Most of the light in the world is only partly coherent, and photonics has mostly treated that as a defect. Lasers — the light source nearly every photonic system is built around — are coherent by design: their waves march in step. But sunlight, starlight, fluorescence, and a laser beam that has passed through a turbulent atmosphere are partially coherent, a mixture of waves with only statistical relationships between their phases. In a paper published November 3 in Nature Photonics, researchers at Stanford University and the Institute of Science and Technology Austria (ISTA) report a chip that does something no device has done before: it takes partially coherent light in, works out its coherence structure on its own, and sends the components out through separate waveguides, in real time, without being told what to look for.
The device is a self-configuring photonic circuit — a two-layer network of Mach–Zehnder interferometers on a foundry-fabricated silicon chip, driven by autonomous control algorithms. In the main experiment, two independent infrared lasers at approximately 1,535 nanometres were overlapped into the chip, forming a controlled partially coherent input. The chip learned the input’s coherence modes and routed them to separate outputs, with fidelities of 97.8% and 99.9% against a reference tomography performed on the same chip. A three-laser experiment on a separate fully reconfigurable 9 × 9 circuit showed the approach extends beyond two modes.
That is a genuine first, and a narrow one. The demonstration used laboratory lasers, not starlight, a microscope sample, or a communications signal. What the team has shown is that a compact chip can physically perform a mathematical operation — diagonalizing a coherency matrix — that until now existed only as theory and simulation. The applications the researchers name are prospects. The chip is the result.
What was actually demonstrated
Precision first. The paper, “Automatic on-chip separation of partially coherent light into its mutually incoherent components,” is authored by Paul-Alexis Mor and colleagues, with Mor and Annie Kroo as co-lead authors, Charles Roques-Carmes as the project lead, and David A. B. Miller, Shanhui Fan, and Olav Solgaard among the co-authors. Roques-Carmes began the project at Stanford with Mor, then moved to ISTA in June 2026 to establish his research group; Mor accompanied him to continue the work there.
The chip itself is small — fingernail-sized, a few square millimetres — and built from a two-dimensional array of nine grating couplers that couple free-space light into the chip, feeding two self-configuring layers of Mach–Zehnder interferometers. Each interferometer contains two independently controlled, thermally actuated phase shifters and two 50:50 directional couplers. The chip was designed and fabricated through a foundry process, then mounted on a printed circuit board with its own electronics and detectors. The paper does not name the foundry.
Two mutually incoherent lasers at approximately 1,535 nm formed a rank-two partially coherent field, and the chip’s job was to find its two coherence modes — the orthogonal eigenvectors of its coherency matrix — and route each to its own output waveguide.
It did so by variational coherence tomography, a method built on a framework Roques-Carmes, Fan, and Miller published in Light: Science & Applications in 2024. The first self-configuring layer optimizes all of its phase shifters to maximize the power at one output; the second layer does the same for the next mode. The optimization is gradient-based: each phase shifter is sinusoidally dithered at its own frequency, and the photodetector signal is demodulated against those drive frequencies to measure all the gradients in a layer simultaneously — up to 14 phase shifters in one step. The gradients feed a gradient-descent optimizer, and in a representative run they converged to zero within approximately 50 iterations.
Two checks established that the separation was real. First, the recovered modes were compared against a reference method — “mixture tomography” — in which each laser was switched on individually and the coherency matrix reconstructed numerically from separate measurements. Against that benchmark, the two recovered modes had fidelities of 97.8% and 99.9%. Second, the two outputs were recombined through an interferometer, and the fringe visibility progressively vanished as the optimization converged, showing the outputs had become mutually incoherent.
A third experiment, on a separate fully reconfigurable 9 × 9 circuit, fed in three lasers and showed three optimized outputs becoming mutually incoherent — the method is not limited to two modes.
The distinction that keeps the claim honest
The most important sentence in the paper, for anyone writing about this result, is one the authors state explicitly: the chip diagonalizes the coherency matrix; it does not necessarily recover the original beams. The coherence modes it separates are the orthogonal eigenvectors of the combined field’s coherency matrix, and in general those modes contain contributions from both input lasers. The modes coincide with the original beams only in the special case where the two input fields are orthogonal — their spatial patterns having zero overlap in the mathematical comparison that includes phase as well as intensity. In the experiments, the overlapping beams satisfied that condition closely enough that the outputs corresponded to the separate lasers, so the result can fairly be described as separating a mixture back into its components. But the general claim is about modes, not sources, and coverage that says the chip “recovers the original lasers” in all cases is overstating the physics.
That distinction matters for the applications, too. A microscope that separates overlapping fluorescent emitters, or a telescope that separates light from a star and an exoplanet, is not recovering clean copies of two sources; it is extracting the independent statistical components of a mixed field — exactly what an intensity-only camera throws away. David Miller put it plainly in the Stanford coverage: “In a camera, we throw away the depth information. By doing this separation of light, we know everything knowable about it.” The chip’s value is that the separated modes remain available as light in waveguides, ready for further optical processing, rather than existing only as numbers in a computer reconstruction.
The self-configuring character is the other half of the result. Nothing about the input is programmed in advance: the chip learns the modes from the light itself, and once configured, the separation persists in the optical network independently of the detectors. The same approach can run in reverse, generating partially coherent light with a chosen coherency matrix — demonstrated in the supplementary information — and continuous optimization can compensate for fabrication imperfections and slow drifts. A photonic circuit that calibrates itself in operation is a different proposition from one that must be characterized once, externally, and hoped to stay put.
Why a chip, and why now
The hardware argument in the paper is a familiar one in integrated photonics, and it is worth taking seriously. The same function built from discrete optics would require hundreds of mirrors and other components spread across an optical table, aligned and stabilized to interferometric precision. On a chip a few square millimetres in size, those components sit in fixed relative positions, turning the alignment problem into a fabrication problem.
Roques-Carmes made the point directly: “This is a rare case in which the hardware industry was ready before the application. Thanks to advances in the photonic chip industry, we can now integrate hundreds of optical components onto a single chip just a few square millimeters in size.” The chip is foundry-fabricated, which is what makes that claim more than rhetoric — a foundry device and a one-off cleanroom device have different scaling stories.
There is also a scaling argument in the method. For a general N-dimensional field, a full coherency-matrix reconstruction requires on the order of N² measurements. But in many practical settings the matrix is effectively low rank — rank r, much smaller than N — and the team’s method exploits that: each of the r sequential optimizations uses at most on the order of N physical components, so the total component count scales as rN, not N². For the low-rank fields that dominate real sensing and imaging problems, that is the difference between a tractable chip and an impossible one.
Where this could go — and what it has not done
The applications the researchers name are real research directions, and none has been demonstrated.
Microscopy and imaging are the nearest prospect. Under a microscope, two fluorescent particles that overlap spatially can appear as one in a conventional image. A coherence-sorting chip could, in principle, distinguish the light from the different sources even where they overlap, because the information is in the light’s statistical structure, not its intensity profile. The team is working toward a microscope with ISTA biologists. No microscope has been built, and no biological sample has been imaged with this method.
Satellite communication is the second named application. Laser signals passing through Earth’s atmosphere lose coherence to turbulence, and a receiver that can analyze the resulting partially coherent field might handle those disturbed signals better. The present experiments did not demonstrate a working satellite link, did not pass light through turbulence, and established no performance under real atmospheric conditions.
Astronomy is the most speculative and the most evocative. Miller, Solgaard, and Kroo are collaborating with NASA to explore whether the technology could probe light from distant galaxies, where a single blob in a telescope might contain distinct strands from separate stars and exoplanets. “We want to make our chips as efficient as possible so that every photon counts,” Roques-Carmes said. No astronomical observation was part of this demonstration. Optical computing is also named in the paper’s framing, but no computing operation was demonstrated and no performance figure for a computing task is reported — this is a coherence-processing result, not a photonic computing result.
What is undisclosed, and why it matters
An honest accounting of what the paper and the coverage do not contain is part of the story, not a footnote to it.
First, no end-to-end loss figure is reported in the main text. How much of the incoming light survives the grating couplers, the two interferometer layers, and the routing to the outputs is not stated as a headline number. The paper addresses loss only in the supplementary information, where numerical simulations quantify the method’s robustness to realistic losses, thermal cross-talk, and finite detection signal-to-noise — but a simulation of robustness is not a measured insertion loss, and this article asserts none. For any sensing application, where the selling point is “every photon counts,” the loss budget is the number that will decide whether the method is practical, and it is not yet on the record.
Second, no wall-clock reconfiguration speed is reported. Convergence within approximately 50 optimization iterations is stated, but the time per iteration is not given, so no claim about how fast the chip can track a changing field can be made from this paper. Thermal phase shifters are power-hungry and slow relative to electro-optic alternatives, and the paper gives no power budget. Whether the loop can keep up with a turbulent atmosphere or a moving microscope sample is an open question.
Third, the foundry is not named, and no fabrication yield, wafer count, or cost figure is given. “Foundry-fabricated” establishes a commercial process, not a research cleanroom, but not that the process is qualified, repeatable at volume, or available to other groups. The next generation of chips is being taped out for operation at ISTA, which is the concrete near-term manufacturing signal.
Fourth, a factual correction to the trade coverage: the Nanowerk report of November 4 describes “overlapping ultraviolet beams from two independent lasers.” The paper and the Stanford coverage are clear that the experiment used infrared light at approximately 1,535 nm. The ultraviolet description appears to be an error in the Nanowerk piece, and this article follows the paper.
Finally, the scale of the demonstration is two modes in the main device and three in the 9 × 9 circuit. Real fields — a microscope image, a turbulent beam, a galaxy — have far higher-rank coherency matrices. The rN scaling argument says the component count grows manageably, but no high-rank demonstration exists and fidelity at higher rank is untested. The gap between three modes and a useful instrument is the work ahead.
What to watch
The signals that will move this story from demonstration to device are specific. First, a loss figure: a measured end-to-end insertion loss, with the grating-coupler contribution separated from the circuit contribution. Second, a speed figure: reconfiguration time, and whether the loop can track a changing input rather than a static one. Third, the next-generation chips being taped out for ISTA — whether they scale the mode count, and whether the foundry is named. Fourth, an application demonstration: a microscope image, a turbulence-distorted beam, or an astronomical measurement where the chip’s output changes a result intensity-only measurement could not obtain. Fifth, a power figure for the thermal phase-shifter control loop, which will decide whether the architecture can leave the bench.
The fair summary is a narrow one. Researchers at Stanford and ISTA have built a foundry-fabricated photonic chip that takes a partially coherent input, learns its coherence modes without prior knowledge of the field, and routes them into separate waveguides with fidelities of 97.8% and 99.9% against on-chip reference tomography — demonstrated with two lasers at approximately 1,535 nm, and extended to three modes on a separate 9 × 9 circuit. No field, imaging, or communications application has been demonstrated, no insertion loss or reconfiguration speed is reported, and the foundry is unnamed. That is a modest claim — and precisely because it is modest, it is believable. A theoretical operation from the 1980s, formalized in 2024 and now physically realized on a chip, is the story. What it becomes is the next one.
Sources and further reading
Mor, P.-A., et al., “Automatic on-chip separation of partially coherent light into its mutually incoherent components,” Nature Photonics (2026), DOI: 10.1038/s41566-026-02035-6, published November 3, 2026 — foundry-fabricated photonic integrated platform for spatially partially coherent light; nine grating couplers feeding two self-configuring Mach–Zehnder interferometer layers; two mutually incoherent beams at λ ≈ 1,535 nm; fidelities of 97.8% and 99.9% against mixture-based tomography on the same chip; convergence within approximately 50 iterations; rank-3 verification on a fully reconfigurable 9 × 9 circuit; coherence-entropy discrimination of co-localized beams; self-configuring separation of orthogonal mutually incoherent beams; r sequential optimizations, component scaling O(rN); loss robustness addressed in supplementary simulations, no headline insertion-loss figure in the main text; foundry unnamed.
Roques-Carmes, C., Fan, S., and Miller, D. A. B., “Measuring, processing, and generating partially coherent light with self-configuring optics,” Light: Science & Applications 13, 260 (2024), DOI: 10.1038/s41377-024-01622-y — conceptual framework for the method: cascaded self-configuring layers, sequential optimization of average power outputs, diagonalization of the input coherency matrix, numerical simulations of Mach–Zehnder interferometer arrays; background for the 2026 experimental demonstration.
Stanford / phys.org, “Fingernail-sized chip untangles mixed light into clean beams,” November 3, 2026 — Stanford coverage of the Nature Photonics paper; two infrared lasers mixed and separated; third laser added and sorted; quotes from Annie Kroo, Charles Roques-Carmes, Olav Solgaard, David A. B. Miller, Shanhui Fan, and Paul-Alexis Mor; NASA collaboration on astronomical applications; theoretical origins in the 1980s.
Nanowerk, “Splitting light on a chip,” November 4, 2026 — trade coverage of the Nature Photonics paper; quotes from Roques-Carmes and Mor; chip described as a few square millimetres with hundreds of optical components; next-generation chips being taped out for operation at ISTA; microscopy work with ISTA biologists. Note: the Nanowerk piece describes the input beams as ultraviolet; the paper and Stanford coverage state infrared light at approximately 1,535 nm, and this article follows the paper. No imaging, field, or communications demonstration is reported in any source cited here, and none is asserted in this article. This article is analysis, not investment advice.