Nonlinear photonic processors
PhotonQSC designs nonlinear photonic chips that sense and compute in the same optical circuit. Inference runs on the light itself, so weak signals are processed before detection and digitization bury them.
The problem
A conventional receiver converts light to current, amplifies it, digitizes it, and only then runs a model. Each of those steps adds noise. When the incoming signal is weak, the information is gone before the processor sees it.
What we build
Optical nonlinearity is the compute primitive. Resonators confine light until it interacts with itself, which gives the circuit the nonlinear response that linear optics cannot provide.
Kerr effect · resonant enhancement · all-optical activation
The sensing element and the processor share the same waveguides. There is no conversion step between measuring the field and computing on it, so nothing is lost in between.
in-sensor computing · single optical layer · one readout
We train with the noise in the loop: shot noise, thermal drift, and fabrication variation. The chip learns the decision that survives them, at signal levels where a digitized waveform carries too little to work with.
noise-aware training · photon-starved inputs · robust to drift
How we train it
Training a physical system usually means simulating it on a computer and hoping the simulation matches the device. We train on the hardware. Our physical adjoint method reads the gradient from the same chip that runs inference, so the model learns the device as fabricated, with its real losses and its real noise.
The method covers nonlinear circuits and nonreciprocal ones, the two cases where simulation-based training is least reliable.
dχ/dp = −2 Re[ cᵀ (∂A/∂p) e ]What we are building
A receiver that sends the incoming wave back along the path it arrived on, weighted by the link error. The radio channel then computes the beam update itself. A large phased array can follow a moving receiver and route around blockage without estimating the channel or mapping the room.
retrodirective return · reciprocal channel · no explicit channel estimation
A terahertz front end that computes on the incoming wave before it reaches a mixer or an ADC. The circuit extracts features at the carrier frequency and digitizes only the result, at bandwidths where fast converters are costly and power-hungry.
sensing and computing in one front end · analog-domain inference
Where it fits
Keep a high-gain beam locked on a moving receiver through blockage and multipath, with the adaptation computed by the link itself.
Return signals weaken quickly with distance. Processing them on-chip, before detection, extends the range at which a target can still be classified.
Recover symbols close to the sensitivity limit with part of the equalization done optically, ahead of the receiver electronics.
Detect trace absorption or small index shifts when the sample limits how much optical power you can use.
Classify wideband signals in the optical domain, where bandwidth is plentiful, and digitize only the answer.
Specifications
Each design is specified against your signal and your noise. These are the parameters we fix together at the start of a project.
| Material platform | [e.g. silicon nitride, thin-film lithium niobate] |
|---|---|
| Operating wavelength | [X nm] |
| Minimum input SNR | [X dB] |
| Inference latency | [X ns] |
| Optical input power | [X µW] |
| Die footprint | [X × X mm] |
| Packaging | [e.g. fiber-attached module] |
Working with us
You bring the measurement and its noise sources. We define the task, the SNR range, and the decision the chip has to make.
We lay out the photonic circuit and simulate it against your noise model until it meets the target.
The design is taped out at [FOUNDRY PARTNER] on a standard process, then diced and packaged.
We train each chip in hardware against real signals and deliver it with measured performance data.
Contact
Tell us what you are measuring and where the noise comes from. An engineer will reply with whether a photonic front end can help.
Or write to us directly hello@example.com