The work addresses a central challenge in the study of coupled resonant systems. When two or more optical cavities interact, their resonances can shift, broaden, split, or interfere with one another. These changes produce complicated transmission spectra that contain information about the system’s intrinsic resonant frequencies, losses, coupling strengths, and transmission phases. In principle, researchers can recover these parameters by fitting measured spectra to theoretical formulas. In practice, however, the process can be slow, computationally demanding, and vulnerable to multiple solutions.

The problem arises because different combinations of physical parameters can generate spectral responses that look remarkably similar. A fitting algorithm may therefore identify several plausible answers, making it difficult to determine which set of parameters accurately represents the real optical system. This ambiguity is particularly serious in advanced resonant platforms associated with exceptional points, topological photonics, bound states in the continuum, high-sensitivity sensing, optical signal processing, and low-threshold lasers.

To overcome these limitations, a team led by Da-Quan Yang of Beijing University of Posts and Telecommunications, Yi Xu of Guangdong University of Technology, and Hua-Shun Wen of Nankai University designed a neural network trained with data generated by coupled mode theory. Rather than relying only on large collections of experimental measurements, the system learns from low-cost theoretical datasets produced by a physical model. This gives the network access to many possible combinations of resonant-system parameters without requiring researchers to measure every case in the laboratory.

The network also incorporates physical constraints based on the eigenvalues of the coupled system directly into its loss function. In a conventional deep-learning model, the loss function measures how far a prediction is from a target answer. Here, the loss function additionally penalizes predictions that violate the physical behavior expected from the resonator. This physics-guided design helps the network distinguish between mathematically possible solutions and solutions that are consistent with the underlying dynamics of coupled optical modes.

In simulations and experiments, the method retrieved four important parameters in complex coupled resonant systems: intrinsic resonant frequency, intrinsic loss, coupling strength, and transmission phase. According to the researchers, the average computation time was reduced by three orders of magnitude compared with traditional fitting methods, while predicted performance improved by more than two orders of magnitude. Relative to a conventional data-driven neural network without the same physical constraints, the prediction error for physical parameters fell by as much as 81.59 percent.

The result is significant because it shows how machine learning can be used not simply as a fast pattern-recognition tool, but as a way to preserve and exploit the physical structure of a problem. In optical systems, where a small change in geometry, material properties, or cavity spacing can produce a large and highly nonlinear spectral response, this combination may be especially valuable. A model that understands both the measured signal and the equations behind it can be more robust when data are limited, noisy, or affected by multiple interacting resonances.

The researchers next tested the framework in a displacement-sensing experiment involving two coupled microcavities. As the displacement changed, it modified the system’s underlying physical parameters, producing spectral signatures that included resonance shifts, linewidth broadening, and mode splitting. The researchers fed the measured transmission spectra into the pre-trained network, then used the predicted parameter changes to infer displacement. The system maintained strong performance across different displacement conditions, with reported coefficients of determination, or R² values, above 0.979.

This approach could have applications beyond the particular microcavity platform tested in the study. Fast retrieval of hidden physical parameters could help researchers characterize integrated photonic circuits, tune coupled-resonator devices, monitor microscopic mechanical motion, and develop intelligent sensors for nanoscale particles. The framework may also support non-invasive morphological measurements of on-chip optical systems, where extracting information from spectral changes without physically disturbing the device is highly desirable.

The researchers say the method’s broader potential lies in its scalability. By combining physics-based simulation, experimental data, and neural-network inference, the framework could be adapted to increasingly complicated systems containing multiple cavities or several sensing modalities. Future integration with optoelectronic chips and multimodal sensing technologies may create compact platforms capable of analyzing several physical properties at once. For a field in which interpreting every spectral feature can be computationally expensive, a neural network that learns the language of coupled light could turn complex resonances into rapid, actionable measurements.

Subject of Research: Physics-data co-driven deep neural networks for analyzing coupled optical resonant systems and nanoscale displacement sensing

Article Title: Deciphering optical coupled resonant systems with physics-data co-driven deep neural networks

Web References: https://doi.org/10.1038/s41377-026-02389-0

References: Light: Science & Applications, DOI: 10.1038/s41377-026-02389-0

Image Credits: Hua-Shun Wen et al.

Keywords

Optical resonators, coupled mode theory, deep neural networks, physics-informed machine learning, photonics, microcavities, resonant sensing, nanoscale displacement detection, coupled resonant systems, optical spectroscopy

Machine learning has emerged as a crucial driver for the rapid advancement of nanophotonics, facilitating inverse design strategies that transcend traditional trial-and-error methods. By harnessing data-driven techniques, researchers can optimize complex optical devices with unprecedented speed and accuracy. This capability is vital for engineering metasurfaces and integrated photonic circuits that exhibit tailored electromagnetic responses, enabling functionalities such as all-optical signal processing, adaptive imaging, and robust environmental sensing. The interplay of computational intelligence and nanostructure engineering creates a landscape where photonic devices evolve from passive components to intelligent systems capable of dynamic reconfiguration.

One of the most compelling arenas within intelligent photonics is optical neural networks, which exploit photons as information carriers to perform neural computations intrinsically. Unlike their electronic counterparts, optical neural networks benefit from parallelism and ultrafast propagation speeds while consuming markedly less energy. These networks operate through diffractive layers, metasurfaces, or integrated waveguide arrays engineered to mimic the function of artificial neurons and synapses. Despite significant progress, realizing scalable and versatile optical computing platforms necessitates breakthroughs in integrating these networks with existing digital architectures, addressing challenges in programmability, noise resilience, and fabrication tolerances.

The integration of sensing and computing on a single photonic platform marks another paradigm shift enabled by intelligent photonics. Metasurface-based neural networks harness multiple degrees of freedom in light—including phase, polarization, and orbital angular momentum—to capture rich environmental information. This multiplexing capability allows for simultaneous data acquisition and pre-processing, drastically reducing latency and bandwidth requirements in edge devices. Applications extend to advanced imaging systems capable of real-time scene understanding, bio-optical sensors with enhanced specificity, and telecommunication networks that dynamically adapt to varying channel conditions. Such tightly coupled sensing-computing systems anticipate a future where smart optoelectronics permeate everyday technology.

Despite the promise, intelligent photonics faces formidable theoretical and practical obstacles. The complexity of accurately modeling light–matter interactions at the nanoscale, combined with the intricacies of deep learning optimization, demands sophisticated algorithms and simulation tools. Fabrication challenges arise from the nanometric precision required for metasurfaces and photonic circuits, where slight deviations can degrade performance. Furthermore, operational stability under varying environmental conditions remains a crucial hurdle. Addressing these issues mandates collaborative efforts merging expertise from machine learning, materials science, optics engineering, and manufacturing disciplines.

A research team at the Harbin Institute of Technology, Shenzhen, under the guidance of Professors Jingtian Hu, Shumin Xiao, and Qinghai Song, provides a seminal review articulating the current state and future directions of intelligent photonics. Their comprehensive analysis illuminates how machine learning techniques catalyze advancements across computing, sensing, and dynamic photonic devices. By dissecting the interplay of algorithmic strategies and hardware innovations, the review charts a roadmap for overcoming existing limitations and accelerating practical deployment.

Central to this discourse is the vision of large-scale optical networks as a cornerstone for next-generation, energy-efficient computing. As artificial neural networks grow in complexity and size, traditional electronic infrastructures struggle to meet the rising demands of speed and power consumption. Optical computing platforms, with their inherent parallelism and low dissipation, represent a compelling alternative. The authors stress that realizing this vision depends on co-designing hardware and algorithms, formulating hybrid systems capable of seamless integration with digital environments, and optimizing architectures for real-world tasks.

Beyond centralized computing centers, intelligent photonics fosters the emergence of edge computing devices endowed with integrated sensing and processing capabilities. Such devices exploit the multifaceted nature of light to capture complex signals and extract relevant features through embedded optical neural networks. This approach minimizes data transmission requirements by conducting initial data analysis locally, enhancing privacy and reducing latency. In domains ranging from autonomous vehicles to wearable health monitors, the seamless fusion of sensing and computation facilitated by intelligent photonics drives novel functionalities and improved user experiences.

The implications of intelligent photonics extend well beyond individual devices or applications. The researchers envision profound impacts on the broader technological landscape, including the metaverse, augmented and virtual reality environments, and expansive Internet of Things ecosystems. These emerging platforms demand compact, multifunctional hardware capable of real-time processing and adaptability — criteria that intelligent nanophotonic systems are uniquely positioned to meet. By enabling versatile, miniaturized components with unprecedented operational flexibility, this field stands to accelerate the evolution of immersive and interconnected digital worlds.

Underlying these advancements is a commitment to sustainability and efficiency, as the AI industry grapples with escalating energy costs associated with large-scale machine learning models. Intelligent photonics offers a path toward greener AI by harnessing photonic circuits’ ultralow energy consumption and high throughput. This transition not only addresses environmental concerns but also unlocks new possibilities for deploying AI technologies in resource-constrained environments, facilitating broader accessibility and impact.

The path forward for intelligent photonics is defined by interdisciplinary collaboration and innovation. Bridging gaps between theory, fabrication, and application, researchers must develop robust frameworks that account for material nonlinearities, fabrication imperfections, and practical integration challenges. Simultaneously, advancing machine learning methodologies tailored for photonic systems will amplify design capabilities and operational robustness. Collectively, these efforts will transform intelligent photonics from a nascent concept into a foundational pillar of future information technology.

This review, published in the journal eLight, encapsulates the exciting frontier of intelligent nanophotonics and machine learning convergence. It shines a spotlight on the promising advances, substantial challenges, and transformative prospects that define this rapidly evolving field. By inspiring cross-sector collaboration and knowledge sharing, it lays the groundwork for the next wave of innovations poised to reshape computing, sensing, and communication paradigms in the digital age.

In sum, intelligent photonics epitomizes a new computational paradigm where light’s physical properties, enhanced by machine learning, empower devices with unprecedented speed, efficiency, and intelligence. As this interdisciplinary field matures, it is set to redefine the boundaries of what photonic technologies can achieve—ushering in a future where intelligent, adaptive, and sustainable optical systems become ubiquitous across scientific, industrial, and consumer landscapes.

Subject of Research: Intelligent nanophotonics and machine learning integration

Article Title: Intelligent nanophotonics: when machine learning sheds light

Web References: http://dx.doi.org/10.1186/s43593-025-00085-x

Image Credits: by Nanfan Wu, Yuxiang Sun et al.

Keywords: intelligent photonics, nanophotonics, machine learning, optical neural networks, metasurfaces, optical computing, edge sensing, integrated photonics, energy-efficient AI, diffractive optics, photonic circuits