The researchers experimentally reprogram one thin-film lithium-niobate slab waveguide to route spatial modes, sort beam shapes, multiply intensity vectors by matrices, and control wavelength response. Projected illumination changes the refractive-index landscape, while physics-aware training combines measured outputs with gradients from a calibrated beam-propagation model. Demonstrated functions include four- and eight-mode permutations, broadband operation at 1500–1600 nm, and separation of 1310 nm and 1550 nm. The present index contrast of about 1.7 × 10⁻³ limits device compactness and trans

✳ A programmable photonics experiment

What if an optical chip could change its job—just by changing the light shining on it?

One slab. No etched routes. A projected green pattern reshapes the landscape that an infrared wave travels through. Draw your own and see where the signal goes.

Project a pattern. Green light makes a covering layer more electrically conductive.

Change the index. A voltage shifts the lithium niobate’s refractive index: how light advances through it.

Let waves interfere. The new landscape changes the signal’s phase, reshaping where its energy emerges.

The virtual optical workbench

Your pattern. The wave’s response.

Paint directly on the green projection. The adjacent view solves how an infrared wave propagates through the index you create. Try a guide, a splitter, or a pattern learned on the real chip.

Where did the light go?

Power arriving in four equal output regions.

A smooth illuminated track guides the beam toward output A.

Try this: switch to a blank slab, paint a slanted band, then turn programming strength down to zero. The green pattern stays; its optical effect disappears.

This is a scalar, paraxial beam-propagation model based on Eq. (1) in the paper. It advances a complex optical field through the chip with a split-step Fourier method: transverse diffraction alternates with a phase shift caused by the local refractive index. The intensity is |E|².

The teaching window is 600 μm wide and 10.74 mm long (12.12 mm for the wavelength-sorting pattern), with the source calibration’s background index of 1.94. Illumination is blurred and mapped through a saturating response to a maximum index increase of 0.0017. Absorbing edges suppress artificial reflections. Output percentages divide the surviving signal among four equal regions; they are simulated intensity fractions, not the paper’s mode-overlap measurements.

The real programmable region spans about 8 × 9 mm. Source patterns are cropped to the central 600 μm and resampled while preserving physical feature size. Their outputs in this reduced model are illustrative. The authors’ calibrated model also learns background distortions and an illumination correction network, which are not included here. The optional inverse-design search uses adjoint gradients in this simulation to increase the input power reaching a selected output region. It runs up to 24 projected gradient updates, starting from the current drawing. The paper’s optimization instead includes measured outputs from the real chip. The real chip settles after a pattern change in about 1.5 seconds. The browser shows the new steady-state solution immediately.

How the researchers find a pattern

Start with the destination. Learn the landscape.

You have been designing forward: paint a pattern and watch the result. The paper reverses the question. Given a desired output, which illumination pattern makes the physical chip deliver it?

Choose the job

Specify which input should reach which output—or choose a desired mixture of signals.

Ask the real chip

Project the current pattern. Send in the laser. Measure the emerging intensity.

Update the pattern

A calibrated digital wave model supplies gradients: how to adjust the illumination toward the target.

Measure → update → repeat. The functions generally converged in about 100 iterations over 10–20 minutes. The paper’s four-mode routing example took about 7 minutes.

What happened on the physical chip

One chip. Several different jobs.

The researchers reprogrammed the same slab for routing, separating differently shaped waves, mixing signals, and sending different wavelengths to different outputs.

Measured · 1550 nm

Reassign the routes.

Four Gaussian-shaped input beams were sent to four chosen output positions. Reprogramming the illumination changed this permutation without fabricating a new chip.

Maximum crosstalk for the four-mode permutation. Unwanted channels were about 500 times weaker in the reported measurement.

An eight-mode spatial permutation was also demonstrated, with maximum crosstalk of −13 dB. Crosstalk describes unwanted signal leaking into other channels.

Why a rewritable slab matters

Design becomes an experiment you can revise.

Conventional inverse design often ends with a fixed, fabricated structure. Here, the chip stays programmable: the experiment can adjust to the physical device and then reuse it for another optical function. Fabrication and calibration become resources shared across many designs.

Continue into the research

Yiqi Zhao, Martin M. Stein, Fan O. Wu, Ryotatsu Yanagimoto, Benjamin A. Ash, Mandar M. Sohoni, Logan G. Wright, Tatsuhiro Onodera & Peter L. McMahon.

arXiv preprint · 2026

Explore the source material Experimental datasets, trained patterns & figure code ↗

Trained illumination grids: Zhao et al., Zenodo 22293276. Cropped to the central 600 μm, resampled and quantized. Data and source figures: CC BY 4.0. Figures are cropped and resized; scientific annotations unchanged.

Graphical abstract: AI-generated conceptual illustration informed by the paper. The live workbench is an educational wave model.