Researchers show how nonlinear optics paves the way for intelligent photonic systems.

Artificial intelligence is transforming the world – but it requires vast amounts of energy. Modern language models and image analysis systems now consume as much electricity as entire cities. Researchers at Leibniz IPHT are therefore pursuing a different approach: rethinking computation itself – using light.

Optical fibers can do more than transmit data. They can also process information. This is demonstrated by the Smart Photonics junior research group led by Prof. Mario Chemnitz. The researchers take an unconventional approach, investigating how the natural mixing of light frequencies can filter and weight information to solve tasks that are typically handled by digital neural networks. Their work provides a building block for future intelligent photonic systems.

When Light Performs Computation

In their experiments, the team with first author Sobhi Saeed sent short, encoded light pulses through an optical fiber. Each pulse carries a dataset embedded in its spectral phase. Depending on the encoded intensity and temporal structure, the light spectrum evolves in complex ways as it propagates through the fiber. The resulting output spectrum contains structurally filtered information about the input data and can be used for tasks such as classification, segmentation, or prediction.

The fiber responds nonlinearly: small differences in the input lead to large changes in the output. Effects that are typically undesirable in classical signal transmission are here used deliberately – as computational operations. “We use the dynamics of light as a computational resource,” explains Mario Chemnitz. “Light interacts with the fiber in a way that maps complex patterns – without digital processors or programmed parameters. The result emerges directly in the light itself.”

At the fiber output, the spectrum is analyzed, filtered, and linearly combined in simple steps tailored to the task at hand. In their experiments, the researchers used different types of optical fibers and tested how well they could distinguish between input signals. The stronger the nonlinear interactions, the more precisely the system could separate patterns.

A New Metric for Intelligence in Light

One challenge remains: the architecture presented by Saeed and the team links information in a largely uncontrolled way – a typical feature of socalled reservoir computing systems. As a result, the exact flow of information is not directly accessible. To address this, the researchers introduce a new concept in photonics: nonlinear inference capacity. This metric describes how effectively a physical system can disentangle complex patterns – in other words, how “deeply” it can process information.

For the first time, this makes it possible to compare the performance of analog physical systems with that of digital neural networks. The data analysis and machine learning methods were developed in collaboration with Prof. Thomas Bocklitz and his research department Photonic Data Science.

“We wanted to understand how deeply a physical system can compute,” says Chemnitz. “Our results show that nonlinear optics is a genuine computational resource. It can represent learning processes that electronics cannot achieve in the same way.” In tests, the optical system even outperformed digital networks with multiple hidden layers. The reason: many interactions occur simultaneously within the fiber. Weighting and filtering processes run in parallel rather than sequentially, as in electronic circuits.

Analog Instead of Digital

While digital systems break information down into zeros and ones, the optical fiber processes data continuously—in an analog flow of energy and time. This resemblance to how the brain operates makes the system particularly efficient: relationships are encoded in transient information flows and can be accessed directly. “The fiber is a neural network made of light,” says Sobhi Saeed. “It recognizes patterns by changing its physical states – not through programming, but through intrinsic interactions.”

The researchers tested their concept using a new, standardized dataset designed to quantify the computational power of analog systems. For the first time, they were able to show that the nonlinear dynamics of light can match – and under certain conditions even exceed – the “computational depth” of artificial neural networks.

Photonics as a Foundation for Intelligence

This work is the first publication of the Smart Photonics junior research group, founded in 2022 at Leibniz IPHT and funded by the Carl Zeiss Foundation through the CZS Nexus program. It marks the beginning of a research direction that brings together physics, artificial intelligence, and hardware development: intelligent photonic systems.

“Our vision is hardware that can sense, process, and understand information without ever converting it into electrical signals,” says Chemnitz. “Light is inherently fast, parallel, and energy-efficient. If we harness these properties, we can build future computing and machine- learning systems that operate directly in the optical domain.” In the long term, this research could enable new forms of analog, learning-capable information processing – not confined to data centers, but operating directly where data is generated, such as in sensors, laboratory instruments, or diagnostic devices.

 

Original publication: https://doi. org/10.1515/nanoph-2025-0045