Abstract
Our visual perception of our surroundings is ultimately limited by the diffraction limit, which stipulates that optical information smaller than roughly half the illumination wavelength is not retrievable. Over the past decades, many breakthroughs have led to unprecedented imaging capabilities beyond the diffraction-limit, with applications in biology and nanotechnology. In this context, nano-photonics has revolutionized the field of optics in recent years by enabling the manipulation of light-matter interaction with subwavelength structures. However, despite the many advances in this field, its impact and penetration in our daily life has been hindered by a convoluted and iterative process, cycling through modeling, nanofabrication and nano-characterization. The fundamental reason is the fact that not only the prediction of the optical response is very time consuming and requires solving Maxwell's equations with dedicated numerical packages. But, more significantly, the inverse problem, i.e. designing a nanostructure with an on-demand optical response, is currently a prohibitive task even with the most advanced numerical tools due to the high non-linearity of the problem. Here, we harness the power of Deep Learning, a new path in modern machine learning, and show its ability to predict the geometry of nanostructures based solely on their far-field response. This approach also addresses in a direct way the currently inaccessible inverse problem breaking the ground for on-demand design of optical response with applications such as sensing, imaging and also for plasmon's mediated cancer thermotherapy.
Introduction: The Design Bottleneck in Nano-photonics
Nano-photonics allows for the manipulation of light using structures smaller than the wavelength of light itself. This capability is essential for advanced sensing, imaging, and therapeutic applications. However, the design process is traditionally hindered by two primary computational challenges.
The Direct Problem involves predicting the optical response of a known geometry by solving Maxwell’s equations. This process is computationally intensive and slow, even when using high-end numerical solvers like Finite Element Method (FEM) or Finite Difference Time Domain (FDTD).
The Inverse Problem is significantly more difficult: determining the specific geometry required to produce a desired optical response. Because the relationship between geometry and light-matter interaction is highly non-linear, this "holy grail" of design typically requires a slow, iterative cycle of modeling, fabrication, and characterization. While evolutionary algorithms offer some shortcuts, they must re-search the entire parameter space for every new design query, making them impractical for real-time applications.
Methods: A Bi-directional Deep Learning Approach
Researchers at Tel Aviv University have developed a novel bi-directional deep neural network (DNN) architecture to overcome these hurdles. The system consists of two interconnected paths:
- Geometry-Predicting-Network (GPN): This "inverse" path takes two transmission spectra (horizontal and vertical polarization) and material properties as input to predict the nanostructure's geometry.
- Spectrum-Predicting-Network (SPN): This "direct" path takes the geometry predicted by the GPN and predicts the resulting spectra.
A key innovation is co-adaptation: the networks are trained together, with the GPN's output feeding directly into the SPN. This joint training allows the networks to learn the underlying physics more effectively, significantly reducing error compared to training them in isolation. Furthermore, the SPN utilizes shared weights for both polarizations, mimicking the physical coupling of free electrons in 2D nanostructures.
The model was trained on a synthetic dataset of 15,000 experiments generated via COMSOL Multiphysics, focusing on a versatile "H-shaped" geometry family where leg lengths and angles are varied.
Results and Discussion
The bi-directional approach achieved a Mean Squared Error (MSE) of 0.16, a substantial improvement over the 0.37 MSE observed when training separate networks. Beyond simulation, the model was validated against physical gold nanostructures fabricated on Indium Tin Oxide (ITO) glass. Scanning Electron Microscopy (SEM) confirmed that the DNN successfully retrieved physical dimensions with high accuracy.
One of the most significant advantages of this method is speed. While the initial training phase is computationally expensive, the subsequent "query phase"—actually designing a structure—is quasi-instantaneous, taking only seconds. This was demonstrated by designing a complex gold plasmonic sensor for Dichloromethane, which required specific, different resonances across orthogonal polarizations.
Architecture Optimization
The researchers also explored the impact of network depth on performance. They found a "sweet spot" at six "join" layers; fewer layers lacked the necessary complexity to model the physics, while additional layers led to vanishing gradients that hindered effective learning.
Conclusion
By shifting from iterative numerical simulations to a physics-informed deep learning framework, this work provides a path toward on-demand design of nano-photonic structures. This approach not only solves the inverse problem for complex geometries but also enables real-time characterization and design for applications ranging from bio-sensing to cancer thermotherapy. Future work will look to expand the geometry families and include more variable physical parameters to further generalize the model's utility.
Summary updated