AI Designs Light at the Nanoscale: Russian Scientists Create the World's First Generative Model for Ultra-Sensitive Optical Engineering

Scientists at ITMO University's New Physics Institute have developed MetaDiT, the world's first generative AI model capable of automatically designing metasurfaces. These ultrathin optical components manipulate light at the nanoscale. The work, carried out jointly with a research center in Qingdao, China, has already been accepted for presentation at the prestigious AAAI-26 international conference.

Metasurfaces are flat optical components whose surfaces consist of arrays of nanoscale structures. Those structures determine how light passes through a material or reflects from it: they can focus a beam, redirect it, alter its polarization, or reshape its spectral response. Over time, such components could replace or complement bulky combinations of lenses, mirrors, and filters in cameras, virtual reality headsets, and medical instruments, making future devices lighter, more compact, and more capable.

AI as a Nanoengineer

Before MetaDiT, designing metasurfaces largely resembled a brute-force search through millions of possible configurations. The new model does more than generate a nanostructure pattern. It simultaneously defines its physical characteristics, including thickness, the material's refractive index, and lattice period. That addresses one of photonics' fundamental challenges known as inverse design. Instead of manually exploring enormous design spaces, researchers specify the desired spectral response, and the AI proposes an optimal structure, reducing work that once required years of calculations to just a matter of hours.

Technological Sovereignty and the Next Generation of Smart Glasses

What does this mean in practice? For Russia's technology sector, it demonstrates that domestic researchers can build highly specialized neural networks capable of solving some of engineering's most demanding design problems. MetaDiT could become the core of Russian computer-aided photonics design platforms, reducing reliance on foreign engineering software.

The technology also points toward a new generation of consumer devices. Metasurfaces can replace bulky lenses, mirrors, and optical filters. In the future, that could lead to virtually weightless virtual and augmented reality glasses, ultra-precise medical analyzers capable of detecting DNA molecules, and compact sensors for autonomous drones.

An Evolution Measured in Nanometers

MetaDiT represents the latest milestone in ITMO University's long-term research program. In 2023, the university introduced perovskite nanolasers and launched the METANANO school in Qingdao. In 2024, researchers developed nanostructures for medical biomaterial analysis. In 2025, the university deployed the TensorTrain method to compress complex matrices and improve the efficiency of silicon photonic chips. Now, in 2026, that accumulated expertise in data and algorithms has converged into a single generative AI breakthrough.

The next challenge, however, lies in moving from virtual designs to physical manufacturing. An AI-generated structure that performs perfectly in simulation may prove too difficult or too expensive to fabricate using nanolithography. The next step is to adapt the model so that it designs within real manufacturing constraints while accounting for available materials.

The technology also has significant export potential, particularly across Asian markets. Collaboration through the joint research center in Qingdao has already created an established platform for international partnerships. MetaDiT could eventually be offered as a cloud-based software service. Even so, an AI model alone cannot build an industry. Commercial success will also require domestic manufacturing capacity, advanced metrology, and strong intellectual property protection.

A Paradigm Shift

MetaDiT marks a fundamental turning point. Artificial intelligence is no longer acting solely as an analytical tool but is becoming a full engineering co-designer. Where human researchers and conventional computing methods struggle with billions of possible combinations of geometries and materials, the neural network can identify elegant solutions. Over the medium term, systems like MetaDiT are likely to reshape both research laboratories and advanced manufacturing, demonstrating that Russia can develop specialized AI not only for conversational systems but also for breakthroughs in hardware engineering and fundamental physics.

In metasurfaces, the overall spectral profile and its fine details are equally important. That is why we built two levels of control into the architecture. The first, a coarse level, defines the model's overall objective by specifying which regions of the spectrum should transmit light efficiently and which should not. The second, a fine-grained level, ensures that local features such as narrow resonances, peaks, and dips are preserved, because they often determine whether a device will function at all. On their own, neither the global picture nor the fine details provide sufficient accuracy. In the first case, the model may overlook critical features. In the second, it may lose the broader context. By combining both levels, we enable the model to retain both simultaneously. That is why MetaDiT is less like an artist completing a predefined pattern and more like an engineer designing a structure for a specific physical problem

Andrey Bogdanov

PhD in Physics and Mathematics, Head of the MetaDiT Development Team, Leading Research Fellow, Faculty of Physics, ITMO University