High-speed non-volatile barium titanate field-programmable photonic gate array

Abstract

Programmable integrated photonics aims to replicate the versatility of field-programmable gate arrays in the optical domain. However, scaling these systems has been prevented by the high power consumption and thermal crosstalk of conventional volatile phase shifters. Here we introduce a non-volatile field-programmable photonic gate array, implemented on a hybrid silicon–barium titanate platform, which overcomes the power scaling limitations of previous technologies. Unlike traditional thermo-optic devices that require constant power to maintain a state, our device utilizes ferroelectric domain switching to provide non-volatile memory, allowing optical circuits to be programmed and retained without any holding power or electrical bias. The hexagonal waveguide mesh integrates 58 programmable unit cells and 116 actuators, achieving nanosecond-scale switching speeds of 80 ns while reducing static power consumption to negligible levels (560 nW per π phase shift). To validate this platform, we configured the mesh to perform diverse signal processing functions, including tunable filtering, 4 × 4 linear unitary transformations and optical routing. This work establishes non-volatile ferroelectric silicon photonics as a scalable, heat-free platform essential for the next generation of energy-efficient photonic computing.

Barium titanate ferroelectric is used in photonic phase shifters arranged in a hexagonal waveguide mesh, forming a programmable photonic array. The device achieves sub-100-ns switching speeds, while each individual gate retains its state without continuous powering, enhancing overall energy efficiency.

Main

Programmable integrated photonics (PIP)¹ promises to bring the versatility of reconfigurable electronics into photonic circuits², enabling flexible, low-power and cost-effective platforms for a myriad of applications, including 6 G communications³, data centre interconnections⁴, the Internet of Things⁵, high-tech medicine⁶, artificial intelligence^7,8, photonic processing^9–11 and photonic^12–14, neuromorphic^15–19 and quantum computing^20–23. To meet the ever-growing requirements of these application areas, programmable photonic circuits will need to scale to integrate 10⁴ actuators and beyond in the coming years, which will require addressing several challenges simultaneously^24,25 that so far have remained elusive to the current state-of-the-art silicon photonic thermo-optic (TO) and micro-electromechanical (MEMS) phase shifting technologies^26–29. In essence, programmable unit cells (PUCs) with minimal footprint (below 100 μm), loss (below 0.15 dB), power consumption (below 1 mW/π), and nanosecond-scale tuning speeds will be required to implement complex multiport interferometers and waveguide meshes. TO-based phase shifters incur moderate static power consumption (1–10 mW/π), while those based on mechanical effects require a constant bias when the device is in the standby state. Hence, the static energy consumption can easily scale with the number of actuators, leading to an important increase in the power budget of the programmable processor.

Non-volatile materials such as phase-change compounds, ferroelectrics and memristive oxides have been proposed as alternatives for the implementation of integrated phase shifters, as their optical state can be maintained without any power consumption or bias. Among the different options, phase-change compounds, ferroelectrics and memristive oxides have been researched during the past years as alternatives, offering reversible switching with ultralow power and potentially fast response³⁰. While each of these material platforms has shown promising advances, their respective limitations have restricted experimental demonstrations to only modestly sized circuits, far below the scale required for fully programmable, general-purpose photonic processors.

Phase-change materials (PCMs), such as Ge₂Sb₂Se₅ (GST), are chalcogenides that can be electrically switched between bistable microstructural states of ordered crystalline and disordered amorphous states, providing a phase shift. Integration of PCMs in a silicon waveguide has been reported, where the phase shifting is achieved via the TO effect^31–33, and the technology shows fast switching speeds (0.1 μs, 1 mW) for the set-up of the amorphization phase, as well as potentially low footprints (L_π = 10–50 μm). However, the switching speed for the transition to the crystallization phase is on the order of milliseconds, leading to power consumption in the tens of milliwatts. Another disadvantage stems from the high losses (>0.33 dB) arising from the proximity of the PCMs to the optical mode in the silicon waveguide. These limitations hamper the scalability of photonic processors based on this technology.

Memristors typically consist of a conductor–insulator–conductor trilayer³⁴. The conductor layers can be made of metals or semiconductors, whereas the insulator is a dielectric material. The application of enough voltage forms a conduction filament that switches the memristor from an initial high-resistance state into a low-resistance state. The memristor can be returned to the high-resistance state by reversing the bias polarity, which breaks the filaments. The combination of this mechanism with the free carrier dispersion effect is then exploited to provide a controllable phase shift. A recent experiment³⁵ reported a silicon photonic phase shifter based on the heterogeneous integration of GaAs/Al₂O₃/Si memristor in a 10-μm microring resonator featuring low insertion loss (<0.05 dB). While results in terms of switching speed (<1 ns), switching powers (150–360 μW) and insertion losses (0.27 dB) are promising, the compactness of phase shifters is limited by the large value of L_π (350 μm).

Ferroelectrics and, in particular, barium titanate (BTO), have recently emerged as a promising material for efficient phase shifters. BTO features one of the highest Pockels coefficients reported (r₄₂ of ~923 pm V⁻¹), and its potential for large-scale integration with silicon photonic circuits has recently been demonstrated through a combination of epitaxial growth and direct wafer bonding³⁶. BTO also allows the non-volatile storage of optical information by switching non-volatile ferroelectric domains through the application of an electric field. The first demonstration resulted in multilevel storage over eight levels³⁷, although initial devices developed in this platform showed large L_π (1 mm) and moderate switching speeds (on the order of microseconds) for non-volatile operation. This platform has a clear potential for improvement to obtain compact footprints, ultralow insertion loss, submilliwatt switching, nanosecond reconfigurability and non-volatility.

Here, we present a non-volatile field-programmable photonic gate array (FPPGA) based on BTO (BaTiO₃) actuators, establishing a scalable, heat-free alternative to traditional architectures. The FPPGA integrates 116 actuators or phase shifters in a flattened hexagonal waveguide mesh configuration and achieves nanosecond-scale switching with submicrowatt power consumption per actuator, two orders of magnitude below the state of the art. We experimentally demonstrate several key signal processing applications, including reconfigurable filters, unitary linear transformations, optical routing and splitting. Moreover, we show the capability of the individual PUCs to set and retain a specific state after the external bias is turned off, owing to their non-volatile behaviour. These characteristics mark a significant step forward towards large-scale, power-efficient programmable photonic systems.

Results

Programmable photonic integrated circuits use meshes of interconnected waveguides, where reconfigurable 2 × 2 elements known as PUCs or gates define and control the amplitude and phase of the optical signals that propagate through the network. Following the philosophy of electronic field-programmable gate arrays, an FPPGA can be reprogrammed to realize a variety of photonic circuits and linear multiport transformations by configuring its PUCs and selecting the appropriate ports^38,39. A common implementation of a PUC is based on a balanced Mach–Zehnder interferometer (MZI) consisting of two 50:50 multimode interferometers and an electrically biased phase shifter on each arm. Using a PUC with non-volatile actuators (NV-PUC) as the basic building block eliminates static power consumption, as each state can be set once and retained without the need to apply a continuous bias.

Figure 1 illustrates the design principle and physical realization of the non-volatile FPPGA. The schematic in Fig. 1a depicts the FPPGA based on a hexagonal waveguide mesh topology, enabling software-defined programming of specific photonic circuits. The underlying mechanism of this reconfigurability, detailed in Fig. 1b, stems from the implementation of PUCs using hybrid silicon nitride (SiN)–BTO phase shifters on a silicon platform. By applying voltage pulses, the ferroelectric domains within the BTO layer are reoriented, altering the effective refractive index and providing non-volatile phase shifting. The realization of this design is presented in Fig. 1c, which shows the fabricated and wire-bonded chip on a printed circuit board. Optical microscope images provide further detail on the device specifics: Fig. 1d displays a close-up of a single PUC, while Fig. 1e illustrates the large-scale integration of these cells into the interconnected hexagonal mesh with a flattened design⁴⁰. The footprint of the fabricated circuit is 2 × 10 mm² and includes 58 PUCs, 116 BTO phase shifters and 34 optical input/output ports.

Fig. 1. Design principle and physical realization of the non-volatile FPPGA.

a, Artistic rendering of the programmable photonic chip based on a hexagonal waveguide mesh topology. b, Illustration of the non-volatile phase-shifting mechanism on the hybrid Si–BTO platform. Voltage pulses reorient the ferroelectric domains (blue), modifying the effective refractive index without static power consumption. c, Photograph of the fabricated and wire-bonded chip mounted on a printed circuit board. d, Optical microscope image showing a detailed view of a single NV-PUC. e, Zoomed-out microscope image of the fabricated hexagonal mesh, displaying the large-scale integration of unit cells and interconnected BTO actuators. E-field, electric field.

Non-volatile PUC

The NV-PUC serves as the basic reconfigurable element of the mesh circuit. Phase shifters are implemented using SiN waveguides on top of a ferroelectric BTO film, which are integrated with the silicon routing waveguides, following the fabrication process described in Methods. Programming of the unit cell relies on electrically controlling the BTO actuators applying a transverse electric field. By adjusting the differential phase between the interferometer arms, the NV-PUC can switch between its bar and cross states or provide analogue coupling tuning.

The integration of BTO actuators enables the NV-PUC to operate in both volatile and non-volatile modes depending on the characteristics of the applied voltage. When the NV-PUC is operated in volatile mode, a d.c. voltage modifies the refractive index of the BTO through the Pockels effect. When operated in non-volatile regime, the change in the configuration of its ferroelectric domains, produces a linear and reversible phase change. The combination of both effects creates a butterfly-shaped hysteresis loop in the effective refractive index. The history-dependent optical response can be experimentally observed in the interferometric curves of the PUCs. When the voltage sweep begins at 0 V, where only one domain configuration is stable, we observe the response of a standard MZI. However, when the sweep starts from a previously biased state, either positive or negative, the BTO film presents a range of possible refractive index configurations. As the voltage is swept, part of the ferroelectric domain population begins to reorient. When the direction of the applied electric field is reversed, the hysteresis curve of the refractive index is transferred to the interferometric response creating a similar butterfly shape. Figure 2a shows the MZI response when we apply a d.c. voltage, sweeping from 0 to 12 V and from +12 (−12) to −12 (+12) V. The interplay between the instantaneous Pockels effect and the slower domain reorientation produces the characteristic butterfly-shaped hysteresis observed when sweeping from positive to negative voltages or vice versa.

Fig. 2. Characterization of the NV-PUC.

a, Static transmission response of the MZI showing the characteristic butterfly-shaped hysteresis loop resulting from ferroelectric domain reorientation under d.c. voltage sweeps. b, Volatile electro-optic response demonstrating symmetric rise and fall times of ~82 ns, driven by the Pockels effect. The shaded areas in the main graph highlight the time windows for the rise and fall transitions. The two plots on the right are zoomed-in views of these shaded areas, detailing the 10% to 90% rise and fall times of approximately 82 ns. c,d, Tuning of non-volatile states starting from state 0 (c) and state 1 (d) by applying pulse trains with varying LL and HL voltages. Inset: a diagram of the electrical pulse sequence, pointing out the HL and LL voltages and the corresponding starting state. e, State saturation behaviour as a function of the number of applied pulses (N), showing stability beyond 10² pulses. Inset: a diagram of the electrical pulse sequence, indicating the total number of pulses applied and the starting state. f, Multilevel linear programming demonstrating a stable ‘staircase’ response with 16 distinct phase levels. g, State distribution histograms obtained from 600 random transitions, validating the accuracy and reproducibility of the 16 programmable states.

Building on this static characterization, the BTO phase shifters result in an ultralow static power consumption of just 560 nW/π corresponding to a V_π = 5 V. This represents an outstanding improvement for the scalability of programmable photonics, where standard TO heaters typically require 1.3–20 mW per phase shift, leading to significant thermal management challenges in large-scale circuits. Another significant performance advantage of the BTO platform is observed in its temporal dynamics. Figure 2b illustrates the optical response of the MZI driven by 2-MHz square pulses with a peak-to-peak voltage (V_pp) of 4.8 V and 1.2 V of bias. This specific operating point was chosen along the initial 0–12 V static curve (Fig. 2a) because it aligns with the steepest region of the interferometric response (quadrature point), maximizing modulation depth while ensuring the device operates strictly within the ultrafast, volatile Pockels regime without triggering slower ferroelectric domain reorientations. The device demonstrates highly symmetric switching with rise and fall times around 82 ns (measured from 10% to 90%). This nanosecond-scale response is intrinsic to the electro-optic nature of the Pockels effect, offering a speed advantage of three to four orders of magnitude compared with conventional TO phase shifters (2–100 µs), which are physically limited by slow thermal diffusion and are also sensitive to thermal crosstalk^3,26. The fabricated NV-PUC exhibits an insertion loss of 1.48 dB and a basic delay unit of 12.73 ps, with a total length of 950 μm, of which 350 μm corresponds to the phase shifters and the remainder to the multimode interferometers and bends.

In non-volatile operation, the NV-PUC takes advantage of the fact that the ferroelectric domain configuration in BTO can be permanently modified, which changes the effective Pockels coefficient and therefore, the refractive index at a given applied bias. For any non-zero voltage, the material can support multiple stable refractive index states depending on how the domains are distributed. To program these states, we apply short electrical pulses (on the order of nanoseconds to microseconds) at the desired bias level and voltage. These pulses are shorter than the characteristic domain switching time of the BTO film, so instead of driving the material into a fully poled state, they produce controlled mixed-polarization domain configurations. Adjusting the bias and amplitude voltages (or low and high voltage amplitudes/levels) at which the pulses are delivered allows access to different domain mixtures, each corresponding to a distinct refractive index state that remains stable after the pulse sequence is removed.

To overcome the path dependence inherent to ferroelectric domains, we implemented a reset protocol to ensure identical initial conditions for every measurement. We established two different boundary states, designated as state ‘0’ and state ‘1’, which frame the accessible operation range. These states are prepared via a two-stage sequence: first, the previous domain history is erased using a decaying alternating electric field; second, the polarization is saturated using a train of 10⁴ initialization pulses at +5 V (for state 0) or −5 V (for state 1), with a pulse duration of 300 ns. Without this preconditioning, the device would settle into a random intermediate state determined by its prior usage. Once the device is initialized, the ferroelectric polarization saturates, meaning subsequent pulses of the same polarity induce no further refractive index modification. Consequently, tuning is achieved by applying 10⁴ pulses of the opposite polarity while maintaining a bias matching the initialization sign (see Supplementary Note 1 for details on the procedure for erasing and setting the states). Figure 2c characterizes the tuning dynamics starting from state 0. In this regime, while holding a positive high level (HL) bias between 1 V and 5 V, we apply a pulse train that sweeps the low level (LL) voltage from 0 V down to −5 V. The complementary behaviour is shown in Fig. 2d, which illustrates tuning from state 1. Here, the device is tuned by sweeping the HL voltage from 0 V to 5 V against distinct negative LL baselines. Under this optimized driving scheme, a phase shifter with an active length of just 350 µm achieves a full non-volatile π phase shift with an extinction ratio of 12 dB.

We also investigated the device’s non-volatile behaviour as a function of the number of applied pulses (N). To do this, we applied consecutive 300-ns pulses of constant amplitude while varying N. Figure 2e shows that for intermediate states (between 0 and 1), the achieved state saturates beyond N = 10² pulses. This means that, to establish a state in the device, the total required pulse train time would be 60 µs.

The ability to retain multiple intermediate states allows this device to act as a tunable building block for programmable photonics. To demonstrate this accurate control, we initialized the NV-PUC to state 0 and applied a sequence of specific pulses. The amplitudes of these pulses were precalculated using the data from Fig. 2c,d to ensure that the optical steps were linear and equally spaced. We programmed 16 distinct states, holding each one for 1.2 min to verify its stability before moving to the next. The sequence was then reversed using pulses of the opposite polarity. As shown in Fig. 2f, the device produced a stable, linear ‘staircase’ response, successfully resolving 16 equally spaced states in both directions. To ensure that any state can be set reliably regardless of the previous value, we performed a random switching test involving 600 transitions. In this experiment, we randomly selected target states from the 16 available levels. To guarantee accuracy, we reset the device to state 0 before every new setting. The resulting performance is shown in Fig. 2g. The histogram displays 16 clearly separated peaks, proving that the device creates consistent and reproducible states.

It is worth noting that the global downward slope visible during the hold periods in the descending curve (Fig. 2f) is a systematic measurement deviation caused by slow mechanical fibre coupling drift over the 40-min measurement. Because our test structure lacked an adjacent passive reference waveguide to continuously monitor and normalize out these input power fluctuations, this uncompensated insertion loss manifests as a global slope and does not represent a relaxation of the non-volatile states. Furthermore, the broader state distribution observed at higher transmission levels (closer to state 1) in both the staircase measurements and the corresponding histogram (Fig. 2g) is not a result of thermal or state instability. Rather, it is a direct consequence of the flattened nonlinear transfer function of the MZI near maximum transmission. Because the required voltage difference between adjacent states becomes small near state 1, these higher levels are more sensitive to standard experimental noise.

Hexagonal mesh programmable circuit

We initiated the study with the static characterization of the hexagonal mesh by applying d.c. voltage sweeps to the constituent PUCs to generate their optical response curves. The sequence of characterization was defined using a custom graph-based algorithm (see Supplementary Note 2 for set-up and characterization algorithm details). The resulting parameters were subsequently stored in a database, allowing us to precisely configure the specific coupling ratios required for later experiments. Following system calibration, we evaluated the spectral operating range of the input and output grating couplers. Our measurements indicate that the device achieves peak efficiency in the 1,565 ± 15 nm window.

The hexagonal waveguide mesh can be configured to emulate the hardware architecture of a diverse array of optical circuits, functioning as a general-purpose photonic processor. The mesh supports, among other functions, traditional feedforward and feedback finite impulse response (FIR) and infinite impulse response (IIR) filters, as well as complex multiple-input/multiple-output linear optical transformers.

The unbalanced MZI (UMZI) works as a periodic notch filter. These devices are fundamental components in the construction of lattice filters and FIR transversal filters. By manipulating the PUC settings, we successfully synthesized UMZI structures with path length differences equivalent to two and four times the standard PUC length. The experimental data for these configurations are displayed in Fig. 3a. These path imbalances resulted in free spectral ranges (FSRs) of 43.65 GHz and 19.02 GHz, respectively.

Fig. 3. Implementation of reconfigurable optical filters on the hexagonal mesh.

a, Synthesis of UMZIs with path length differences of 2L_PUC and 4L_PUC, showing the measured normalized transmission spectra and phase response. b, Optical ring resonators configured with cavity perimeters of 6L_PUC and 10L_PUC. c, Demonstration of IIR filter tunability achieved by independent control of the PUC phase shifters to sweep the resonance peak. The colour gradients in the graphs represent the gradual shifting of the resonance peak as different voltage levels are applied to the phase shifter to tune the filter, with the darkest shade corresponding to 0 V and the lightest shade to 7 V. d, Configuration and measured spectral response of a complex two-order RAMZI filter. BS, bar state; CS, cross state; RAMZI, ring-assisted MZI.

We also implemented ring cavities, which operate as periodic filters. Using these structures, we can implement all-pole IIR notch filters, and combining them with other structures, they can realize both IIR notch and hybrid FIR + IIR bandpass filters. These resonators are essential for assembling complex architectures, such as coupled resonator optical waveguides or ring-assisted MZIs (RAMZI). Through precise tuning of the unit cells, we realized single optical ring resonators with cavity perimeters of six and ten PUC lengths. Figure 3b illustrates the measurements for these cavities, which exhibited FSR values of 13.27 GHz and 7.95 GHz, respectively. The synthesized 6-PUC ring resonator show a measured quality factor of 3.45 × 10⁴ and a round-trip amplitude transmission of α = 0.36 (corresponding to ~8.9 dB loss). Similarly, the larger 10-MZI ring exhibited a quality factor $Q_f=5.196\times {10}^4$ and an α = 0.18 (~14.9 dB loss). In both cases, the round-trip loss is completely dominated by the accumulated insertion loss of the MZIs forming the loop. Figure 3c demonstrates the tunability of the IIR filter. This is achieved by tuning both actuators of the PUC acting as phase shifters, which allows independent control of the coupling and output phase. By varying the voltage injected into both phase shifters, the resonance peak can be swept across a complete spectral period.

Leveraging this flexibility, we built more complex filters, such as the two-ring RAMZI filter depicted in Fig. 3d. The measurement results for this structure display both the bar and cross output ports, which are complementary and feature an FSR of 13.14 GHz (see Supplementary Note 3 for additional filter measurements). However, the experimental cross-port spectrum is degraded. This happens because the high round-trip loss in the 6-PUC ring weakens the resonance, preventing the ideal phase and amplitude interference needed for a perfect rectangular shape.

Beyond filtering applications, we leveraged the hexagonal waveguide mesh to implement versatile multiple-input/multiple-output linear optical transformations. The mesh’s programmability allows it to emulate various connectivity schemes, ranging from unitary matrix operations to complex routing and broadcasting networks.

We programmed the hexagonal mesh to implement a rectangular unitary architecture, creating a 4 × 4 programmable linear optical transformer. By defining six active tunable couplers (S1–S6), as shown in Fig. 4a, we established a reconfigurable switch capable of executing various unitary operations. We synthesized three distinct transfer matrices (M) to demonstrate the device’s switching capabilities. The corresponding bar graphs display the normalized output power for each configuration. The device exhibited exceptional fidelity, with normalized output powers consistently exceeding 0.98 and a high transmission ratio between desired and undesired connections (~25 dB).

Fig. 4. Linear transformations and routing on the photonic processor.

a, A 4 × 4 programmable linear optical transformer. Left: mesh schematic defining six active tunable couplers (S1–S6). Right: target unitary transfer matrices and measured normalized output powers for three different configurations, demonstrating high fidelity. b, Configuration of the mesh as an optical router, allowing the simultaneous routing of four independent signals without interference. The bar chart displays the normalized output power for each target port. c, Demonstration of a 1 × 8 optical splitter implemented via a cascading tree structure. The graph shows the superimposed transmission spectra of the eight output ports (o21–o28), indicating uniform power distribution and broadband operation.

To validate the platform’s interconnect capabilities, we programmed the mesh as a flexible optical router. Figure 4b presents a multipath scenario where four independent optical signals are routed simultaneously across the chip lattice without interference. We established four distinct trajectories: port 2–15, 7–27, 10–24 and 32–19. The measured normalized output powers for these paths were 0.892, 0.91, 0.88 and 0.954, respectively. These high transmission values indicate that the mesh can support multiple intersecting signal flows, dynamically creating optical links between arbitrary points on the perimeter.

Finally, we demonstrated the mesh’s capacity for broadcasting applications by configuring it as a 1 × 8 optical splitter. As shown in the schematic in Fig. 4c, a single input signal is distributed through a cascading tree structure to eight distinct output ports (o21–o28). The spectral response measurement shows a highly uniform power distribution across all eight channels in the 1,560–1,570 nm wavelength range. The overlapping traces indicate balanced splitting ratios and broadband operation, proving the mesh is suitable for one-to-many signal distribution tasks.

Discussion

We have demonstrated an extended programmable photonic processor capable of non-volatile operation, providing zero-static power advantages previously unavailable to address the power wall that currently constrains the scaling of PIP. By integrating ferroelectric BTO actuators into a silicon photonic hexagonal waveguide mesh, we realized a platform that combines the zero-static power advantages of non-volatile memory with the high-speed performance required for dynamic signal processing. While non-volatile functionalities were recently explored in reduced scale demonstrations using PCMs⁴¹, our work advances the field by presenting a fully reconfigurable, general-purpose hexagonal FPPGA based on ferroelectric actuators.

The central achievement of this work is the decoupling of circuit complexity from energy consumption. In conventional programmable circuits based on TO phase shifters, maintaining a specific circuit configuration requires continuous electrical biasing. With typical consumption rates of 1–10 mW per actuator, a large-scale mesh comprising thousands of unit cells would consume tens of watts, necessitating complex thermal management solutions to prevent crosstalk and instability. By contrast, our BTO-based architecture consumes very low power (560 nW/π per actuator) when driven in a volatile mode. Furthermore, if it is driven as a non-volatile, once programmed, the state is retained by the ferroelectric domain orientation without external bias. This ‘set-and-forget’ capability suggests that future processors could scale to thousands of actuators without a corresponding increase in the static power budget.

To contextualize the performance of our hybrid SiN–BTO phase shifters integrated on silicon, Table 1 benchmarks our device against the leading phase-shifting technologies currently used in programmable photonics.

Table 1.

Device-level performance benchmark of phase-shifting mechanisms for large-scale photonic circuits

Technology	Physical mechanism	Volatility	Static power consumption	Switching speed	Switching energy	Insertion loss	Actuator length	Thermal crosstalk
TO (without trenches)26	Temperature dependence	Volatile	~25 mW/π	2.69 µs	67.25 nJ	0.23 dB	61.6 µm	High
TO (with trenches)3	Temperature dependence	Volatile	1.3 mW/π	100 µs	130 nJ	0.48 dB	~220 µm	Low
MEMS28,43	Electrostatic actuation	Volatile	<1 µW	1.54 µs	0.2 nJ	0.33 dB	50 µm	Zero
PCM: GST44,45	Phase transition (amorphous/crystalline)	Non-volatile	Zero	100 ns and 0.6 ms	180 pJ and 17 nJ	1.6 dB	5 μm	High
BTO (this work)	Ferroelectric	Volatile	~560 nW/π	~80 ns	44.8 fJ	1.48 dB	350 µm	Zero
		Non-volatile	Zero	~60 µs

As shown in Table 1, our device bridges the performance gap between existing switching technologies. PCMs typically rely on transitions between amorphous and crystalline states, which can introduce significant optical absorption (>0.33 dB per element) and are limited by slower crystallization speeds (milliseconds). By contrast, our BTO platform leverages the Pockels effect, enabling switching speeds of 80 ns, several orders of magnitude faster than both PCMs and TO heaters.

However, regarding optical loss, our current prototype exhibits an insertion loss of 1.48 dB per PUC. While this is currently higher than the intrinsic loss of optimized PCM or MEMS devices, it is crucial to distinguish the source of this attenuation. Unlike PCMs, where loss is often intrinsic to the material state, the losses in our device are primarily due to the hybrid integration process. Specifically, the mode transitions (tapers) between silicon and SiN–BTO waveguides. Because BTO itself is inherently transparent at telecom wavelengths, there is a clear engineering path to reduce these losses to <0.5 dB in future generations by optimizing the taper geometry and buffer layer thickness.

Despite the current insertion loss, the experimental demonstrations, ranging from tunable FIR/IIR filters to rectangular unitary architectures, validate the optical robustness of the platform. However, it is important to clarify the operational mode used in these system-level experiments. While the individual unit cells have been proven to support non-volatile memory storage, the complex mesh configurations presented here were characterized using volatile actuation. Controlling the entire mesh in non-volatile mode requires an electronic interface capable of distributing specific programming pulses to all electrodes independently. In our current experimental setup, which relies on a single pulse source, such addressing is limited to manual operation. Consequently, the development of an integrated electronic switching matrix to automate pulse distribution is a critical next step to fully unlock the platform’s zero-static power potential in large-scale circuits.

Finally, the reliance on the Pockels effect fundamentally improves the thermal management strategy compared with conventional platforms. Unlike TO phase shifters, which function by locally heating the waveguide and inherently suffer from thermal crosstalk between adjacent components, our ferroelectric actuators are electric-field driven and generate no waste heat. This effectively eliminates intradevice thermal crosstalk, a major bottleneck in dense photonic integration. Furthermore, electrical crosstalk between adjacent actuators is negligible. The 3-µm buried oxide layer (Methods) and the large spatial separation between the 950-µm unit cells prevent static field leakage or substrate conduction. Most importantly, during non-volatile programming, any transient parasitic coupling to adjacent devices remains orders of magnitude below the coercive field threshold required for ferroelectric domain switching, ensuring that neighbouring non-volatile states are completely undisturbed. However, the ferroelectric material remains sensitive to global ambient temperature fluctuations (see Supplementary Note 4 for a characterization of the device’s temperature dependence). Therefore, while the system is free from self-induced heating instabilities, precise stabilization of the chip substrate using a thermoelectric cooler (TEC) is required to maintain performance against external environmental drift. Importantly, implementing such TEC stabilization is a standard practice in commercial optical modules, which, when combined with straightforward equation-fitting calibration and the negligible material fatigue of BTO (10⁹–10¹² cycles)⁴², fully preserves the platform’s zero-static power advantage for real-world deployment.

Although the inclusion of this TEC, along with the driver circuits, introduces system-level power overheads beyond the intrinsic device-level metrics (Table 1), the platform’s scaling advantages remain robust. Because the actuators generate no localized heat, the TEC solely compensates for environmental changes, drastically reducing the power required for thermal stabilization compared with TO-based systems. Furthermore, in the non-volatile ‘set-and-forget’ regime, the higher system-level energy from pulse generation is a one-time cost rapidly amortized over long holding periods. Once configured, peripheral electronics can enter standby mode, maintaining zero-static power. Even under continuous volatile operation, a 10,000-actuator BTO mesh would consume mere milliwatts statically, avoiding the tens of watts demanded by TO arrays.

In conclusion, the realization of a non-volatile, field-programmable waveguide mesh fundamentally alters the scaling prospects of integrated photonics. By eliminating the static power consumption and thermal crosstalk inherent to TO approaches, this platform overcomes the primary bottlenecks preventing the development of large-scale optical processors. While future efforts must address insertion loss and electronic co-integration, the results presented here provide a clear blueprint for sustainable, high-speed photonic computing. This technology is poised to play a critical role in next-generation hardware, where energy efficiency and adaptability are paramount.

Methods

Fabrication

The 200-mm base wafers used for the BTO integration process were fabricated at CEA-Leti using a 220-nm silicon-on-insulator platform with a 3-µm buried oxide layer. To prepare a surface suitable for BTO bonding, we targeted a top oxide thickness of 350 nm, achieved through sequential oxide deposition and chemical–mechanical polishing steps. Final measurements confirmed a resulting thickness of approximately 365 nm (Supplementary Note 5).

Subsequently, the wafers were transferred to Lumiphase for back-end processing. First, a 200-nm-thick BTO film is transferred onto the planarized wafers through direct wafer bonding. After removing the donor wafer, the BTO layer is patterned and etched so it only remains in the functional device regions. Next, a 200-nm-thick SiN layer is deposited and patterned with high resolution to form 800-nm-wide waveguides directly on top of the BTO. SiN is chosen because its refractive index of 2.1 helps confine the optical mode, achieving an optimal 45% overlap with the active BTO layer to efficiently exploit the Pockels effect (see Supplementary Note 5 for full fabrication flow and cross-section).

For the interlayer coupling between the bottom Si passives and the top hybrid SiN–BTO phase shifters, a taper of the silicon waveguide was designed to transfer the mode. This taper uses a 150-µm-long optimized middle transition section to minimize optical insertion loss (detailed in Supplementary Note 5). Finally, the devices are completed with a three-level metallization process (M1 to M3) to provide flexible electrical routing.

Characterization and measurements

The characterization of the NV-PUC was performed using test structures located on the same chip as the hexagonal mesh. The initial experimental set-up consists of a tunable continuous wave laser (EXFO T100S-HP) followed by a polarization controller. Light is coupled into and out of the PUC’s grating couplers using two cleaved fibres mounted on 8° tilted wedges with tapered V-groove fibre holders. At the output of the PUC, the detected optical power is measured using a programmable power meter (EXFO FTBx-1750). To control the PUC actuators, RF probes with a 250-µm pitch were used to apply two different types of signal from a wavefunction generator (Teledyne T3AFG120). These signals are used to set and erase the various states of the PUC.

The procedure for setting a state begins by erasing the ferroelectric domain state. To achieve this, a sine signal of 1.3 MHz, amplitude-modulated with a down-ramp waveform (33.33 Hz), is applied for 2 s to reorganize the domains into an arbitrary distribution. Subsequently, two specific ferroelectric domain states (0 and 1) were defined by applying a sequence of 10⁴ pulses, each with a duration of 300 ns at 5 V or −5 V (depending on the state 0 or 1). This pulse combination results in a saturated state for both polling directions. After setting the initial state (0 or 1), a train of N pulses (300 ns duration) is applied with specific upper and lower amplitudes (V_HL and V_LL). Once the state is set, the voltage source is turned off, and the optical output power is measured (Supplementary Note 1). The erase signal and the pulse train for setting the states were generated using the same waveform generator, one per channel of the device. An electrical amplifier was placed at the output of the erase signal, and its output, along with the second channel (containing the pulse train), was connected to a coaxial switch (Radiall R570412000LP) used to sequentially apply the two signals.

For the mesh measurements, the device was first automatically characterized. To do this, a bank of power meters (EXFO LTB-12 and FTBx-1750) was used to monitor the optical power at all mesh output ports. This automated characterization routine relies on a graph representation that replicates the mesh topology. We selected a single fixed input port and, using the graph, determined the shortest paths to the remaining ports. By optimizing the output power for each path, we sequentially characterized the individual MZIs along that path, iterating the process for all other ports (Supplementary Note 2). Finally, a passive component tester (EXFO CT440) was used to acquire transmission spectra around 1,560 nm with a spectral resolution of 1 pm.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41566-026-01934-y.

Author contributions

J.C. and D.P.-L. conceived of the concept and supervised the overall project. J.R.R.-C. designed the layout of the hexagonal mesh. Q.W., J.F.-T. and B.C. were responsible for the silicon wafer fabrication and planarization processes. F.E., L.G., C.C. and J.F. carried out the BTO integration. C.C.-L. developed the control software, performed the characterization of the programmable mesh and carried out the experimental demonstration of the applications. C.C.-L. and J.R.R.-C. performed the measurements of the non-volatile test structures. J.C. wrote the introduction, and C.C.-L. wrote the original draft of the paper. C.M. and A.B. participated in the discussion of the results and provided comments on the paper. All authors reviewed and approved the final version.

Peer review

Peer review information

Nature Photonics thanks Hanke Feng, Karanveer Singh and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Funding

This work was supported by the European Commission through the Horizon 2020 project NEOTERIC (grant agreement no. 871330) and the Horizon Europe project PROMETHEUS (grant agreement no. 101070195). J.C. and D.P.-L. acknowledge funding from the European Research Council (ERC) under grant agreement nos. 101097092 (ANBIT), 101241773 (TRANSBIT) and 101076175 (LS-Photonics).

Data availability

The data that support the findings of this study are available from the corresponding authors upon reasonable request.

Code availability

The code that supports the plots and data analysis of this study is available from the corresponding authors upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41566-026-01934-y.