A physics-informed network shapes the statistics of light in 3D

A physics-informed neural network has helped researchers control how strongly light fluctuates through a three-dimensional region. Ze-Huan Zheng and Jin-Hui Chen report experimental speckle fields with suppressed intensity variations, alongside fields whose enhanced bright peaks persist as light propagates. The result extends the design target beyond the appearance of a single illuminated plane.[1]

The work appeared in Optics Letters on 16 October. Speckles are the bright and dark grains formed when scattered light waves reinforce or cancel one another. Although their positions look random, their intensity distribution follows statistical rules. The researchers’ method aims to prescribe a different distribution while accounting for how the optical field moves through space.[1]

The difference between a picture and a volume

In an ordinary two-dimensional display, a designer can ask for a bright spot at a particular pixel. A three-dimensional optical field is more difficult. Its appearance changes as light propagates, and imposing a desired arrangement at one plane may inadvertently change it elsewhere. Controlling the distribution of speckles through an entire volume requires accounting for propagation, interference and constraints of what an optical system can physically produce.[1]

Typical fully developed speckle intensities follow a well-known statistical pattern often described by Rayleigh field statistics and an exponential intensity distribution. Unusually concentrated bright peaks or unusually suppressed fluctuations are different regimes. Super-Rayleigh speckles have more pronounced extreme intensity variations; sub-Rayleigh designs suppress parts of that variation. The point is not simply choosing whether an image looks grainy. Researchers want to prescribe statistical behaviour while the field still obeys the physics of light.[1]

Designing backwards from the desired field

There are many ways to shape a laser wavefront, but finding an optical pattern that satisfies constraints at several depths can be a demanding inverse problem. In a forward problem, one starts with a known wavefront and calculates where the light will go. In the inverse version, researchers describe what kind of light distribution they want and seek a wavefront likely to create it. A neural network can help search that design space.[1]

A physics-informed network incorporates optical propagation into the design process, rather than treating each depth as an independent image. The proposed field must still be produced by real optical hardware, where diffraction and interference determine the result. The reported experiments therefore test a physical field, extending the evidence beyond a simulated pattern.[1]

Depending on a measurement task, a researcher may want comparatively even illumination or more pronounced local peaks. Designing the statistics offers another variable to control in an optical experiment. Whether a particular field improves sensing or imaging would need a comparison in the relevant instrument.[1]

From a controlled field to a measurement tool

Useful application tests would examine how closely measured statistics match the requested distribution across depths, how sensitive the field is to alignment and whether the effect repeats across experimental sessions. The publisher abstract supports the two proof-of-concept demonstrations but does not provide device-level speed, efficiency or imaging gains.[1]

The study illustrates a role for machine learning in which the output is a physical distribution rather than an image label. By making unusual speckle statistics an explicit design objective across space, researchers gain another way to compose the illumination used in an experiment.[1]

Report an error or suggest a correction