conventional experimental or simulation approaches. Test cases will show it provides a viable path to conduct spectroscopic studies of structure and chemical and electronic properties in a reactive environment and is general enough to be deployed for other spectral imaging tools in material research. |
Thursday, March 17, 2022 10:00AM - 10:12AM |
S32.00007: Elucidating proximity magnetism through polarized neutron reflectometry and machine learning Nina Andrejevic, Zhantao Chen, Thanh Nguyen, Mingda Li Polarized neutron reflectometry (PNR) facilitates structural characterization of multilayered materials with depth sensitivity, enabling the study of interfacial phenomena such as the magnetic proximity effect, a promising pathway for magnetizing topological insulators (TIs) and advancing TI-based device applications. However, PNR profiles often inhabit a complicated objective function landscape using traditional fitting methods, posing a significant challenge to parameter retrieval. Here, we develop an alternate, data-driven framework to retrieve the parameters of candidate proximity-coupled systems from their PNR profiles with minimal user intervention. Using a variational autoencoder, we map PNR profiles to a low-dimensional latent space from which the true sample parameters can be readily obtained. The decoded profiles directly inform the suitability of the parameter space through the reconstruction quality and are robust to moderate perturbation of the inputs. Importantly, we find that the latent mapping naturally bypasses the issue of multiple local minima, and is both well-organized and visually interpretable in terms of physical parameters. We evaluate our model by recovering the sample parameters from experimental PNR profiles of two candidate proximity-coupled systems. |
Thursday, March 17, 2022 10:12AM - 10:24AM |
S32.00008: Predicting X-Ray Absorption Spectra of Materials Using Graph-based Neural Networks Fanchen Meng, Matthew R Carbone, Deyu Lu X-ray absorption spectroscopy (XAS) is an important characterization technique in condensed matter physics, materials science and chemistry for resolving local structural and electronic properties. First-principles simulations of XAS have played a critical role in interpreting the otherwise abstract spectral features and obtaining physical insight. However, the widely used simulation methods are computationally expensive, especially when computing the near edge region. Machine learning (ML) models trained from first-principles simulation data have been widely employed on a variety of problems to speed up inference while maintaining high fidelity. However, to the best of our knowledge, ML has not been applied to predicting the XAS for periodic materials. In this work, we use the graph-based neural networks to predict the X-ray absorption near-edge structure (XANES) spectra of a wide range of materials from the Materials Project. Our models are accurate and highly efficient, and therefore can enable high-throughput spectrum sampling in the large materials phase space, which could allow for efficient structure refinement and may offer new routes for real time spectral interpretation in autonomous experimentation. |
Thursday, March 17, 2022 10:24AM - 10:36AM |
S32.00009: Identifying charge density and dielectric environment of graphene using Raman spectroscopy and deep learning Zhuofa Chen, Yousif Khaireddin, Anna K Swan Raman spectroscopy is used to evaluate graphene and its interactions with its surroundings such as strain, charge density, and dielectric environment, reflected in peak positions, shape and intensity. In addition, substrate interference effects and experimental alignment affects collected spectroscopic data causing variation even for similar environments. Such variations, artifacts, and environmental differences pose a challenge in accurate spectral analysis. In this work, we developed a deep learning model to overcome the effects of such variations and classify graphene Raman spectra according to different charge densities and slightly varying dielectric environments. We demonstrated the spectra classification with 99% accuracy using the proposed CNN model. This CNN model is able to classify Raman spectra of graphene with different charge doping levels (< 2X1012cm-2) and even subtle variation in the spectra between graphene on SiO2 and graphene on silanized SiO2. Our proposed model shows high reproducibility and stability. Our approach has the potential for fast and reliable estimation of graphene doping levels and dielectric environments. The proposed model paves the way for achieving efficient analytical tools to evaluate the properties of graphene. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700