Integrated quantum photonics has rapidly developed in recent years, as it holds great promise for realizing complicated quantum photonic applications with small footprints, therefore overcoming the limitations of traditional bulk optics [1]. Single-photon detectors are essential components of quantum photonic systems [2,3], making the on-chip integration of these detectors a pressing requirement that can enhance scalability and eliminate lossy fiber interconnections [4,5]. With outstanding performance, superconducting nanowire single-photon detectors (SNSPDs) are widely utilized in quantum photonic experiments [6–8], exhibiting ultrahigh detection efficiency [9,10], negligible dark count rate [11,12], and low timing jitter [13,14], which makes them among the most attractive candidates for co-integration with quantum photonic circuits [15].

Suppression of photonic noise is necessary for SNSPDs to be properly operating, particularly when the pump power is substantially higher than the generated photon pairs [15,16]. In most experiments related with photon pair generation, pump rejection is achieved by off-chip filters before detection [17,18], thereby hindering the compactness of the whole system. On-chip pump filters have been demonstrated through various strategies on silicon photonics platform, such as cascaded microrings [19], cascaded Mach–Zehnder interferometers [20], and contra-directional couplers [21], and some of them are monolithically integrated with photon pair sources [20,22] as entanglement suppliers. However, in future large-scale quantum networks, the receivers may require pump filters with different wavelengths and rejection ratios for various applications; thus it would be more flexible to integrate different pump filters on receiver side, allowing for unique demands. But when it comes to entanglement receivers, integrating pump filters with SNSPDs on a single chip still remains challenging due to the stringent requirements of cryogenic environments, which limit the use of active tuning elements. Silicon Bragg grating filters are entirely passive, making them cryogenic-compatible. Moreover, their robustness to fabrication imperfections [23] allows for cascading multiple sections to obtain a high pump rejection ratio and provides tolerance to additional fabrication process induced by SNSPDs.

In this work, we demonstrate the heterogeneous integration of SNSPDs and pump rejection filters (PRFs) on silicon photonics platform, with the aim of removing strong pump light prior to entanglement characterization. Our on-chip SNSPDs exhibit saturated detection efficiency, and by cascading seven sections of Bragg grating filters, we achieve on-chip pump rejection exceeding 56 dB. This integration combines the advantages of high-performance SNSPDs and PRFs on a single chip, eliminating the need for fiber or bulk optical filters and the associated separate interfaces. In addition, by feeding correlated photon pairs from a silicon microring resonator [24], we proceed to analyze energy-time entanglement using a dense wavelength division multiplexer (DWDM) and Franson-type interferometers [25], showing two-photon interference fringes with visibilities of $92.85\%\pm 5.95\%$ and $91.91\%\pm 7.34\%$, along with a coincidence-to-accidental ratio (CAR) that reaches 32. Our results effectively show the successful removal of pump light before detection, bringing a step closer to future scalable quantum information processing.

As Fig. 1(a) shows, the integrated chip consists of seven cascaded Bragg gratings functioning as PRF, along with a waveguide-integrated SNSPD at the output of the gratings. Scanning electron microscope (SEM) images of the Bragg gratings and SNSPDs are shown in Figs. 1(b)–1(d). Our photonic circuit is fabricated on a standard silicon-on-insulator (SOI) wafer by electron beam lithography (EBL) and inductively coupled plasma (ICP) etching, with a cross-sectional dimension of $500\mathrm{nm}\times 220\mathrm{nm}$ for single-mode waveguides. TE-polarized light from single-mode optical fibers is coupled to the waveguides through grating couplers, with an average coupling loss of $-3.9\mathrm{dB}$ at 1550 nm.

To achieve a high pump rejection ratio, the Bragg gratings have been carefully designed to reflect light around the pump wavelength into the second-order mode (see Appendix B for additional details on the gratings). These backreflections are radiated away in single-mode waveguides interconnecting adjacent gratings, thus preventing coherent interactions and enabling effective cascading for high rejection [23].

To characterize the central wavelength, rejection ratio, and bandwidth of the PRF at cryogenic condition, we cool the chip down to 2.2 K and measure its transmission spectrum using a tunable continuous-wave (CW) laser (Keysight N7776C) and an optical power meter (Thorlabs PM100D). For comparison, the transmission spectrum at room temperature is also recorded. The PRF has a central wavelength of 1556.1 nm at room temperature and shifts to 1544.0 nm at a cryogenic temperature of $\sim 2.2\mathrm{K}$, which is caused by the decrease in the refractive index of silicon at low temperatures [26]. Further, the rejection ratio and bandwidth remain nearly unchanged between room and cryogenic temperatures. A rejection ratio exceeding 55 dB is obtained, with a 3 dB bandwidth of 4.9 nm [Fig. 2(a)]. However, due to optical crosstalk between the input and output waveguides, and the noise floor of the power meter, the measured rejection ratio result is not precise enough. To obtain a more accurate estimation of the actual rejection ratio, we record the counting trace of the on-chip SNSPD while scanning the CW laser wavelength across the rejection band of the PRF (see Appendix C), resulting in a rejection ratio of 56.5 dB at 1544.5 nm [inset of Fig. 2(a)].

SNSPD is formed in U-shaped niobium nitride (NbN) nanowire atop the waveguide, with a thickness of $\sim 6\mathrm{nm}$ and width of $\sim 80\mathrm{nm}$. To prevent the detector from latching, two inductors made of wider ($\sim 150\mathrm{nm}$) meander nanowires are added [27,28]. A 5 nm thick ${\mathrm{SiO}}_2$ layer is deposited prior to the sputtering of NbN to protect the surface of the SOI wafer during our top-down fabrication process (see Appendix A for more details). The nanowire length is set to 100 μm to ensure sufficient absorption of light through the evanescent field of the waveguide. To evaluate the performance of the on-chip SNSPD, we mount it in a two-stage GM cryocooler, which has a base temperature of 2.2 K, and a critical current of 9.7 μA is recorded.

For the detection efficiency of the SNSPD, a CW laser light at 1550 nm is attenuated to a photon flux of ${10}^5\text{photons}/\mathrm{s}$ and then coupled into the waveguides by a fiber array. The system detection efficiency (SDE) is calculated as a function of bias current, as illustrated in Fig. 2(b). Our detector exhibits a maximum SDE of 20.1%, and the plateau of the saturating SDE curve indicates a high internal quantum efficiency [29]. It should be noted that the coupling loss of the grating coupler is approximately $-3.9\mathrm{dB}$, and the propagation loss from the optical waveguide is approximately $-3.0\mathrm{dB}$ (see Appendix C for further details). Both types of losses could be significantly reduced to achieve a high SDE.

Since our PRF is aimed for the removal of pump light used in photon pair generation, we perform time correlation measurements to obtain coincidence counts of the photon pairs. Figure 3 shows the experimental setup, where photon pairs are generated by spontaneous four-wave mixing (SFWM) in a 15-μm-radius silicon microring resonator with a quality factor of $2.2\times {10}^4$ [30]. A CW laser light at 1544.5 nm serves as the pump, matching the resonant wavelength of the microring [Fig. 4(a)], which is also in the rejection band of the PRF. The pump light passes through a bandpass filter to eliminate spontaneous noise and a polarization controller to keep TE mode polarization, before being coupled to the microring by fiber array. The generated signal-idler photon pairs are input into a 32 channel DWDM, which directs the signal and idler photons to appropriate channels with a 100 GHz bandwidth. The output photon pairs are then coupled to two receiver chips by fiber arrays, which are glued to the chips by cryogenic-compatible epoxy, and detected by the on-chip SNSPDs; meanwhile the residual pump has been filtered by the PRFs. Finally, a time tagger (Swabian Instruments Time Tagger 20) connected to the SNSPDs is employed for recording coincidence counts.

The DWDM has an extinction ratio of 40 dB. Along with the 56.5 dB from our PRFs, a total rejection ratio up to 96.5 dB should be provided. By biasing the two SNSPDs at 95% of their critical currents, the measured coincidence peak shown in Fig. 4(b) clearly indicates that the pump is sufficiently removed before detection, confirming the effectiveness of the PRFs. Further, using on-chip PRFs instead of off-chip alternatives reduces the extra noise generated in interconnecting fibers and helps improve the correlation [30].

To further evaluate our device, we extract another figure of merit, i.e., the coincidence-to-accidental ratio (CAR), which is the ratio of net coincidence counts to accidental coincidence counts [31]. Typically, CAR depends on pump power as well as noise rejection [20]. In our case, we perform a Gaussian fit to the histogram in Fig. 4(b), and the full width at half-maximum (FWHM) of the peak represents coincidence counts, while the averaged background indicates accidental counts. A calculated CAR of 32 is obtained; we note that this value is not particularly high due to the dark counts of our SNSPDs, which are approximately 100 Hz because of the imperfections in our chip packages. The absence of a shielding metal block introduces background light noise and could be improved in the future. The dark counts of the detectors contribute to the background counts in the coincidence histogram, thus resulting in high accidental counts and low CAR. Although a higher CAR could be achieved with lower pump power, we opt for this power level owing to the trade-off between noise and photon pair flux. With reduced pump power, the generated photon pair flux will also decrease, resulting in fewer coincidence counts and longer integration time. In addition, Raman scattering produced in the connecting fibers also contributes to the noise.

The generated photon pairs are expected to be energy-time entangled, which could be verified by a Franson-type two-photon interference experiment [25]. As depicted in Fig. 5(a), the signal and idler photons pass through two unbalanced Mach–Zehnder interferometers (UMZIs) before reaching our integrated chips. The relative time difference between two arms of the UMZIs is set to be $\mathrm{\Delta }T=400\mathrm{ps}$, which is longer than the coherence time of the signal and idler photons but shorter than the coherence time of the pump laser [17,32]. In this case, the signal and idler photons can travel along either the long ($L$) or short ($S$) path of the UMZIs, consequently allowing four possible combinations ($L$ & $L$, $S$ & $S$, $L$ & $S$, $S$ & $L$), and producing three peaks in the coincidence count histogram. The side peaks arise from photons traveling along different paths ($L$ & $S$ or $S$ & $L$), and the central peak occurs when both photons take the same path ($L$ & $L$ or $S$ & $S$). Because the two cases contributing to the central peak are indistinguishable, the two-photon state could be expressed as $$|\psi ⟩=\frac{1}{\sqrt{2}}(|LL⟩+e^{i{\phi }_{i+s}}|SS⟩),$$(1)where ${\phi }_{i+s}$ is the sum of the phase difference acquired by signal (${\phi }_s$) and idler (${\phi }_i$) photons passing through the long ($L$) and short ($S$) arms of the UMZIs. The relative phase differences of the UMZIs could be tuned by resistive heaters and are proportional to the squares of the voltages applied on the UMZIs.

The coincidence counts of the central peak will vary sinusoidally with ${\phi }_{i+s}$, showing a two-photon interference fringe. If the fringe visibility exceeds $1/\sqrt{2}(70.7\%)$ under two nonorthogonal bases, it shows a violation of Bell’s inequality and verifies energy-time entanglement [33,34].

In our experiment, the coincidence counts are measured under different ${\phi }_s$, whereas ${\phi }_i$ is fixed at ${\phi }_i=0$ and ${\phi }_i=\frac{\pi }{4}$. By varying ${\phi }_s$, the single-side count rates at the outputs of each UMZI remain constant, as shown in Fig. 5(b), confirming that the time difference is much longer than the coherence time of the signal and idler photons. The two-photon interference fringes fitted by sinusoidal curves, are shown in Fig. 5(c), with raw visibilities of $92.85\%\pm 5.95\%$ and $91.91\%\pm 7.34\%$, respectively, confirming the violation of Bell’s inequality. Notably, these raw visibilities are calculated without subtraction of photonic noise or detector dark counts.

These results confirm that our integrated chips serve as effective receiver chips for removing the strong pump light and distributing energy-time entanglement resources, which will be promising for complex quantum information processing.

We experimentally demonstrate on-chip SNSPDs integrated with pump rejection. Our PRF, composed of cascaded Bragg grating filters, exhibits a pump rejection beyond 56 dB in an all-passive manner and completely filters out the pump light with standard fiber demultiplexers. The successful verification of energy-time entangled photon pairs reveals the effectiveness of our pump rejection method [35,36], while also avoiding the extra noise typically introduced by off-chip filters. The on-chip SNSPDs exhibit favorable performance, with a saturating detection efficiency that indicates the reliability of our top-down fabrication process.

To move a step further, bandpass filters (BPFs) based on Bragg gratings [37] could also be added to our PRFs, in which way the signal and idler photons could be demultiplexed and routed to different SNSPDs [38] on the same chip. Moreover, by cascading the BPFs and PRFs, we could achieve a pump rejection ratio above 100 dB. This strategy would enable the photon pair sources (e.g., ring resonators and spiral waveguides) to be integrated together with pump rejection and detection on a single chip, realizing a fully integrated quantum photonic system [1,5,39].

Currently, the traditional U-shaped architecture of our SNSPDs may not provide the optimal solution for overall performance [40,41]. To address the trade-off between efficiency and timing metrics, one potential strategy is to embed superconducting nanowires into subwavelength grating structures [42]. By carefully designing the grating geometry, high absorption efficiency could be attained even with ultrashort nanowires, thus simultaneously optimizing efficiency and timing performance.

In summary, we have demonstrated on-chip SNSPDs integrated with pump rejection exceeding 56 dB in an all-passive way, along with a saturated SDE of 20.1%. Our integrated chip verifies the generation of energy-time entangled photon pairs with visibilities up to $92.85\%\pm 5.95\%$ and $91.91\%\pm 7.34\%$, indicating the sufficient removal of pump and the potential for further quantum state manipulation.