Session Y61: Deep Learning for Spectroscopy

Live

Tracking the structure of functional nanomaterials under operando conditions is a challenge due to the paucity of experimental techniques that can provide atomic-level information for active metal species. As a result, the search of activity descriptors relies almost exclusively on theoretical prediction and human expertise. X-ray absorption fine structure spectroscopy (XAFS) stands out as an element-specific method that is very sensitive to the local geometric, dynamic and electronic properties of the metal atoms and their surroundings and, is, therefore, able to track local structure modifications in operando conditions. Here we report on the use of X-ray absorption near edge structure (XANES) spectroscopy and supervised machine learning for investigating the information content “hidden” in the spectra. Using an autoencoder-based approach, we zoom in on the latent space for obtaining the number of key descriptors that affect the spectrum. Using a multilayer perceptron, we determine the values of key descriptors in metal nanocatalysts and size-selective clusters. We demonstrate that these methods are superior with respect to commonly used extended X-ray absorption fine structure (EXAFS) analysis of local structural properties in many cases, notably when EXAFS data quality is limited by low catalyst loadings and harsh reaction conditions in operando experiments. In the both cases, we train the artificial neural network on theoretical XANES simulations and use it to “invert” the experimental spectrum and obtain the corresponding structural properties. Several applications of these methods to the determination of catalytic descriptors in operando conditions, such as studies of synthesis, nucleation, growth and reactivity of metal catalysts will be demonstrated.   

Live

X-ray absorption fine structure (XAFS) is a premiere technique for the characterization of nanoscale systems. In many important cases, such as nanocatalysts with low metal loading, under high pressure and/or temperature, and under in situ/operando experimental conditions, XAFS spectral quality is limited to the region near the absorption edge. That region, also known as X-ray absorption near edge structure (XANES) has information about the local environment and electronic properties of the absorbing atom that, until recently, could not be reliably extracted by “inverting” the spectrum, only – by direct modeling using a limited number of candidate structures. In this work, we applied supervised machine learning and unsupervised machine learning approach to do the quantitative analysis of structural descriptors and explore what XANES features are embedded into a “bottleneck” representation. By combining principal component analysis with the autoencoder, we find that the latent variables are linearly separable, opening the door for their subsequent interpretation in terms of structural descriptors.   

Live

Low dimensional latent representations of data make it easier to extract hidden patterns and these representations can then be further used to build classifiers and other predictors. In this work, we will underscore the importance of applying probabilistic generative models to analyze X-ray absorption fine structure (XAFS) data and develop pathways for good latent representations. Inverting XAFS spectra to structural descriptors and mapping the evolution of changes in these descriptors during the in-situ experiment or under varying experimental conditions are of significant interest to catalysis community. By using deep learning and variational inference, we show that Variational Autoencoders (VAE) can help disentangle non-linear interactions between underlying explanatory factors. Furthermore, we would show that low dimensional latent representations can also be utilized to invert the experimentally obtained XAFS spectra to structural and electronic properties of the catalysts in Pd nanoparticles under hydrogen atmosphere at elevated temperatures.   

Live

X-ray Absorption Near Edge Spectroscopy (XANES) is frequently used to unravel the local electronic structure of atoms and studying oxidation state changes. The local structure of atomic impurities such as Cu or As in CdTe-based solar cell materials, for example, can be probed using XANES. In this work, we used first principles computations to generate Cu and As K-edge XANES data for point defects and defect complexes in bulk and grain boundary structures of CdTe, as well as various relevant compounds of the impurity atoms, with the idea of capturing the structural diversity likely to be found in the solar cell material. Using a massive computational dataset derived in this fashion, we developed a machine learning (ML) framework for accurately predicting the coordination number (CN) around the central Cu or As atom by applying regression techniques ranging from Gaussian processes (GP) to random forests (RF) to neural networks (NN). NN and GP models can predict the CN with a root mean square error of < 0.05 for a dataset with a CN range of 3 to 12. We further studied the effect of noise in the XANES data on the ML models, and used the final, optimized models to predict the coordination environment in dozens of samples with measured XANES spectra.   

Live

Raman spectroscopy is commonly used for characterization of graphene. The correlations between Raman peak attributes (e.g. peak width, position, asymmetry) indicate important information about physical properties of graphene, such as layer thickness, defects, doping levels, dielectric screening, and strain. Here we use machine learning techniques to find and study the relevant correlation of Raman spectral parameters, and reveal the properties of graphene in different dielectric environments. Graphene has shown different doping levels and screening effects on different substrates. We found that the 2D peak asymmetry reveals the charge doping level with a strong correlation. We analyze Raman spatial mapping data of graphene in different dielectric environments to find an efficient way to evaluate the phonon properties of graphene with Raman spectra data, which is an important reference for fabricating graphene-based devices. This method helps us understand correlations in graphene in different screening environments.   

Live

Neural networks are computational systems that rely on a series of weighted algorithms to processes input data and give an output. A common type of neural network used for image processing is a convolutional neural network (CNN). Due to their effectiveness at image classification, CNN’s have great potential to be useful in analysis of reflection high energy electron diffraction (RHEED) patterns of complex oxides. This potential is realized by creating a CNN that takes RHEED images as input and outputs a predicted x-ray phtotoelectron spectroscopy (XPS) spectrum of the material. Neural network performance depends on the weight values of the network, which are found by training the neural network. A problem that arises when training such a CNN is the limited availability of consistent XPS spectra to compare to the output of the neural network when training. This problem is overcome by using BriXias software to simulate a wide variety of XPS spectra. BriXias software utilizes a database of material characteristics to evaluate the inelastic mean free path (IMFP) and transport mean free path (TMFP) of electrons traveling within a material. It then uses the IMFP and TMFP, along with specified model parameters and XPS geometry, to simulate XPS data of material.   

Live

Topological insulators (TIs) are bulk semiconductors that manifest spin-helical conducting surface states due to the topological bulk-boundary correspondence principle. Control and deeper understanding of the surface environment is thus of vital importance for long-term technological goals. In this talk, I will describe a data acquisition and analysis framework termed sparse big data (SBD) spectroscopy, in which one rapidly maps a sample surface to resolve the variation of electronic structure as a function of local environment. Leveraging the large data set enables one to realize the analog of a multi-dimensional doping series. I will present examples demonstrating that the fine ‘control’ this entails can be used to unlock new experimental observables such as a hybridization-like interplay between the surface and bulk electronic states in a TI. I will also touch on likely avenues for future development involving convolutional neural networks and new data visualization strategies.   

Live

X-ray spectroscopy methods probe a synthesized material to provide detailed information about its structure at multiple levels of granularity. However, these methods are expensive and in some cases need to be performed at a synchrotron motivating rapid analysis of the data coming from these methods.

X-ray diffraction is critical to examine crystal structure of the material synthesized and also acts as a link between experiment and theory. In recent years, high-throughput measurement of x-ray diffraction has given us large volumes of data that are challenging to analyze. We show that by using physical constraints, matrix factorization, and introducing known knowledge we can rapidly map out quternary phase diagrams from experimental data.

X-ray absorption spectra are criticial to capture local chemical information of a material which are of great importance for functional properties such as catalysis, photocatalysis, battery electrodes etc. However, intepretation of these spectra is very time consuming and automated interpretation is likely to help maximize the use of expensive synchrotron resources. By combining random-forest and physically meaningful featurizations we show that we can automatically capture coordination number, bader charge, and nearest neighbor distances.   

Live

The determination of atomic-scale properties such local environment and spin states of functional materials is of great importance to the materials physics community, yet the difficulty in extracting these properties from characterization data such as X-ray spectroscopy poses challenges to effective data analysis. We will discuss how machine learning models are used to extract those properties from X-ray spectra. Examples include the use of random forest models for local environment prediction from X-ray absorption spectroscopy and extracting the electronic structure change of a representative Ni-Co-Mn-based cathode material through X-ray emission spectroscopy. These findings indicate that the combination of computational spectroscopy and machine learning techniques will be an invaluable resource by greatly enhancing the efficiency at which experimental X-ray spectra can be analyzed.   

Live

Simulations of excited state properties, such as spectral functions, are often computationally expensive and therefore not suitable for high-throughput modeling. As a proof of principle, we demonstrate that graph-based neural networks can be used to predict the x-ray absorption near-edge structure spectra of molecules to quantitative accuracy. Specifically, the predicted spectra reproduce nearly all prominent peaks in O and N K-edge spectra, with 90% of the predicted peak locations within 1 eV of the ground truth. Besides its own utility in spectral analysis and structure inference, our method can be combined with structure search algorithms to enable high-throughput spectrum sampling of the vast material configuration space, which opens up new pathways to material design and discovery.   

Live

First-principles prediction of nuclear magnetic resonance chemical shifts plays an increasingly important role in the interpretation of experimental spectra, but the required density functional theory (DFT) calculations can be computationally expensive. Promising machine learning (ML) models for predicting chemical shieldings in general organic molecules have been developed previously, though the accuracy of those models remains below that of DFT. The present study demonstrates how much higher accuracy chemical shieldings can be obtained via a Δ-ML approach. Specifically, an ensemble of neural networks (NN) is trained to correct PBE0/6-31G chemical shieldings up to PBE0/6-311+G(2d,p) which can predict ¹H, ¹³C, ¹⁵N, and ¹⁷O chemical shieldings with root-mean-square errors of 0.12, 0.79, 1.82, and 2.66 ppm. Furthermore, the ability to estimate the uncertainty in the predicted shieldings based on variations within the ensemble of NN models is also assessed. Finally, it is also demonstrated that the ML model predicts experimental solution-phase NMR chemical shifts in drug molecules with only modestly worse accuracy than the target DFT model.