🔬 physics

Mie Optical Computing

This paper proposes a compact neuromorphic optical computing architecture that encodes trainable transformations within the multipolar response of a single Mie scatterer, demonstrating that such a device can achieve classification performance comparable to a single-layer neural network while overcoming the parameter-density limitations of conventional diffractive processors.

Original authors: Mihail Petrov, Vsevolod Kleshchenko, Vladimir Igoshin, Costantino De Angelis

Published 2026-09-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mihail Petrov, Vsevolod Kleshchenko, Vladimir Igoshin, Costantino De Angelis

Original paper licensed under CC BY 4.0 (https://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

In the quest to make computers faster and more efficient, scientists have long looked to light as a potential replacement for electricity. Traditional computers process information using electrons moving through silicon chips, but light travels faster and can carry more data at once without generating as much heat. This field, known as optical computing, aims to build machines that process images and data using beams of light instead of electrical currents. For years, the leading designs for these light-based computers have relied on complex stacks of layers, similar to a multi-story building where light bounces between floors. Each layer acts as a filter or a mirror, tweaking the light slightly before it passes to the next. While effective, these systems are bulky, difficult to shrink down, and require a lot of space to perform even simple tasks. The fundamental challenge has been how to pack enough computing power into a tiny, single piece of material without needing a massive, sprawling structure.

A team of researchers from ITMO University in Russia and the University of Brescia in Italy has proposed a radical solution to this problem: instead of building a multi-layered machine, they suggest using a single, tiny particle to do all the work. In their study, they demonstrated that a solitary speck of material, roughly the size of a wavelength of light, can be trained to recognize patterns just as well as a much larger, more complex system. They focused on a specific type of interaction called Mie scattering, which occurs when light hits a particle that is comparable in size to the light's wavelength. When light strikes such a particle, it doesn't just bounce off; it excites a complex internal vibration within the particle, causing it to re-emit the light in a highly specific, patterned way. The researchers realized that if they could carefully control the internal structure of this particle, they could program it to perform mathematical calculations simply by how it scatters light.

The team tested this idea by asking the particle to perform a classic task: identifying handwritten numbers. They took images of digits from zero to nine, encoded them into the phase of a light beam, and focused that beam onto a single, virtual particle. The particle acted as a processor, transforming the incoming light based on its internal properties. On the other side, a set of detectors measured the intensity of the light scattered by the particle. The researchers then trained the particle's internal structure using a computer simulation, adjusting its properties until the pattern of scattered light correctly identified the number being shown. In these simulations, a single particle with a size parameter of 15 (a measure of its size relative to the light's wavelength) achieved a test accuracy of about 90 percent. This performance is comparable to a single-layer artificial neural network, a standard type of computer algorithm used for pattern recognition, proving that a single, compact object can indeed perform non-trivial computing tasks.

One of the most significant findings of the study is how much more information a single particle can handle compared to traditional layered systems. In a standard optical computer, the amount of information a layer can process is limited by the number of pixels it contains, which is roughly determined by the wavelength of light. However, a single particle can support a vast number of internal vibration modes, or "channels," that interact with each other. The researchers found that the number of adjustable settings, or parameters, inside a single particle grows much faster than the number of pixels in a flat layer as the particle gets larger. Specifically, the density of these trainable settings in a particle scales with the fourth power of its size, whereas in a flat layer, it scales with the square. This means that for a particle only slightly larger than the wavelength of light, the amount of computing power it can pack into its tiny volume is orders of magnitude higher than what a flat, layered design could achieve in the same space.

The study also explored the physical limits that nature imposes on such a device. Real-world materials must obey certain rules, such as reciprocity (the idea that light behaves the same way traveling forward or backward) and passivity (the rule that a passive object cannot create energy, it can only absorb or scatter what it receives). The researchers tested how these physical laws affected the particle's ability to compute. They found that while requiring the particle to be passive did reduce its maximum accuracy slightly, it also made the system more robust. A passive particle was less likely to be thrown off by small errors or noise in the light, acting as a natural stabilizer. Furthermore, they showed that even with these strict physical constraints, the particle could still be trained to recognize numbers with high accuracy, reaching up to 84 percent when designed as a non-absorbing dielectric object.

To prove that this concept was not just a mathematical trick but something that could exist in the real world, the team designed a specific physical shape for the particle. They created a virtual, inhomogeneous sphere, meaning a ball made of different materials mixed together in a precise pattern. By adjusting the density of the material at different points inside the sphere, they could tune its scattering properties. The simulations showed that this custom-designed sphere could successfully classify the handwritten digits. The optimal shapes were complex and varied, relying on coordinated interactions between many different internal resonances rather than a single simple effect. While such a finely detailed structure would be difficult to manufacture with current technology, the study serves as a proof of principle. It demonstrates that the complex calculations required for optical computing do not necessarily require a large, multi-layered device, but can be concentrated into the multipolar response of a single, compact scatterer.

This work opens a new path for designing optical computers that are incredibly small and efficient. By concentrating the entire computing process into a single object, researchers can potentially integrate these processors directly onto computer chips, bridging the gap between light-based processing and electronic circuits. The study suggests that the future of optical computing might not lie in building bigger, more complex stacks of mirrors and lenses, but in mastering the intricate physics of a single, tiny speck of matter. As the researchers continue to refine these designs, the possibility of building ultra-fast, low-energy optical processors that fit on the head of a pin moves from the realm of theory closer to reality.

Technical Summary: Mie Optical Computing

Problem Statement
Current optical neural networks (ONNs) and diffractive optical processors typically rely on spatially distributed architectures, such as cascaded passive or active diffractive layers. While these systems offer high-speed, low-latency processing, they are fundamentally limited by the physics of wave propagation. Specifically, increasing computational complexity in these distributed systems requires extending free-space propagation distances, adding more diffractive layers, or stacking multilayer metasurfaces. This leads to large physical footprints and imposes minimum size constraints on the system thickness for nonlocal operations. Furthermore, the density of trainable parameters in conventional diffractive layers is limited to approximately $1/\lambda^2$ per unit area. The paper addresses the challenge of implementing complex, trainable optical transformations within a single, compact optical element while maintaining high connectivity and parameter density.

Methodology
The authors propose a "Mie Optical Computing" architecture where the entire trainable transformation is implemented by a single compact scatterer operating in the Mie regime (particle size comparable to the wavelength). The core methodology involves:

Vector Spherical Harmonics (VSH) Basis: The system represents input and output optical signals as structured fields expanded in a basis of vector spherical multipoles. The input image (phase-encoded MNIST digits) is focused onto the scatterer, and the output is the scattered field.
T-Matrix Formalism: The optical processor is modeled as a trainable electromagnetic T-matrix. This matrix maps the incident multipole coefficients to the scattered multipole coefficients. Unlike distributed layers, this transformation occurs within a single finite object.
Parameter Density Scaling: The number of trainable parameters in the T-matrix scales as $n_{max}^4$ , where $n_{max}$ is the multipole truncation order. Since $n_{max} \propto ka$ (where $k$ is the wavenumber and $a$ is the particle radius), the total parameter count scales as $(ka)^4$ . However, when compared to the parameter density of conventional diffractive layers (scaling as $1/\lambda^2$ ), the proposed architecture provides a parameter density increase of the order of $(ka)^2$ .
Physical Constraints: The training process incorporates fundamental electromagnetic constraints:
- Reciprocity: Imposes symmetry relations on the T-matrix elements, reducing the number of independent parameters by roughly half.
- Passivity: Enforces energy conservation (scattered power $\le$ incident power) by constraining the spectral norm of the scattering matrix ( $\|S\|_2 \le 1$ ).
- Symmetry: Axial symmetry is imposed to reduce computational complexity, resulting in a block-diagonal T-matrix and a parameter scaling of $n_{max}^3$ .
Inverse Design: To demonstrate physical realizability, the authors perform inverse design on an inhomogeneous, non-absorbing dielectric particle. The particle is discretized into pixels with trainable relative permittivity, and the T-matrix is calculated using the Invariant Imbedding T-matrix Method (IITMM).

Key Results

Classification Performance: The trained T-matrix successfully classifies phase-encoded MNIST images using scattered-field intensity (in both far-field and near-field configurations). At a particle size parameter of $ka = 15$, the system achieves approximately 90% test accuracy, which is comparable to a single-layer artificial neural network (ANN).
Impact of Constraints:
- Reciprocity has a mild impact on accuracy, as the model retains sufficient degrees of freedom.
- Passivity causes a noticeable drop in accuracy compared to unconstrained models. This is attributed to the restriction on the operator spectrum; passive models cannot amplify specific modes to separate classes as aggressively as non-passive models. However, passivity significantly improves robustness against noise and perturbations.
- Axial Symmetry reduces the parameter count by an order of magnitude but maintains competitive accuracy scaling.
Physical Realizability: The inverse-designed dielectric scatterer achieves 84% test accuracy. The results show that high accuracy is achieved only in narrow resonant spectral ranges where the material distribution and size parameter jointly realize a favorable multipolar response. The optimized structures are highly inhomogeneous, relying on coordinated coupling among multiple resonant channels rather than single Mie resonances.
Near-Field Integration: The study demonstrates that similar performance can be achieved using near-field detection ( $kr \sim n_{max}$ ), suggesting potential for on-chip integration where the scattered field is collected directly by detectors or photonic circuits without extensive propagation distances.

Significance and Claims
The paper claims that this architecture demonstrates that nontrivial neuromorphic transformations can be encoded within the multipolar response of a single compact scatterer. The primary significance lies in the ability to concentrate a dense many-to-many modal transformation into a volume much smaller than conventional diffractive optical neural networks. By leveraging the $(ka)^2$ increase in trainable parameter density, this approach offers a pathway to compact optical computing units.

The authors emphasize that while the single scatterer reduces the volume of the trainable transformation itself, it does not eliminate the longitudinal space requirements for the entire system (e.g., the need to focus the input image onto the particle). However, the ability to perform computation within the near-field domain opens new avenues for integrating optical inference directly with integrated photonic circuits. The work establishes that physically constrained, passive, and reciprocal scatterers retain sufficient modal capacity to perform competitive classification tasks, bridging the gap between abstract operator-level optimization and physically realizable dielectric structures.

More like this