The paper asks whether ferroelectric barium titanate (BTO) phase shifters can remove the static-power and thermal-crosstalk barrier that limits scaling of heater-based programmable photonic circuits. The authors fabricated a hybrid silicon–BTO hexagonal waveguide mesh with 58 programmable unit cells, 116 actuators and 34 optical ports on 2×10 mm². Measurements on individual unit cells show ~82 ns rise/fall times, 560 nW per π static consumption in volatile (Pockels) mode, and non-volatile operation in which 300 ns pulse trains set ferroelectric domain configurations that persist with the bias

Non-volatile barium titanate photonic gate array · Nature Photonics 2026

A programmable photonic chip steers light through a honeycomb of tiny interferometers. Usually each one is held in place by a heater that must stay switched on. Catalá-Lahoz and colleagues built the switches from barium titanate, a ferroelectric crystal whose domains remember how they were set. Program the two simulated chips below, then cut their power.

Thermo-optic heaters

Barium titanate (this paper)

Simulated chips with the paper’s 58-cell layout and measured 1.48 dB loss per cell. Each cell is an ideal tunable coupler; unpowered cells fall back to random fabrication-set couplings. Non-volatile retention was measured on single cells; the paper’s full-mesh demonstrations used volatile driving. Heat glow is illustrative.

The work in brief

A gate array for light, and the switches that hold it back

Electronic FPGAs are chips that can be rewired by software after they are made. Programmable photonics tries to do the same with light, for 6G links, data centres, artificial intelligence and quantum computing. This paper tackles the part that stops such chips from growing: the switches.

1 · The idea A honeycomb of identical cells The chip is a hexagonal mesh of waveguides. Every edge of the honeycomb is a programmable 2×2 cell, a Mach–Zehnder interferometer: light splits into two arms and meets again, and a phase shift on the arms decides which way it leaves. Set all the cells and the same chip becomes a filter, a switch, a router or a splitter.

2 · The problem Heaters must stay on Most meshes shift phase with tiny heaters. A heater draws about 1–10 mW per π for as long as it holds its setting, needs 2–100 µs to change, and warms its neighbours. The authors expect future chips to need 10⁴ actuators or more, so this standing power grows with every switch added.

3 · This paper Switches made of barium titanate Barium titanate (BaTiO₃) has one of the strongest electro-optic (Pockels) effects known. A steady voltage shifts the phase in about 80 ns while holding it costs only 560 nW per π; short pulses instead reorient its ferroelectric domains, which then hold the phase with no power at all. The team built 116 of these actuators into a 58-cell mesh on a 2×10 mm² chip.

Inside one cell

Inside one switch: domains that stay where you put them

Every cell of the mesh is a Mach–Zehnder interferometer: light splits into two arms and meets again. Each arm carries a barium titanate phase shifter between gold electrodes; here we drive the upper one. Voltage pulses flip a fraction of its ferroelectric domains; that mixture sets the phase between the arms and therefore how much light leaves each output. Switch the voltage off and the mixture stays — and so does the output.

Illustrative model, calibrated to the shapes of Fig. 2c–e of the paper; the tiles are a cartoon of the domain mixture, not a measured map.

Protocol (Supplementary Note 1)

10⁴ pulses at +5 V, 300 ns · 10⁴ pulses at −5 V

Low level High level Pulses N

4 · Bias off, measure

Voltage on the electrodes

Sixteen levels, one cell

Like Fig. 2f: negative pulses (blue) step the state up, positive pulses (orange) step it down. In the experiment each level was held 1.2 min.

Any level, any time

Like Fig. 2g: each transition starts from State 0 and lands on one of 16 levels; the peaks stay apart.

Volatile mode: sweep the DC voltage

With a steady voltage the Pockels effect shifts the phase almost instantly — the paper measured ~82 ns rise and fall times (Fig. 2b) — while the domains slowly follow the field. Sweeping up and down traces a butterfly-shaped loop.

Illustrative model: reproduces the butterfly shape of Fig. 2a, not its exact notch positions.

Methods

How the chip was built, tuned and measured

The mesh combines a standard silicon photonics wafer with a thin layer of barium titanate. Silicon guides and splits the light; barium titanate, placed only where the phase must change, does the switching. Build the layer stack one step at a time.

Illustration of the layer stack, not to scale. Layer thicknesses and the order of steps are from the paper’s Methods.

Silicon-on-insulator wafer, CEA-Leti

Planarised oxide

Barium titanate layer, Lumiphase

Taper moves light into the hybrid waveguide

Cladding, vias and electrodes

One cell, by the numbers Each cell is a balanced interferometer: two 50:50 multimode splitters with a barium titanate phase shifter on each arm. The whole cell is 950 µm long, of which 350 µm are phase shifter; the rest are splitters and bends. A transverse field across the barium titanate sets the phase between the arms. Loss per cell1.48 dB Delay per cell12.73 ps Voltage for π5 V

Writing a state Single cells were tested on structures on the same chip. Erase: a 1.3 MHz sine, amplitude-modulated by a 33.33 Hz down-ramp, is applied for 2 seconds. Initialise: 10⁴ pulses of 300 ns at +5 V (State 0) or −5 V (State 1). Write: a train of 300 ns pulses between a low and a high level. Then the voltage is switched off and the light is measured. This is the protocol you can run in “Inside one switch”.

Calibrating the whole mesh For the mesh, a program represents the chip as a graph. From one fixed input it finds the shortest path to each other port, maximises the power arriving there, and so characterises the cells along that path one by one. The settings go into a database used to set any coupling later. A bank of power meters watches every output. Spectraaround 1560 nm, 1 pm steps Grating couplers work best1565 ± 15 nm

How the mesh was driven: the filters, switch, router and splitter were measured with steady (volatile) voltages, because the lab setup had a single pulse source. Non-volatile writing was demonstrated on individual cells.

Why it scales

Holding a circuit should cost nothing

A heater burns power for as long as it holds a phase, and it warms its neighbours. The barium titanate cells switch with an electric field: fast in volatile mode, and at zero holding power once their domains are written.

How fast each kind of switch changes state

Switching times from Table 1 of the paper (values for other technologies are the paper’s cited benchmarks). Log time axis.

Static power to hold every actuator at π

Assumption: every actuator holds a π shift; power per π from Table 1 (thermo-optic 25 mW and 1.3 mW; MEMS < 1 µW; phase-change 0; barium titanate 560 nW volatile, 0 non-volatile). 116 is the number of actuators on this chip.

“a large-scale mesh comprising thousands of unit cells would consume tens of watts”

One mesh, many circuits

One chip, many circuits

The same 58 cells can be set into very different circuits. Pick one: the simulated mesh shows where the light goes at the probe laser’s frequency, and the chart shows the simulated response.

On the real chip the team also built a two-ring RAMZI filter (free spectral range 13.14 GHz) and a 4×4 unitary switch whose normalised output powers exceeded 0.98 — those are measured results, not simulated here.

From the paper

The real chip

The interactive scenes above are models built from the paper’s numbers. These are the authors’ own figures: the fabricated chip and its measurements.

Key numbers

58 programmable cells

116 actuators

34 optical ports

2×10 mm² footprint

~82 ns rise and fall

560 nW/π volatile static power

0 W hold, non-volatile

16 programmable levels

44.8 fJ per switch

1.48 dB loss per cell

Conclusions

Size no longer has to cost standing power

The authors’ central point is that their chip separates how complex a programmable photonic circuit is from how much power it needs to stay programmed. Here is what the paper shows, and what is still open.

Built and measured A first non-volatile gate array for light To the authors’ knowledge, this is the first fully reconfigurable, general-purpose hexagonal photonic gate array driven by ferroelectric actuators: 58 cells and 116 barium titanate phase shifters on one chip.

Measured on single cells Set and forget Driven with a steady voltage, a phase shifter needs only 560 nW per π. Written with pulses, a cell keeps its state with the voltage off, in 16 reproducible levels. That is why the authors expect future processors to scale to thousands of actuators without a matching rise in standing power.

Measured on single cells Fast, and no heat between neighbours Switching takes about 80 ns, three to four orders of magnitude faster than thermo-optic heaters. Because the actuators work by electric field, they produce no waste heat, removing the crosstalk between neighbouring cells that limits dense heater meshes.

Measured on the mesh, steady drive One chip, many circuits The same mesh was programmed as delay-line and ring filters, a two-ring RAMZI filter, a 4×4 switch with normalised outputs above 0.98, a four-way router and a 1×8 splitter.

What comes next, according to the authors

Pulses for every electrodeThe full-mesh demonstrations used volatile driving: their setup had a single pulse source. Controlling the whole mesh in non-volatile mode needs an integrated electronic switching matrix that distributes programming pulses to every electrode.

Less loss per cellEach cell loses 1.48 dB, mostly in the tapers between silicon and silicon–barium titanate waveguides. Barium titanate itself is transparent at telecom wavelengths, so the authors see a path to below 0.5 dB.

A steady temperatureThe actuators make no heat of their own, but the chip still feels the room: a 5 °C change shifted a cell’s null by 0.24–0.45 V, so the chip was held at 25 °C with a thermoelectric cooler.

“the results presented here provide a clear blueprint for sustainable, high-speed photonic computing.”

Credits

Catalá-Lahoz, C., Rausell-Campo, J. R., Pérez-López, D., Güniat, L., Convertino, C., Eltes, F., Fompeyrine, J., Mesaritakis, C., Bogris, A., Wilmart, Q., Faugier-Tovar, J., Charbonnier, B. & Capmany, J. High-speed non-volatile barium titanate field-programmable photonic gate array. Nature Photonics (2026).

Published article: doi.org/10.1038/s41566-026-01934-y

Preprint read for this page: Research Square, doi.org/10.21203/rs.3.rs-8551853/v1

The preprint is licensed CC BY 4.0; its figures are reused above with credit.

Simulations on this page are models built from values reported in the paper.