A new machine learning workflow helps decode low-loss EELS data, turning noisy nanoscale spectra into spatial maps of the optical resonances that shape next-generation hybrid nanophotonic materials.
Paper: Highly efficient machine learning strategy for low-loss eels characterization: nanophotonic resonances as a case study. Image Credit: AI-generated image / OpenAI
In a recent research article published in the journal npj Computational Materials, researchers developed a highly efficient machine learning strategy that combines unsupervised and supervised algorithms to classify and interpret low-loss electron energy loss spectroscopy (EELS) data and to spatially map complex nanophotonic resonances in silicon/gold nanopillars.
Nanoscale Spectroscopy Challenges
The characterization of nanoscale materials often relies on probing their physical, chemical, and electronic properties with high spatial resolution. Electron energy-loss spectroscopy (EELS) combined with scanning transmission electron microscopy (STEM) is an invaluable technique in this regard, providing spatially resolved spectra that reflect material properties.
Particularly, the low-loss region of EELS spectra (below 50 eV) contains information on collective excitations, such as plasmons and Mie resonances, as well as inter- and intra-band electronic transitions and phonon excitations. These low-energy excitations are crucial for understanding nanophotonic effects in hybrid metal-semiconductor nanostructures.
However, analyzing low-loss EELS data is challenging due to several factors inherent to the nanoscale regime: overlapping spectral resonances, the intense zero-loss peak (ZLP) that masks nearby features, and low signal-to-noise ratio in small volumes.
Traditional analysis methods such as Principal Component Analysis (PCA) and Non-negative Matrix Factorization (NMF) assist with denoising and dimensionality reduction but often lack adaptive learning and predictive capabilities.
With recent advances generating large, complex datasets, implementing robust, automated machine learning (ML) strategies tailored to low-loss EELS is necessary to enable more accurate and efficient characterization of nanophotonic resonances and other excitations.
Schematic representation and SEM images of the fabrication process of the Si/Au nanopillars by colloidal lithography. (a) Array of polystyrene nanospheres homogeneously self assembled on the Au surface on top of the silicon wafer. (b) Au nanodiscs obtained by Ar RIE using the nanospheres as masks. (c) Obtained Si/Au nanopillars by deep Si RIE using the Au nanodiscs as masks. (d) Nanopillars after their release from the substrate by ultrasounds.
Machine Learning Workflow
The first step employs Uniform Manifold Approximation and Projection (UMAP), a nonlinear dimensionality reduction algorithm, to transform the high-dimensional spectral data into a lower-dimensional space, preserving the complex nonlinear relationships characteristic of low-loss EELS signals.
Subsequently, Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), an unsupervised clustering method, is applied to the reduced representations to identify clusters corresponding to different local spectral profiles.
Due to the presence of outliers or rare spectral patterns, a final supervised classification step using Support Vector Machines (SVM) is introduced to reclassify these ambiguous points, leveraging the clusters obtained by HDBSCAN as labeled training data. This hybrid unsupervised-supervised workflow enables near-real-time classification of large datasets and supports transferability to new EELS maps acquired under comparable experimental conditions without retraining from scratch.
Experimentally, Si/Au nanopillars were fabricated using colloidal lithography and reactive ion etching, producing p-doped Si nanopillars approximately 200 nm in diameter and 2 μm long, capped with 75 nm thick Au nanodiscs. STEM-EELS measurements were performed at 60 keV with high spatial resolution, yielding spectrum images that capture spatially localized nanophotonic resonances.
Complementary finite-difference time-domain (FDTD) simulations modeled the optical absorption of the nanopillars under plane wave illumination to correlate EELS resonances with electromagnetic modes, while recognizing that EELS can probe additional modes beyond those excited by plane waves.
Resonance Mapping Insights
Applying the combined UMAP-HDBSCAN-SVM analysis to high-resolution low-loss EELS data revealed distinct clusters that could be associated with spatially localized nanophotonic resonances within the Si/Au nanopillars. The method effectively separated vacuum/background and nanopillar regions from areas exhibiting characteristic plasmonic, hybrid, and dielectric resonances after ZLP exclusion.
Notably, the EELS spectra near 2.45 eV localized at the Au nanodisc were interpreted not as a single optical mode but rather as a cluster of closely spaced resonances, consistent with FDTD simulations and possibly influenced by known interband transitions in Au. This illustrates the ML strategy's ability to resolve complex nanoscale resonance clusters that are difficult to distinguish with conventional approaches.
A reproducibility test on an independent spectrum image from the edge of another nanopillar confirmed the robustness of the strategy, successfully identifying similar characteristic clusters and related nanophotonic modes.
Additional analysis of lower-magnification images, where spatial resolution and signal strength were reduced, demonstrated that, despite these limitations, the approach still detected and localized nanophotonic resonances, highlighting its sensitivity and adaptability to varying experimental conditions. However, the lower-resolution data could not resolve the finer resonance details observed in high-resolution measurements.
The supervised SVM stage proved instrumental in extending the unsupervised clustering by assigning outlier spectra within the dataset to appropriate classes, thus enhancing the completeness and accuracy of the classification. Its ability to generalize from one dataset to independent acquisitions under similar experimental conditions further suggests applicability for rapid on-the-fly EELS data interpretation.
The combination of high spatial-resolution STEM-EELS with this efficient ML framework opens new avenues for unraveling complex nanophotonic phenomena, including plasmon-Mie hybridization and dielectric mode coupling in hybrid nanostructures.
Schematic of Si/Au nanopillar fabrication. Ar RIE to create Au nanodiscs on the Si wafer using polystyrene beads as masks. Deep RIE (Bosch process) to form Si nanopillars using the Au discs as masks. Release of the Au/Si nanopillars from the Si wafer by ultrasonication and schematic of the dimensions. Image Credit: Adapted from Costa-Ledesma, V., del-Pozo-Bueno, D., Coll, C. et al. (2026). Highly efficient machine learning strategy for low-loss EELS characterization: nanophotonic resonances as a case study. npj Computational Materials. DOI: 10.1038/s41524-026-02171-1. Licensed under CC BY 4.0.
Advances and Applications
This study demonstrates a highly efficient machine learning strategy combining UMAP for dimensionality reduction, HDBSCAN for unsupervised clustering, and SVM for supervised refinement to analyze low-loss EELS spectrum images of nanoscale hybrid metal-semiconductor structures.
The methodology exhibits robustness across different datasets and experimental conditions, including lower-magnification measurements, and provides a transferable model for rapid classification of new EELS acquisitions when experimental conditions are comparable.
By enabling automated, near-real-time classification and supporting the interpretation of complex excitation modes at the nanoscale, this work paves the way for advanced characterization of nanostructured materials.
Download your PDF copy by clicking here.
Source:
- Costa-Ledesma V., del-Pozo-Bueno D., et al. (2026). Highly efficient machine learning strategy for low-loss EELS characterization: nanophotonic resonances as a case study. npj Computational Materials. DOI: 10.1038/s41524-026-02171-1, https://www.nature.com/articles/s41524-026-02171-1