Coupled nanolaser networks as substrates for photonic neuromorphic computing
Researchers have demonstrated the feasibility of neural networks based on coupled nanolaser networks, a breakthrough that paves the way for the use of symmetry-protected photonic modes for information processing in AI.
References :
Photonic neuromorphic computing using symmetry-protected zero modes in coupled nanolaser arrays, Kaiwen Ji, Giulio Tirabassi, Cristina Masoller, Li Ge, Alejandro M. Giacomotti, Nature Communications 16, 9203. Published: 16 October 2025.
DOI: 10.1038/s41467-025-64252-x (open access publication)
In the quest for energy-efficient artificial intelligence systems, photonic neuromorphic computing offers an interesting alternative to traditional artificial neural networks. Indeed, a classic neuromorphic architecture is composed of a large number of neurons organized in layers, whose dimensions are constantly increasing, at the cost of a growing energy impact. In this context, photonic devices, such as semiconductor nanolasers, are promising systems due to their low energy consumption and high operating speed. However, the robustness of such systems against external disturbances remains a major challenge to be addressed in order to guarantee high-fidelity information processing.
The present study was carried out in the following CNRS laboratory:
Laboratoire Photonique Numérique & Nanosciences (LP2N, CNRS/Institut d'Optique Graduate School/Université de Bordeaux)
In a recent study, researchers experimentally demonstrated that a low-dimensional network composed of coupled nanolasers, acting as a hidden photonic layer, can intrinsically perform non-trivial classification tasks, such as the XNOR logic gate (which tests the equality between two bits presented to it). The complexity of this task cannot be solved by a single-layer neural network, regardless of its size. The robustness of the photonic layer is ensured by the excitation of so-called “zero modes,” which are unique in that they are “protected” from manufacturing imperfections that affect their coupling parameters. The approach is based on a hybrid architecture, where a digital neural layer is used to optimize the shaping of data projected onto the photonic layer. The researchers were thus able to demonstrate the feasibility of binary classification tasks by a neuromorphic hybrid system with high accuracy.
The team then validated the computing power of their photonic architecture by classifying highly compressed handwritten digits, where the boundaries between classes overlapped significantly. These results highlight the potential of photonic architectures based on the symmetry and topology of coupled networks to optimize the size of hidden layers, paving the way for compact and energy-efficient neuromorphic systems. These results are published in the journal Nature Communications.
