A photonic chip learns to separate partially coherent light

A silicon photonic circuit can now do something that optics has mostly treated as a mathematical exercise: take partially coherent light, learn its underlying coherence modes, and physically route those modes into separate outputs. The result, published October 30 in Nature Photonics by researchers from Stanford, ISTA and collaborating institutions, is a laboratory demonstration rather than a finished imaging system—but it turns a long-standing way of describing mixed light into an on-chip operation.

What the chip is actually separating

Coherence describes how predictably different parts of a light field oscillate relative to one another. Laser light is highly coherent; ordinary sources such as bulbs are much less so; many real optical fields sit between those extremes. When several contributions overlap, their intensity pattern alone can hide information about how those contributions relate in phase.

The new device works with the coherency matrix that describes those statistical relationships. A self-configuring network of Mach–Zehnder interferometers adjusts itself layer by layer, using measured optical power gradients to find the matrix’s eigenmodes. Those orthogonal coherence modes are then routed into distinct waveguide outputs instead of merely being reconstructed later in software.

The experiment moved the decomposition onto hardware

In the main experiment, two independent infrared lasers illuminated a foundry-fabricated silicon photonic circuit through a nine-input spatial sampling geometry. The system learned and separated the two strongest coherence modes even when their spatial intensity patterns overlapped. The paper also reports a rank-3 experiment on a reconfigurable 9 × 9 circuit, showing operation beyond two modes.

When the researchers benchmarked the variational method against an on-chip mixture-based tomography method, they reported eigenvector fidelities of 97.8% for the first learned mode and 99.9% for the second. In a separate case where two incident beams were spatially orthogonal and mutually incoherent, the chip acted as a self-configuring demultiplexer and measured cross-talk as low as −13.0 ± 2.6 dB.

There is an important qualification: the chip is not generally recovering the original physical sources. It diagonalizes the coherency matrix. The separated coherence modes can contain contributions from both input beams and coincide with the original sources only under particular conditions, such as when those sources are already orthogonal.

Why physical separation matters

Traditional coherence analysis can measure a field and numerically reconstruct its modes. This architecture leaves the learned modes physically separated after optimization, so each mode remains available for further optical processing. The authors frame that as useful for programmable coherence analysis and control rather than just a new way to calculate the same answer.

That distinction could matter in measurements where overlapping light carries information that a conventional intensity image collapses together. Stanford’s team points to microscopy, sensing and astronomy as possible applications. In principle, two fluorescent emitters that overlap in an image or several astronomical sources that appear blended could contain separable coherence information even when their brightness patterns are difficult to distinguish.