A tellurium thin film that briefly melts under laser pulses delivers the first sub-nanosecond all-optical neuron, operating 100 times faster than previous designs.
(Nanowerk Spotlight) Photonic neural networks should be able to run artificial intelligence at the speed of light. Optical waveguides can carry out the heavy matrix arithmetic that dominates deep learning far faster and more efficiently than electronic chips. But between every layer of calculation sits a small, essential step called nonlinear activation, the operation that gives a neural network its ability to recognize faces, parse language, or make predictions rather than simply crunch numbers.
No purely optical device has been able to perform this step without converting the signal to electricity, processing it on a conventional chip, and converting it back to light. This detour happens at every layer, sometimes dozens of times per inference, burning time and energy at each handoff. The result is that today’s photonic neural networks operate far below their theoretical speed, throttled not by the light itself but by the electronic neuron in the middle. Even the most advanced ultrafast photonic processors for AI still rely on electronic components for this critical nonlinear step.
What has been missing is a material that changes how it transmits light, sharply and rapidly, when struck by an optical pulse. Without that, there is no way to produce the nonlinear response a neuron requires while staying in the optical domain. Thermo-optic approaches, which use absorbed light to heat a waveguide and shift its refractive index, are limited by thermal diffusion to microsecond response times.
Phase-change materials such as Ge₂Sb₂Te₅ (GST) shortened that to hundreds of nanoseconds. Free-carrier dispersion effects, including those explored in silicon microresonator-based photonic neurons, reached about 10 nanoseconds but demanded bulky resonators or interferometers that inflate device size.
A study published in Advanced Materials (“A Phase-Transition-Driven All-Optical Neuron with Sub-Nanosecond Nonlinear Activation”) reports that elemental tellurium can fill this gap through a mechanism none of these prior approaches exploited: a light-triggered solid-to-liquid phase transition. When an optical pulse strikes a thin tellurium film on a silicon nitride waveguide, the material absorbs enough energy to melt locally.
This abrupt change of physical state sharply increases tellurium’s extinction coefficient, the property governing how strongly it absorbs light, producing a fast nonlinear drop in optical transmission.
Once the pulse ends, the molten region cools and recrystallizes on its own, resetting the device without any external signal. That self-resetting property sets tellurium apart from conventional phase-change materials. GST switches between two stable solid states, amorphous and crystalline, at room temperature. Because both states persist indefinitely, a GST-based neuron needs an extra reset pulse after each firing, adding time and energy to every cycle.
All-optical neurons. (a) Schematic diagrams of optoelectronic and all-optical deep neural networks are shown, wherein the optoelectronic network consists of photonic synapses (Syn.), electrical neurons (Neu.), and optoelectronic conversion (OE / EO Conv.) systems. The all-optical network, on the other hand, contains both synapses and neurons in the optical domain. Replacing electrical neurons with all-optical neurons can effectively reduce the speed and energy bottleneck caused by OE / EO conversions, thereby enhancing the energy efficiency of neural networks. (b) Conventional all-optical neurons based on thermo-optic and free-carrier dispersion effects. (c) A phase-transition-driven, all-optical neuron employing elemental tellurium. (Image: Reproduced with permission from Wiley-VCH Verlag) (click on image to enlarge)
Tellurium’s liquid phase exists only while energy is being supplied. Remove the light and the material returns to its crystalline ground state, mimicking the fire-and-reset rhythm of biological neurons.
Confirming that the nonlinearity stems from a genuine phase transition rather than ordinary heating required careful testing. When input energy stayed below the threshold of about 1.21 picojoules, tellurium’s crystalline grain structure remained unchanged and waveguide transmission held steady, ruling out thermal attenuation at that level.
Above the threshold, regions closest to the optical input showed large changes in grain orientation, a signature of melting and recrystallization that simple heating could not produce. Optical-thermal simulations confirmed that those same regions reached tellurium’s melting temperature of roughly 447 °C while more distant areas did not.
The speed implications are striking. The tellurium device achieved a total response time below 260 picoseconds, the first time any all-optical neuron has broken the nanosecond barrier and a nearly 100-fold gain over the fastest prior designs. That speed matters because it begins to close the gap between neurons and the photonic synapses they must keep pace with.
Energy consumption fell in proportion: threshold energies reached as low as 0.4 picojoules for a 3.3 µm device under 280-picosecond pulses, roughly 100 times lower than competing technologies. And because the phase transition directly changes the extinction coefficient without needing a resonator or interferometer, the active footprint shrank to between 1.56 and 4.29 µm², small enough to integrate densely on chip.
The neuron’s response curve also proved tunable, resembling the rectified linear unit (ReLU) function at lower input energies and shifting toward a sigmoid shape at higher energies, allowing the activation behavior to be adjusted simply by changing pulse power.
To test practical utility, the team integrated tellurium neurons into a three-layer deep neural network designed to classify handwritten digits from the MNIST dataset. A single tellurium neuron processed nonlinear activations for 50 images in about 20 microseconds using 2-nanosecond optical pulses.
An equivalent optoelectronic system, limited by its analog-to-digital converters, required roughly 5 milliseconds for the same task, making the all-optical approach about 100 times faster. Four-channel wavelength division multiplexing extended throughput further, handling 200 images in the same time window.
Recognition accuracy reached 96.4%, modestly below the 98.2% of a purely electronic implementation but well above the 92.5% of a network with no nonlinear activation. The gap traces primarily to precision: tellurium neurons operate at roughly 6-bit precision, compared with 32 bits in software, introducing small rounding errors that accumulate across layers.
Durability testing showed the neurons sustained more than 5 × 10⁹ switching cycles under 280-picosecond pulses, surpassing GST-based neurons by two orders of magnitude. After a week of shelf time, response curves showed no measurable drift. The neurons also responded consistently across the telecommunication C-band, from 1528.8 to 1550.9 nm, making them compatible with dense wavelength division multiplexing systems supporting 80 or more parallel channels.
The researchers went a step further by fabricating a fully integrated all-optical computing chip combining GST-based photonic synapses for linear weighting with tellurium neurons for nonlinear activation. This chip achieved a compute density of 3.60 TOPS/mm² and an energy efficiency of 13.65 TOPS/W, exceeding comparable optoelectronic networks by one to two orders of magnitude.
Unlike earlier light-based chips for AI computation, this system keeps the entire inference pipeline in the optical domain. Insertion loss from the tellurium film, measured at 2.1 to 2.7 dB/µm, is somewhat higher than GST’s ~1.5 dB/µm but remains manageable because the per-neuron loss is distributed across many channels in a full network.
Further optimization, including sub-picosecond laser pulses and refined device geometries, could push response times below 35 picoseconds and threshold energies under 0.1 picojoules. At that point, photonic neurons would nearly match the speed of the synapses around them, removing the last major obstacle between photonic neural networks and the speed-of-light computing they were always meant to deliver.
For authors and communications departmentsclick to open
Lay summary
Prefilled posts
Nanowerk Newsletter
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
