a, Schematic of the switch array. Silicon waveguides on the bottom layer are shown in grey, whereas the InGaAsP waveguides on the top layer are shown in blue (not pumped) or red (pumped). The pump pattern is depicted via the orange shading. The signal light (depicted with yellow shading) from each input port (I₁–I₈) is routed to the perpendicular direction at the pumped unit, then to the target output port (O₁–O₈). The inset shows the layer information of the waveguides in the coupling region. BOX, buried oxide. b, Working principle of a single switch unit in the ‘off’ state. In the left image, the InGaAsP waveguide is very lossy due to the absence of pump light, resulting in the system working in the PT-broken phase. Most of the light stays in the silicon layer and goes directly to the through port. The right image shows energy flow in the coupling region in the x direction. c, Working principle of a single switch unit in the ‘on’ state. The pump light provides gain to the InGaAsP waveguide on the top layer, resulting in the system operating in the PT-symmetric phase. In the left image, input light from the silicon waveguide on the bottom layer is coupled to the upper InGaAsP layer and then coupled back to the perpendicular bottom silicon layer, and finally travels to the drop port. The right image shows energy flow in the coupling region in the x direction before the crossing. In b and c, the scale bar represents the dimensions in the x direction, whereas the z direction has been stretched by a factor of 1.5 to better visualize the layer information.

In this coupled waveguide system, the non-Hermiticity (that is, g) controls the PT symmetry breaking of the system: \vert g \vert < 2 \vert \kappa \vert for the PT-symmetric phase and \vert g \vert > 2 \vert \kappa \vert for the PT-broken phase. In the absence of external pumping, InGaAsP is inherently lossy with \vert g \vert > 2 \vert \kappa \vert, placing the system in the PT-broken phase where the two waveguides are effectively decoupled. In this state, the system remains ‘off’ with the input light staying in the silicon waveguide and travelling directly to the through port (Fig. 1b). External pumping on InGaAsP turns the material loss into gain, leading to a phase transition into the PT-symmetric phase with \vert g \vert < 2 \vert \kappa \vert. In this state, the coupling between the two waveguides is ‘on’, enabling energy transfer between them (Fig. 1c). According to equation (2), the condition of \cos(qx- \theta)=0 corresponds to complete power transfer from the silicon waveguide to the InGaAsP waveguide, defining the coupling length L needed between two waveguides: L=\frac{\theta +\frac{\uppi }{2}\,}{q}=30\,\upmu {\mathrm{m}} (see Supplementary Section 1).

The hybrid silicon photonic switching network was fabricated using wafer bonding for the heterogeneous integration of InGaAsP with silicon, followed by overlay electron-beam lithography (Fig. 2a; see Methods and Extended Data Fig. 1). To minimize the reflection induced by a complex optical potential²⁷, both ends of the InGaAsP waveguide are tapered to ensure smooth mode transitions (see Supplementary Section 1 for details). The device was characterized using a pulsed input laser at a wavelength of 1,490 nm with a pulse duration of 10 ns, synchronized with the patterned illumination on the curved InGaAsP waveguide from a pump laser at a wavelength of 1,064 nm with a pulse duration of 20 ns. To achieve a high extinction ratio, the thickness of the oxide layer between the two waveguides must be meticulously designed. In the on state, its resultant intrinsic coupling (κ) is sufficiently high to facilitate energy transfer over a short distance, ensuring a compact device footprint for high-density integration. Conversely, in the off state, the large contrast in the imaginary index between silicon and InGaAsP reverses the effective coupling from κ, leading to a sufficiently low coupling strength to minimize modal loss despite the material loss of InGaAsP in the absence of pump. In our case with an oxide layer thickness of 325 nm, the transmission spectrum of a single switching unit, as a function of the pump power that shifts InGaAsP from loss to gain, exhibits high extinction ratios between the on and off states: 15 dB at the through port and 37 dB at the drop port (Fig. 2b; see Methods, Extended Data Fig. 2 and Supplementary Section 2 for more details). Note that in the on state, optical gain amplifies the mode resulting in an increased transmission to the drop port. Moreover, time-correlated measurements (see Methods, Extended Data Fig. 2 and Supplementary Section 3) show an ultrafast switching response (Δt, corresponding to the relative delay time in peak positions) on the order of 100 ps (Fig. 2c,d). As our photonic switching relies on the active excitation of carriers in InGaAsP, the response time is consistent with the carrier lifetime and rise time reported previously in the literature^28,29 and at least an order of magnitude faster than the current state-of-the-art optical switch array^1,3,8,9,11.

Fig. 2: Characterization of a single switch unit.

a, Scanning electron microscopy images of the switch. Left: a single switch unit (outlined by the white dashed box) in the switch array. Top right: expanded and tilted view of area highlighted by the red dashed square in the left image. Bottom right: false-colour image of the area highlighted by the green dashed square in the top-right image, expanded further to show the layer information. E-beam, electron-beam. b, Transmission of the through and drop ports as a function of the pump power. The working ranges of the on and off states are depicted by the red and blue shaded areas, respectively (see Supplementary Section 4 for a more detailed analysis on the insertion loss). c, Time response of a single switch unit measured via time-correlated single photon counting. The blue curve shows the pump pulse collected from the surface of the sample, and the red curve shows the signal pulse collected from the drop port, with a relative peak position delay of 110 ± 38 ps. The time zero is set to be the peak position of the pump pulse. d, Expanded view of the dashed box in c, showing the peak positions of the pump and signal pulses, highlighted by the vertical blue and red dashed lines, respectively.

For our single switch units with the through ports cascading horizontally and the drop ports cascading vertically in two dimensions, we form our non-blocking switch network (Fig. 3). Switching and routing within the network is defined by the pumping strategy. An optical signal from a certain input port remains horizontal before it reaches the pumped switch, and remains vertical afterwards until it reaches the output port. Inside the pumped switch, non-Hermitian PT-symmetric hybrid waveguides bend the light downwards. To verify the non-blocking connectivity, we tested our switches in an 8 × 8 network (Fig. 3a) with an input wavelength of 1,490 nm. The L-shaped regions for optical pumping were generated via a spatial light modulator (SLM). We denote our input state across ports I₁ to I₈ by a vector with the Boolean intensity level (0 or 1) at each port, and we measured the signals across the eight output ports, O₁–O₈, using a monochromator (see Methods, Extended Data Fig. 2 and Supplementary Section 5 for experimental details). In Fig. 3b–d, we show images taken using a charge-coupled device (CCD) camera for three such vectors when the diagonal elements are pumped in the network. Strong signals from the desired output port(s) can be clearly seen, while the rest of the output ports stay dark. For 15 input vectors with one, four or all port(s) launched, we achieved port-to-port non-blocking switching for all of them, with only 3.1% average intensity cross-talk to the undesired ports (Fig. 3e; see Supplementary Section 5 for more information). We attribute this small cross-talk mainly to the unavoidable collection of the scattering light from the input port and the imperfect pump pattern generation from the SLM, where the pump region is blurred and the pump light oversteps its boundary with adjacent switch units. Furthermore, our switch network can be dynamically reconfigured with its non-blocking feature maintained, by redistributing the pump placements (Fig. 3f,g). The new input–output connectivity can also be securely established, with minimal cross-talk (2.7% in Fig. 3f and 3.0% in Fig. 3g).

Fig. 3: Characterization of an 8 × 8 switch.

a, Optical microscope image of the 8 × 8 switch array. b–d, CCD image of the measurement with the diagonal pump but different input vectors: input vector 00000001 with an input only at port I₈ (b), input vector 01010101 with inputs at ports I₂, I₄, I₆ and I₈ (c) and input vector 11111111 with inputs at all of the ports (d). The intensity of photoluminescence has been subtracted. e–g, Output intensity measured at each output port (O₁–O₈) for three different pump patterns. Each output port is normalized individually across 15 different input vectors (I₁I₂I₃I₄I₅I₆I₇I₈). The inset in each figure shows the pump pattern (shaded in red). Our device architecture offers sufficient scalability, where our current 8 × 8 switching network can be fully scalable to a higher dimensionality (see, for example, Supplementary Section 6).

Incorporating spectral parallelism into our switch network, we next demonstrate WSS. Wavelength selectivity is introduced by adding MRRs with various spectral resonances to the input ports (Fig. 4a). With this design, each row in our network will now possess a distinct spectral component. For these experiments, we chose three wavelength (λ_1–3) values, at 1,470, 1,490 and 1,510 nm, a bandwidth that spans a broad range of 40 nm, leveraging the broad gain spectrum (see Supplementary Section 8) of active semiconductors for WSS. These spectral components were multiplexed and simultaneously launched into the switch network from a common input stream (Fig. 4a). Using the same pumping strategy as in Fig. 3, we first establish one-to-one non-blocking connectivity (Fig. 4b), where three wavelengths are switched among five outputs with an extinction ratio (the largest versus the second largest output port intensities at one wavelength) of over 17 dB for all wavelength components. More importantly, with the spectral orthogonality in WSS, more diverse connectivities can be realized by adopting more diverse pumping strategies. By pumping multiple switches within the same column, multiple inputs can now be switched and routed to the same output port (multiple-to-one), where different spectral components are well-separated using the spectrum measurement, with a minimum extinction ratio of 16 dB (Fig. 4c,d). One-to-multiple connectivity is also viable (Fig. 4e), through the wavevector mismatch between the waveguides in two layers. We can control the widths of the waveguides such that PT-symmetric pumping does not bring about complete power transfer to the top layer—non-trivial amount of power still resides in the bottom silicon layer, extending into subsequent switches and making one-to-multiple connectivity possible when multiple switches are pumped in the same row (see Supplementary Section 9). With a versatile pumping strategy, therefore, WSS in our switch network is equipped with a diverse, flexible and reconfigurable input–output connectivity.

Fig. 4: Demonstration of WSS.

a, False-colour schematic of the working principle, which includes three parts: multiwavelength input (with λ₁ = 1,470 nm, λ₂ = 1,490 nm and λ₃ = 1,510 nm), demultiplexing (MRRs filtering different wavelengths to different channels) and switching and multiplexing (with the switch array routing and combining the light to horizontal output ports). b, Schematic of one-to-one switching (left) and the corresponding measured result (right). c–e, Schematics of switching (left) and the measured results (right) for two-to-one switching (c), multiple-to-one switching (d) and one-to-multiple switching (e). In b–e, output signal is normalized by the through port intensity (see Supplementary Section 7 for details).

We have demonstrated a non-Hermitian hybrid silicon photonic switch that is distinct from the existing models and has the potential to boost the on-chip data bandwidth density by around three orders of magnitude (see Supplementary Sections 8 and 10). In terms of the switching mechanism, non-Hermitian phase transition across the exceptional point determines the flow of light exiting the two-layer hybrid silicon switch unit. Compared with non-Hermitian mode switching between two states of the same Hamiltonian^30,31, the switching here in our work occurs between the PT-symmetric and -broken phases of two different Hamiltonians, dynamically tuned by the gain–loss parameter g. In particular, such continuous tunability across the PT phase diagram makes our platform ideal for studies with rich non-Hermitian physics, such as slow light at the exceptional point^32,33. Of the many metrics for evaluating an optical switch, our switch excels in size and switching time, with a compact footprint of only 85 × 85 µm per unit and a switching time on the order of 100 ps, approaching the fundamental limit of the carrier lifetime and rise time in InGaAsP with full utilization of its broad spectral bandwidth. In a network with the large-scale and tunable non-Hermitian functionalities formed by our scalable switch, versatile pumping strategies generate diverse input–output connectivity, where an excellent extinction ratio is sustained across various switching scenarios and multiple wavelengths in the S band. Our non-Hermitian hybrid architecture with InGaAsP integrated on top of silicon photonics is universal, providing opportunities for the exploration of gain materials with nonlinear responses in other visible and optical communication bands^34,35. In addition, our non-Hermitian hybrid switch is compatible with electrically pumped platforms, which can be monolithically integrated with silicon photonics^36,37,38, for large-scale and high-dimensional information processing in optical communication and computing tasks.