Adam E. A. Fouda Site-Selective Classification of Auger-Electron Spectra and Core-Electron Binding Energies; Chicago, IL; binding energy boosts robustness and accuracy.

Adam E. A. Fouda Site-Selective Classification of Auger-Electron Spectra and Core-Electron Binding Energies; Chicago, IL; binding energy boosts robustness and accuracy.

Site-Selective Functional Group Classification of Auger-Electron Spectra and Core-Electron Binding Energies with Convolutional Neural Networks

Site-Selective Functional Group Classification of Auger-Electron Spectra and Core-Electron Binding Energies with Convolutional Neural Networks Adam E. A. Fouda*1 Data Science Institute, The University of Chicago, Chicago, IL, 60637, United States (*Electronic mail: adamfouda@uchicago.edu) (Dated: 16 September 2026) X-ray spectroscopy techniques probe chemical states in systems with atom-site specificity; however, the site-selective characterization of local bond environments in novel research materials often relies on a combination of reference spectra, multiple complementary spectroscopic techniques and electronic structure calculations. Auger-electron spec- troscopy has long accompanied x-ray photoelectron spectroscopy as a second modality to resolve chemical states with overlapping core-electron binding energies. Here I show that the wealth of information encoded in the Auger spectrum offers the opportunity to train convolutional neural networks for site-selective functional group classification in organic molecules directly from the Auger spectrum lineshape. Furthermore, the inclusion of the core-electron binding energy as an additional modality improves the classification, either by augmenting the input fitted intensities with the binding energy or by conditioning the classification on the binding energy via feature-wise linear modulation layers. The latter approach was previously developed for visual reasoning in image classification, and the present results show that this is the more robust approach for including the binding energy. This work demonstrates the potential for new data-driven characterization capabilities in spectroscopic techniques that are information-rich but difficult to interpret. I. INTRODUCTION The local bonding environment around specific atom sites can significantly alter the activity of functional materials and molecules in applications such as catalysis1,2 and organic synthesis3,4. Spectroscopic techniques are used to character- ize chemical systems. Infrared (IR) and Raman spectroscopy excite the vibrational modes between atoms and are widely used for bond-type identification. However, IR and Raman are global probes of the system’s bonding environment, and higher-energy x-rays are required to interrogate the bonding environment around a specific atom site. X-rays excite inner- shell electrons tightly bound to an atom’s nucleus, and the binding energies of the electrons in these core orbitals are affected by the target atom’s local bond environment, an ef- fect known as the chemical shift.5 This phenomenon enables the photoionization of inner-shell electrons by x-ray photo- electron spectroscopy (XPS) measurements to resolve dif- ferent chemical states in materials,6 molecules,7 solutions,8 liquids,9,10 and biological matter,11 with atom-site selectivity. However, the influence of multiple competing mechanisms, such as charge transfer, electric fields and hybridization,5 which contribute to the structural environment effects on the core-electron binding energy, results in the presence of over- lapping peaks in the XP spectrum, which limits the resolution of XPS measurements. The site-selective characterization of bond environments in novel research materials therefore often relies on reliable reference spectra and on the combination of multiple complementary spectroscopic techniques. Following the removal of a core electron by x-ray pho- toionization, the core-hole decay process results in the au- toionization of secondary electrons from the x-ray-irradiated sample.12,13 The detection of these secondary electrons, which appear as Auger features in the XP spectrum, is com- monly used to distinguish oxidation states of active metal sites in catalysts and functional materials.14–22 This decay process, known as Auger-Meitner decay, involves an outer-shell elec- tron filling the core vacancy, and another outer-shell electron being ejected to the continuum. All energetically possible decay channels involving two outer-shell electrons contribute to the spectrum, and the number of possible final states thus scales non-linearly with respect to the system size. The multi- tude of decay channels explicitly involving transitions of elec- trons from valence bonds to the core-hole encodes a wealth of information into the Auger spectrum, making it a sensitive probe of the surrounding electronic structure. Auger-Electron Spectroscopy (AES) is therefore a powerful tool for charac- terizing local bond environments in novel chemical systems. The complexity of the Auger-Meitner decay process en- ables its enhanced sensitivity compared with XPS and other single-electron x-ray techniques. However, its complexity also challenges its experimental interpretation and computa- tional simulation, which limits the widespread adoption of this technique. The theoretical treatment of the large num- ber of final valence two-hole states (including the contribu- tion of multielectron shakeup and shakeoff processes) is non- trivial due to their multiconfigurational character.23–25 Fur- thermore, the accurate determination of the Auger-Meitner decay rate, and thus spectral intensity, requires the treatment of the ejected electron’s continuum wave function. Numerous approaches have been developed for its treatment, including approximating the spectral intensity by an electron population analysis,26–29 and both implicit30–42 and explicit43–50 consid- erations of the continuum electron wave function. Thus, full treatments of the Auger spectrum are usually limited to small molecules, and approximate treatments of the decay channels and rates are often employed for molecules with more than 20 atoms and for non-gas-phase systems. Whilst the analysis of x-ray spectroscopy measurements with electronic structure methods in “forward” (structure-to- spectrum) mapping is well established for techniques such as XPS,10,51–54 x-ray absorption spectroscopy (XAS),55–61 x- ray emission spectroscopy (XES)62–67 and resonant inelastic arXiv:2609.16386v1 [physics.chem-ph] 14 Sep 2026

2 x-ray scattering (RIXS),68–74 the recent rise in data-driven approaches utilizing machine learning (ML) has enabled rapid growth in developments using “backward” (spectrum- to-structure) mapping in XPS,75–79 XAS80–83 and XES84. The information rich and computationally complex nature of AES makes this technique particularly well suited for “backward” mapping. Prior to the recent rise in ML applications in the physical sciences, neural networks have been applied to the classification of single element Auger spectra85, discriminat- ing titanium nitride compositions in thin films,86 and back- ground removal for automatic analysis87. This work explores a novel data-driven approach for the site-specific identification of local bond environments in or- ganic molecules (organic functional groups) directly from the Auger spectrum lineshape with a convolutional neural net- work (CNN), a class of model commonly used in computer vision applications. CNNs extract local patterns from high- dimensional inputs via several layers of convolution filters to form a hierarchical composition of increasingly complex features. One-dimensional (1D) CNNs have demonstrated success in the analysis of spectral signals IR and Raman spectroscopy,88–90 and XAS82 For XPS, Drera et al. used a CNN to quantify elemental stoichiometry from XP spectra,76 whilst Pielsticker et al. used a CNN to determine the con- centration of different chemical phases in transition metal XP spectra.75 Both studies consider the effect of the Auger fea- tures in the XP lineshapes, and Pielsticker et al. found that correctly quantifying the different chemical phases in spectra with multiple elements required the consideration of Auger transitions. To the best of my knowledge no CNN has been previously applied to the analysis of molecular AES. In this work, I present a CNN model for classifying the organic functional groups directly from the carbon 1s Auger spectrum. The absence of large experimental databases in this domain necessitates model training on a dataset of calculated spectra. The spectra were calculated on a subset of molecules from the QM9 database using an electronic structure method which was previously benchmarked against experiment with modest accuracy.25 The level of theory and construction of the dataset are discussed in Section II A. Furthermore, this work shows that the classification performance is improved by including the carbon 1s binding energy (Eb) as an addi- tional input via two different approaches. The first approach is a simple augmentation of Eb to the fitted Auger spectrum intensities. The second approach conditions the classification with the Eb via the inclusion of feature-wise linear modulation (FiLM) layers, which apply a feature-wise affine transforma- tion between the convolution layers. The implementation of FiLM layers into the CNN architecture is discussed in Section II B. FiLM layers were developed to condition image classifi- cation with text inputs to achieve visual reasoning. I show that this mechanism can be translated to chemical applications and provides an new strategy for multi-modal chemical state iden- tification in automated spectroscopy analysis. II. COMPUTATIONAL DETAILS A. Dataset TABLE I. Carbon environment classes identified in the dataset with the SMARTS string pattern matching algorithm, with the merged class names in bold. A symbol representation of the nearest neigh- bor bonding defining the un-merged classes are given along side the counts of each class in the training and validation dataset and the hold-out and evaluation test sets. SM Table S1 contains the SMARTS strings91 used for the class definitions. Environment Bonding Train+val Hold-out Eval heteroaromatic 1979 72 6 arom N N-Car 1314 48 4 arom O O-Car 354 17 1 arom O N O-Car-N 311 7 1 aryl N 207 7 0 aryl amine Car-NH2 187 6 0 aryl nitro Car-NO2 20 1 0 aryl O 240 7 0 phenol Car-OH 187 5 0 aryl ether Car-O-R 53 2 0 aryl F 210 7 0 aryl fluoride Car-F 210 7 0 aryl carbonyl 313 13 0 aryl carbonyl Car=O 313 13 0 hydrocarbon 1603 32 31 methyl -CH3 140 6 10 methylene -CH2- 170 5 10 methine

CH- 65 2 0 quaternary C< 14 0 0 alkyne -C≡C- 774 4 2 vinyl C=C 71 4 2 aromatic Car (all-C ring) 369 11 7 carbonyl 558 19 3 ketone R-CO-R 148 3 1 aldehyde R-CHO 229 10 2 ester carbonyl -CO-OR 111 3 0 ester alkyl C-O-CO-R 65 3 0 carboxylic acid -COOH 3 0 0 carboxylate -COO− 2 0 0 amide carbonyl 167 3 1 amide carbonyl -CO-N 165 3 1 isocyanate O=C=N 2 0 0 nitrile 447 12 0 nitrile -C≡N 447 12 0 imine 104 4 0 imine C=N 104 4 0 oxyl 274 3 0 ether C-O-C 179 1 0 alcohol C-OH 95 2 0 amine 97 0 0 amine C-N 97 0 0 alkyl fluorinated 16 0 1 fluorinated C-F 16 0 1 Recently, we showed that carbon 1s Auger spectra in mod- est agreement with experiment can be produced efficiently

3 c) methane ethane propane butane cyclopropane cyclobutane ethylene acetylene tetrafluoromethane oxazole acetone formaldehyde formamide acetaldehyde benzene Evaluation molecules pyrimidine FIG. 1. a) Bar charts of the merged environment class counts in the training and validation sets (upper panel) and hold-out (lower panel blue) and evaluation (lower panel dark orange) test sets. b) Percentage of molecular sizes in the QM9 subset (training, validation and hold-out dataset) (blue) and evaluation dataset (dark orange). c) Structures of the evaluation dataset molecules used in this work. for a set of 20 organic molecules by excluding the contri- bution of shakeup and shakeoff processes, implicitly treating the continuum with pre-calculated atomic bound-continuum integrals via the one-centre approximation (OCA),36–42,92 and treating the bound-electron structure of the final dica- tion states with the multiconfiguration pair-density functional theory93–96 (MC-PDFT) method.