In a recent article published in the journal Engineering, researchers presented a comprehensive review of the fast-growing field of intelligent photonics, highlighting its potential to significantly advance the integration of artificial intelligence (AI) with optical technologies.
Image Credit: Alexander_Evgenyevich/Shutterstock.com
Background
The foundation of this interdisciplinary field is built on a rich legacy of advancements in optics, materials science, and computational methods. At the same time, breakthroughs in AI, particularly in deep learning, have shown remarkable capabilities in areas like pattern recognition, image reconstruction, and complex data analysis. The convergence of these domains has given rise to metaphotonics, a field that harnesses nanostructured materials to develop compact, tunable optical components designed for AI-driven optimization. This synergy also extends into areas such as holography and quantum photonics, enabling optical systems to take on increasingly sophisticated roles, like 3D imaging, ultra-fast computation, and quantum information processing, with greater efficiency.
Studies Highlighted in this Review
The core methodological approach centers on using deep learning algorithms to enhance the design, optimization, and control of optical devices, particularly metasurfaces and related components. Traditionally, designing metasurfaces has relied on complex, iterative physics-based simulations that are both time-consuming and constrained in exploring the full range of design parameters. To address this, the authors point to neural networks (especially deep learning models) that can learn the intricate relationships between the geometric properties of nanostructures and their optical behaviors. These models make it possible to perform rapid inverse design, where a target optical response is defined and the network quickly predicts the nanostructure configuration needed to achieve it.
The authors also explore the development of optical neural networks (ONNs), which use photonic components, like cascaded Mach-Zehnder interferometers (MZIs), wavelength-division multiplexing (WDM), and programmable metasurfaces, to perform neural network operations with high speed and low power consumption. These systems take advantage of light’s wave properties to enable parallel processing, using interference and diffraction effects to carry out tasks such as matrix multiplication. To physically realize these nanostructures, fabrication methods such as advanced lithography and nanofabrication are employed. The design process itself is often guided by AI algorithms that help optimize the parameters to achieve specific, high-performance functionalities.
The review also covers holography-inspired neural network architectures, where deep learning models are used to design holographic devices capable of producing high-fidelity 3D image reconstruction and display. Another key area is quantum photonics, which brings together quantum effects and AI techniques to enable ultra-fast quantum algorithms and the development of scalable quantum optical circuits.
Discussion
The integration of AI techniques, especially deep learning, into optical device design has led to several notable breakthroughs. Neural networks have demonstrated the ability to learn the inverse mapping between target optical functionalities and the corresponding nanostructure parameters, dramatically speeding up the development process for metasurfaces and other nanophotonic components. This data-driven approach enables the optimal design of metasurfaces capable of complex tasks such as wavefront shaping, polarization control, and spectral manipulation, with wide-ranging applications in advanced imaging, beam steering, and beyond.
In the area of optical neural networks, photonic systems built on MZI meshes, wavelength-division multiplexing (WDM), and tunable metasurfaces show strong potential for ultrafast, energy-efficient AI inference, outpacing traditional electronic approaches. These architectures exploit wave interference to carry out matrix-vector multiplications, the fundamental operation in neural networks, at the speed of light. This enables inherent parallelism with minimal heat generation. Experimental results have demonstrated high accuracy in tasks like image classification, signal processing, and data encoding, highlighting their promise for real-world applications in autonomous systems and high-speed data centers.
Metaphotonics, empowered by AI-driven design, has produced lightweight, multifunctional optical components capable of dynamic modulation and multi-channel capabilities. Such devices miniaturize traditional bulky optics and enable sophisticated control over light fields, including phase, amplitude, and polarization. These advances support applications in compact imaging systems, augmented reality, and biomedical device engineering.
Holography-based neural networks, designed via deep learning, allow for the generation of high-resolution, high-fidelity 3D images with applications in display technology and optical data storage. Quantum photonic systems, optimized through AI strategies, are positioned to accelerate certain quantum algorithms and support scalable quantum optical circuits , highlighting their future promise in tackling computationally intensive tasks such as quantum simulation and secure communication.
Conclusion
The reviewed article underscores the powerful potential of intelligent photonics, driven by the convergence of AI and cutting-edge optical engineering. Through deep learning, researchers are streamlining the inverse design of metasurfaces and nanophotonic components, enabling the creation of devices with capabilities that go well beyond those of traditional optics. The development of optical neural networks presents a compelling path forward for ultra-fast, energy-efficient AI computation, offering solutions to the performance and power constraints of electronic systems, particularly in data centers, autonomous vehicles, and portable electronics.
While these advances hold enormous promise, numerous challenges persist, particularly in fabrication scalability, device robustness, and integration complexity. Future research directions will likely focus on developing resilient algorithms that account for manufacturing tolerances, exploring new materials for reconfigurability, and designing scalable architectures that integrate seamlessly into existing technological ecosystems.
Journal Reference
Xu D., Ma Y., et al. (2025). Intelligent Photonics: A Disruptive Technology to Shape the Present and Redefine the Future. Engineering, 46, 186–213. DOI: 10.1016/j.eng.2024.08.016, https://www.sciencedirect.com/science/article/pii/S2095809924005149