🔬 optics

Extracting dipole orientations from asymmetric plasmonic nanostructures towards machine-learning-assisted spectropolarimetry

This paper presents a machine-learning-assisted spectropolarimetry method that extracts the azimuthal orientations of dipole moments in asymmetric plasmonic nanostructures by simultaneously fitting polarimetric dark-field spectra across multiple analyzer angles, achieving high accuracy that aligns with geometric simulations and electron microscopy data.

Original authors: P. Christian Simo, Michaela Zbytovska, Annika Mildner, Lukas Lang, Melanie Sommer, Dieter P. Kern, Monika Fleischer

Published 2026-09-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: P. Christian Simo, Michaela Zbytovska, Annika Mildner, Lukas Lang, Melanie Sommer, Dieter P. Kern, Monika Fleischer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Light carries more than just brightness; it carries a hidden direction. When light bounces off an object, its waves vibrate in specific patterns, a property known as polarization. For tiny objects smaller than the width of a human hair, this polarization holds a secret map of the object's shape and how it interacts with light. Scientists have long wanted to read this map to understand the orientation of tiny energy sources, called dipoles, inside nanoparticles. Knowing exactly how these dipoles point is crucial for designing better sensors, more efficient solar cells, and even for studying single molecules in biology. However, when these nanoparticles are not perfectly round or symmetrical, reading the map becomes incredibly difficult. The light they scatter gets tangled, and traditional methods often fail to tell the difference between a slightly tilted shape and a perfectly straight one, especially when the objects are too small to see clearly with standard microscopes.

A team of researchers at the University of Tübingen has developed a new way to solve this puzzle by combining careful light measurement with modern computer learning. Instead of trying to take a single, perfect picture of a nanoparticle, they bathed three different types of gold nanoparticles in light and watched how the scattered light changed as they rotated a filter in front of their detector. They tested particles that were slightly uneven, very irregular, and one that was deliberately engineered to have a strange, protruding shape. By measuring the light at twelve different angles of the filter, they gathered a rich set of data for each particle. They then used a mathematical model based on how light waves interfere with each other to guess the direction of the energy dipoles inside. To make sure their guesses were correct, they compared the results against detailed computer simulations of the particles' shapes and images taken with an electron microscope, which can see the tiniest surface details.

The researchers found that simply taking more measurements made their guesses significantly more accurate. When they used data from just a few filter angles, their estimates of the dipole direction varied widely. But as they added more angles to their analysis, the uncertainty shrank dramatically, eventually settling on a precise direction with an error margin of less than one degree. This level of precision allowed them to see that the energy dipoles inside the irregular particles did not always line up with the obvious long or short axes of the shape. In fact, for the most complex particle, which had a distinct bump on its side, the energy flow was tilted in a way that matched the location of that bump, revealing a hidden connection between the particle's physical asymmetry and its internal energy behavior.

To understand which specific measurements mattered most, the team trained a computer learning algorithm to analyze their data. This algorithm acted like a filter, learning which angles of the light filter were most important for figuring out the orientation of the dipoles. It discovered that the most critical measurements came from angles that aligned with the specific irregularities of the particle's shape. For example, if a particle had a protrusion on one side, the light measurements taken from that specific direction were the most valuable for determining the dipole's tilt. This suggests that the method can do more than just measure orientation; it can also reveal hidden details about a particle's shape that are too small to be seen directly. The study confirms that by looking at how light scatters from many different angles, scientists can map the invisible energy directions inside complex nanostructures without needing to destroy the sample or put it in a vacuum. This approach offers a powerful new tool for understanding the tiny, asymmetric building blocks of the future, turning a complex optical problem into a clear, measurable reality.

Technical Summary: Extracting Dipole Orientations from Asymmetric Plasmonic Nanostructures via Machine-Learning-Assisted Spectropolarimetry

Problem Statement
Determining the orientation of dipole moments in nanostructures is critical for understanding radiative properties such as emission polarization, directionality, and coupling. While polarization-resolved detection is a standard method, it faces systematic challenges when applied to asymmetric or chiral plasmonic particles. In these systems, geometric and electromagnetic asymmetries complicate the relationship between the particle's shape and its dipole radiation. Traditional approaches often rely on reconstructing only the dominant dipole moment, leading to significant angular errors for non-dominant or nearly degenerate modes, particularly in low-aspect-ratio or irregularly shaped particles. Furthermore, experimental biases introduced by excitation polarization, anisotropic interfaces, and near-field coupling can distort far-field polarimetry unless explicitly modeled. While deep learning has been applied to orientation recovery, it risks overfitting and unreliability if not grounded in robust physical models or sufficient data.

Methodology
The authors propose a comprehensive spectroscopic method combining polarimetric dark-field spectroscopy with an Analytical Dipole Model (ADM) and machine learning (ML) to extract electric dipole orientations in asymmetric plasmonic systems.

Experimental Setup: The study utilizes three types of asymmetric gold nanoparticles: low-asymmetry colloidal particles, high-asymmetry colloidal particles, and electron-beam lithography (EBL)-fabricated particles. These are placed on an indium tin oxide (ITO) coated glass substrate.
Spectropolarimetric Acquisition: The particles are characterized using unpolarized dark-field spectroscopy to identify resonance positions and dipolar intensities. Subsequently, polarimetric measurements are performed with a fixed incident linear polarization ( $\Phi = 45^\circ$ ) and a variable analyzer angle ( $\Theta$ ) ranging from $0^\circ$ to $165^\circ$ in $15^\circ$ increments (12 total angles).
Analytical Dipole Model (ADM): The collected spectra are fitted simultaneously using an ADM. This model emulates the electric far-field scattering intensity of dipole moments, accounting for the interference between multiple dipoles. The ADM treats the dipole orientations ( $\gamma_k$ ) as fitting parameters, alongside magnitudes and phase responses.
Statistical Permutation Analysis: To evaluate the precision and stability of the method, the authors analyze all non-repeating permutations of the 12 analyzer angles. Sets of $m$ angles (where $1 \le m \le 11$ ) are grouped, and the ADM is fitted to these subsets. This generates a statistical distribution of extracted orientations for each Localized Surface Plasmon (LSP) mode.
Validation and Simulation: The experimentally extracted orientations are compared against:
- Geometric symmetry axes ( $\gamma_{long}$ , $\gamma_{short}$ ) derived from Scanning Electron Microscope (SEM) images via contour analysis.
- Simulated dipole orientations ( $\gamma_{k,sim}$ ) calculated using Boundary Element Method (BEM) simulations based on 3D models reconstructed from SEM contours and grayscale height profiles.
Machine Learning Integration: A Histogram Gradient Boosting Regressor (HGBR) is trained on the filtered permutation data. The model evaluates the "importance" of specific analyzer angles ( $\Theta_j$ ) in predicting dipole orientations. This links the weights of the analyzer angles to the specific geometric asymmetries of the plasmonic systems.

Key Results

Orientation Extraction: The method successfully extracts the azimuthal orientations of dipole moments for particles with varying degrees of asymmetry, including systems with up to three distinct LSP modes (one of which is a mixed dipolar-quadrupolar mode).
Statistical Convergence: As the number of simultaneously fitted spectra ( $m$ ) increases, the standard deviation of the extracted orientations decreases significantly. For $m > 6$ , the standard deviation reaches sub-degree values (e.g., $< 2.5^\circ$ ), and the mean orientations converge toward the simulated values ( $\gamma_{k,sim}$ ) and geometric axes.
Accuracy: The extracted orientations show strong agreement with BEM simulations and SEM-derived geometric axes. For the EBL-fabricated particle, the method successfully identified a clockwise shift in the transversal mode orientation ( $\approx 75^\circ$ ) caused by a specific geometric protrusion, which was also reflected in the simulation.
Machine Learning Insights: The HGBR model revealed that the "importance" of specific analyzer angles correlates with the particle's geometric asymmetry. For instance, analyzer angles aligned with asymmetric protrusions or specific symmetry axes were found to be most critical for accurate prediction. This suggests that the spectral response to specific analyzer angles encodes hidden morphological features of sub-diffraction particles.

Key Contributions

Robust Extraction in Asymmetry: The study demonstrates that combining simultaneous multi-angle fitting with an ADM allows for the reliable extraction of dipole orientations in asymmetric and chiral nanostructures, overcoming limitations of single-dipole or dominant-mode assumptions.
Statistical Framework: By evaluating all permutations of analyzer angles, the authors establish a statistical framework that quantifies the precision of the method and identifies the minimum number of measurements required for reliable results (demonstrating stability with as few as three angles, with optimal precision at higher counts).
ML-Driven Feature Analysis: The integration of HGBR provides a novel way to interpret spectropolarimetric data, linking specific measurement parameters (analyzer angles) to physical geometric features (asymmetries) without requiring prior knowledge of the particle's shape.
Non-Invasive Characterization: The method offers a pathway to determine dipole orientations without the need for vacuum-based SEM observations, which can alter samples, by relying solely on optical spectroscopy.

Significance and Claims
The paper claims that this all-spectroscopic approach improves the accuracy of dipole orientation measurements in complex plasmonic systems. It asserts that the method is particularly valuable for:

Complex Systems: Determining orientations in chiral or asymmetric systems where traditional hardware-software methods may struggle.
Hidden Feature Detection: Using the sensitivity of analyzer angle importance values to reveal sub-diffraction morphological features (e.g., surface protrusions or asymmetries) that influence dipole radiation.
Broad Applicability: The technique is applicable to fields requiring orientation-specific spectroscopy where SEM is impractical, such as chiroptical and fluorescence spectroscopy.

The authors maintain a modest tone regarding 3D orientation, noting that while the current work focuses on in-plane (azimuthal) extraction, the ADM framework is theoretically capable of resolving out-of-plane tilt if the observation angle is varied. The study emphasizes that the robustness of the method relies on the simultaneous fitting of multiple spectra and the statistical validation provided by the large dataset of permutations.

More like this