Photonic neuromorphic processors promise fast and energy-efficient computing, but compact delay lines remain a bottleneck. Alongside this work on the archive are studies of chaos in crystals, tunneling resonances, and optical reservoir computers based on cesium vapors, but they all either use bulky fiber coils or digital buffers that do not preserve optical phase. This work stands apart: it is the first to integrate a passive silicon spiral, fabricated in a standard foundry, into the feedback loop and test it on real computational tasks.
At the LIGENTEC foundry, using the AN350 platform, a silicon nitride chip was fabricated with eight consecutive Archimedean spirals. Each spiral is about 0.22 meters long, and the total optical path reaches 1.76 meters with a group index of 1.981, giving a calculated delay of 11.63 nanoseconds. The chip was inserted into the feedback loop of an optoelectronic reservoir: a continuous-wave laser at 1553 nm is modulated by an electro-optic Mach–Zehnder modulator, light passes through the spiral, reaches a photodetector, and the electrical signal returns to the modulator. The single nonlinear node is expanded into 50 virtual neurons via time multiplexing with a step of 20.48 nanoseconds. Hyperparameters were optimized by Bayesian search directly on the setup, and training was limited to a linear output layer.
With a feedback gain of G = 0.0223, which is 40 times lower than typical values, the system classified sinusoidal and square signals without errors (WER = 0 on all folds, best NMSE = 0.074). Prediction of the Mackey–Glass chaotic series achieved NMSE = 0.0082, which is only 8% worse than the best optoelectronic result at G around 1. Recognition of nine Japanese vowel speakers gave WER = 0.0898, which is twice as good as the previous speckle processor and only 2.5 percentage points worse than a digital delay line at a gain 23 times higher. Additionally, a subcarrier phase encoding method was introduced: deliberate aliasing of a 10 MHz signal at a sampling rate of 976.6 kHz allowed the phase difference between the spiral and a straight waveguide (0.232π) to be used as a computational resource, achieving NMSE = 0.0767 and WER = 0 without increasing the gain.
This work differs from neighboring studies in that it is the first to use a passive integrated delay line on silicon nitride inside the reservoir computer loop, rather than fiber, a digital buffer, or a resonant structure. It proves that even with losses of 23.2 dB and a gain of 0.0223, useful computation can be preserved, and the phase information of the optical path becomes an additional encoding channel unavailable to digital buffers.
The next step is obvious: reduce excess losses from 23.2 dB to a few decibels, which would allow higher gain and likely improve accuracy. More integrated versions could combine the laser, modulator, and detector on a single chip, and the phase encoding method opens the way to new architectures where phase is used directly for signal processing.
The results will influence the development of compact photonic accelerators for time-series tasks, speech recognition, and fast signal classification, especially where small size and low power consumption are important.
Fabricate a second-generation chip with lower losses and test it on more complex tasks, including real-time online signal processing.
For photonics, the key unresolved problem is the integration of passive delays with low loss and phase preservation. This work shows that even with current losses, computation is possible, but the question of the ultimate scalability of such lines on a chip without a critical drop in signal-to-noise ratio remains open.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The work uses a silicon spiral as a delay line in an optoelectronic reservoir computer. Eight cascaded Archimedean spirals provide an optical path of 1.76 m and a delay of 11.63 ns. With a feedback gain of 0.0223, the system classifies signals, predicts the Mackey-Glass series (NMSE 0.0082), and recognizes speakers (WER 0.0898).
Where its strength is
For the first time, a passive integrated silicon nitride delay line is used as a functional element of a reservoir computer, retaining computational capabilities at a gain 40 times lower than usual. Phase encoding of the subcarrier adds an extra channel that digital buffers cannot provide.
Where it could go next
Worth trying
Why this may matter
Compact passive delay lines on a silicon chip could become a key element of scalable, low-power photonic reservoir computers, bringing them closer to practical use in signal processing and recognition.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.52, against a usual 0.64. It looks like few have drilled nearby.
The Archimedean spiral used in the chip is named after the ancient Greek mathematician who described it in the 3rd century BC for solving the problem of trisecting an angle. Now the same geometry compresses 1.76 meters of optical path into a few square millimeters.
The idea of delay-based reservoir computing echoes the 'liquid' neural networks in Greg Egan's novel 'Permutation City', where computational substrates use time delays to process information without traditional logic elements.