Integrated photonics [1] enables the synthesis, processing, and detection of optical signals using the photonic integrated circuit (PIC). The successful translation from laboratory research to foundry development over the past decades has established integrated photonics as a standard technology [2] deployed in high-data-rate telecommunication and datacenters [3]. Foundry-level manufacturing of photonic chips allows fast prototyping or mass production with high yield, high throughput, low cost, and guaranteed performance.

Silicon (Si) and indium phosphide (InP) are two mainstream platforms of integrated photonics [4]. The development of heterogeneous integration [5,6] has married these two platforms and created the first electrically pumped InP/Si lasers [7,8] that have today been used for optical interconnect. Despite that, Si and InP still have many limitations, particularly the high linear and nonlinear losses (e.g., two-photon absorption) that compromise their performance.

To address this challenge, silicon nitride (Si3N4) has emerged as a leading platform for low-loss integrated photonics [9–11]. The 5 eV bandgap of Si3N4 makes it transparent from ultraviolet to mid-infrared, and immune to two-photon absorption in the telecommunication band. Meanwhile, Si3N4 has a dominant Kerr nonlinearity but negligible Raman [12] and Brillouin nonlinearities [13]. In addition, with advanced CMOS fabrication techniques, linear optical loss down to 1 dB/m or even lower has been only achieved in Si3N4 [14–19] among all integrated platforms. All these advantages have triggered the rapid development of Si3N4 Kerr nonlinear photonics [9,10] and have enabled key advances such as optical frequency comb generation [20–22], supercontinuum generation [23–25], and quantum light sources [26–28]. In addition to ultralow loss, tight optical confinement with SiO2 cladding is simultaneously required. Since the refractive index of Si3N4 (n≈1.99) is modestly higher than that of SiO2 (n≈1.45), Si3N4 waveguides require sufficient thickness to achieve tight optical confinement for small mode volume and for bending radii down to 20 μm. In addition, while Si3N4 material has intrinsic normal group velocity dispersion (GVD) at telecommunication bands, Si3N4 waveguides with thickness above 600 nm can obtain net anomalous GVD that is required for phase matching in Kerr parametric processes [29,30].

While CMOS foundries have already developed standard Si3N4 processes [31] to fabricate PIC with typical thickness of 300 nm and loss on the order of 10 dB/m, there has not been a process to simultaneously achieve ultralow loss (e.g., below 3 dB/m) and large thickness (above 600 nm) without crack formation. So far, thin Si3N4 PIC with a width above 5 μm and thickness below 100 nm [17–19] can achieve optical loss below 0.1 dB/m. While this process has recently become an 8-inch (200 mm diameter) foundry process [32], the thin Si3N4 exhibits weak optical confinement and, thus, suffers from exaggerated bending loss with a small footprint. The optical mode also exhibits strong normal GVD due to the waveguide geometry. While coupled waveguide structures can be used to alter local GVD [33,34], they cannot offer anomalous GVD over a wide spectral range. In parallel, a thick Si3N4 PIC of ultralow loss and broadband anomalous GVD has been realized via the subtractive process [35–37] and the photonic Damascene process [14,38]. However, all these reported processes have been developed only in laboratories and have issues with transferring to foundry manufacturing. While high-yield, wafer-scale fabrication of Si3N4 PIC has achieved 1 dB/m loss and anomalous GVD in Ref. [14], the photonic Damascene process has intrinsic limitations. For example, the use of chemical-mechanical planarization (CMP) to remove excess Si3N4 can cause a serious dishing effect in large-area structures (see Appendix C) and, thus, cannot be used to fabricate elements such as arrayed waveguide gratings (AWGs) and multimode interferometers (MMIs). Meanwhile, the aspect-ratio-dependent etch effect prevents the formation of narrow but deep channels [39]. In comparison, with the subtractive process [35–37], a thick Si3N4 PIC of ultralow loss and anomalous GVD has been only achieved with electron-beam lithography (EBL) on 3- or 4-inch wafers, which are not foundry-standard. Meanwhile, an 8-inch foundry process has been developed in Ref. [40], but this process has not yet achieved optical loss below 3 dB/m.

In this work, we overcome the above challenges and demonstrate a foundry-standard fabrication process of tight-confinement, dispersion-engineered, ultralow-loss Si3N4 PIC. The process is based on 6-inch (150 mm diameter) wafers and combines deep-ultraviolet (DUV) stepper lithography [14] and state-of-the-art subtractive process [37], i.e., a DUV subtractive process. We have achieved a linear optical loss of 2.6 dB/m in 810-nm-thick Si3N4 PIC. Finally, we generate single-soliton microcombs of 100.17 GHz and 19.975 GHz mode spacings using these devices.

Figure 1(a) presents a photograph that shows dozens of Si3N4 chips on a 6-inch wafer, which contains microresonators of different free spectral ranges (FSRs) and meter-long spirals. Figure 1(b) presents an optical micrograph that shows a curved bus waveguide slowly approaching a 100-GHz-FSR microring resonator for light coupling. This coupler design can increase coupling strength and ideality, which will be described later. Figure 1(c) presents a scanning electron microscopy (SEM) image that shows the actual Si3N4 waveguide cross section with 85° sidewall angle and overlaid fundamental transverse-electric (TE00) mode. The optical mode is tightly confined in the Si3N4 waveguide core with SiO2 cladding, enabling dispersion engineering and small bending radii.

The Si3N4 PIC is fabricated using the DUV subtractive process. Figure 1(d) shows the subtractive process flow widely used to fabricate PIC based on essentially any material, particularly Si3N4 [35–37,40–44]. First, a Si3N4 film is deposited on a clean thermal wet SiO2 substrate via low-pressure chemical vapor deposition (LPCVD). It is well known that LPCVD Si3N4 films are prone to crack due to their intrinsic tensile stress (typically 1.1–1.4 GPa). The film stress can be relaxed via thermal cycling during Si3N4 deposition in multiple layers [40,41], yielding zero cracks during our fabrication. After SiO2 deposition as an etch hardmask, DUV stepper photolithography is used to expose the waveguide pattern. Via dry etching, the pattern is subsequently transferred from the photoresist mask to the SiO2 hardmask and then into the Si3N4 layer to form waveguides. For superior etch quality and smooth waveguide sidewall, the etchant we use is CHF3 with added O2 to remove fluoride-carbon polymers as etch byproducts.

The etched substrate is thermally annealed in nitrogen atmosphere at 1200°C to eliminate nitrogen-hydrogen bonds in Si3N4, which cause absorption loss. Top SiO2 cladding is then deposited on the wafer, which also requires high-temperature annealing to remove silicon–hydrogen bonds that also cause absorption loss. In specific cases where deuterated plasma-enhanced chemical vapor deposition (PECVD) Si3N4 [44,45] and SiO2 [46] are used, thermal annealing may not be required as these films are intrinsically hydrogen-free.

Platinum heaters [22,47] are deposited on the substrate via an evaporator and patterned via a lift-off process. Due to the thick top SiO2 cladding and tight optical confinement of the Si3N4 core, the presence of metallic heaters does not impact the optical loss of Si3N4 waveguides beneath. Afterward, UV photolithography and deep dry etching of SiO2 and Si are used to define chip size and create smooth chip facets. Finally, the wafer is separated into chips using dicing or backside grinding.

We characterize the optical loss of Si3N4 PIC by measuring the resonance linewidth of Si3N4 microresonators. Light is coupled into and out of Si3N4 chips via lensed fibers and inverse tapers [39]. The fiber-chip edge coupling efficiency is about 60%. We use frequency-comb-assisted diode laser spectroscopy [48] to measure resonance frequency ω/2π and linewidth κ/2π, ranging from 1480 to 1640 nm wavelength. The resonance’s quality factor is calculated as Q=ω/κ. Here we study both the TE00 and TM00 (fundamental transverse-magnetic) modes of the 100-GHz-FSR microresonators of 810 nm thickness and 2.40 μm waveguide width. For each resonance fit [49], the intrinsic loss κ0/2π, external coupling strength κex/2π, and total (loaded) linewidth κ/2π=(κ0+κex)/2π are extracted.

Figure 2(a) shows a typical TE00 resonance with a Lorentzian fit. The resonance is under-coupled (κ0>κex), with fitted κ0/2π=13.8MHz. Figure 2(b) shows a histogram of κ0/2π for 11,741 fitted TE00 resonances from sixty 100-GHz-FSR microresonators. The most probable value is κ0/2π=14MHz, corresponding to a statistical intrinsic quality factor of Q0=1.4×107. The microresonator Q0 and linear optical loss α (dB/m physical length) are linked via α=27.27ngλQ0.(1)At telecommunication band λ=1550nm and with a group index ng=2.09 for the given waveguide geometry, Q0=1.4×107 corresponds to α=2.6dB/m. In comparison, κ0/2π=17MHz is found for the TM00 mode, corresponding to Q0=1.1×107 (see Appendix A). The wavelength-dependent loss of each TE00 resonance is studied in Appendix B, showing no prominent hydrogen-related absorption around 1520–1540 nm.

Next, we investigate wafer-scale fabrication yield. The right of Fig. 2(d) shows our design layout containing 16 chips on the DUV stepper reticle. Each chip has a 5mm×5mm size on the wafer and contains many microresonators. The DUV stepper exposes the reticle uniformly over the 6-inch wafer in discrete fields. The 100-GHz-FSR chips characterized above are labeled as the C11 chips. The most probable κ0/2π values for all C11 chips, as well as their GVD parameters D2/2π (described later), are plotted in each field, as shown in the left of Fig. 2(d). In all measured 20 fields, κ0/2π≤16MHz is found, demonstrating that our foundry process to manufacture ultralow-loss Si3N4 PIC is uniform and near 100% yield.

We also characterize κ0/2π of 100-GHz-FSR microresonators of 2.20, 2.00, and 1.80 μm waveguide width. Again, we create a κ0/2π histogram for each case and look for the most probable values. We plot and compare the most probable κ0/2π values for the three width values and observe a decreasing κ0/2π with increasing waveguide width, as shown in Fig. 2(c). This trend indicates that our optical loss is still dominated by the waveguide’s Si3N4/SiO2 interface roughness and can be further reduced by optimizing the dry etching process or using CMP to reduce top surface roughness [36].

For most applications using Kerr nonlinearity of Si3N4, anomalous GVD is required [29,30]. In addition, for the generation of dissipative Kerr solitons [20,50,51], avoided mode crossings (AMX) induced by spatial mode coupling [52] should be suppressed, as they prohibit soliton formation [53,54] or distort soliton spectra [55,56]. Therefore, next we quantitatively characterize the dispersion profile and investigate AMXs of the high-Q, 100-GHz-FSR microresonators. The microresonator’s integrated dispersion is defined as $$\begin{gathered} Dint(μ)=ωμ−ω0−D1μ \\ =D2μ2/2+D3μ3/6+D4μ4/24, \end{gathered}$$(2)where ωμ/2π is the μth resonance frequency relative to the reference resonance frequency ω0/2π, D1/2π is microresonator FSR, D2/2π describes GVD, and D3 and D4 are higher-order dispersion terms. The top of Figure 3(a) shows a typical Dint profile, with each parameter extracted from the fit Eq. (2). The positive sign of D2 validates anomalous GVD. To reveal AMXs, we remove the D2 term and fit the data with D3μ3/6+D4μ4/24, as shown in the middle of Fig. 3(a). We further remove the D2 and D3 terms, and fit the data with D4μ4/24, as shown in the bottom of Fig. 3(a). The observed AMXs are overall weak and only lead to megahertz-level resonance frequency deviation. It also shows that, with these weak AMXs, our spectroscopic method is sufficiently precise to extract the D4 term.

Furthermore, we experimentally characterize the coupling ideality [57] of our high-QSi3N4 microresonators. In the current case, light in the bus waveguide’s fundamental TE mode (TEb,00) is coupled into the microresonator’s fundamental TE mode (TEr,00) and then coupled out of the microresonator and back into the bus waveguide. In this process, coupling ideality I is defined as I=κexκex+κp.(3)Here κex is the external coupling rate between the TEr,00 and TEb,00 modes, and κp is the parasitic loss rate describing coupling strength to other bus waveguide modes as well as radiation modes into free space. The parasitic loss κp appears as another loss channel in addition to κ0 and κex. Thus, I is a parameter describing how much power is recollected in the TEb,00 mode that is exactly the initial driving mode of the bus waveguide [57,58]. In the ideal case of single-mode bus waveguide and no radiation into free space, I=1 is obtained. In the present case, the bus waveguide is multimode, since it has the same waveguide cross section (thickness × width) as the microresonator waveguide, to obtain phase matching between the TEb,00 and TEr,00 modes for maximum κex. Thus, the TEr,00 mode can couple to other bus waveguide modes (e.g., higher-order TE modes or any TM modes) or radiation modes, which ultimately compromises coupling ideality (I<1).

With a sufficient number of characterized resonances, coupling ideality is evaluated by analyzing the dependence of resonance transmission T and total linewidth κ/2π, as T=|1−2K−1+I−1|2,(4)where K=κex/κ0 describes the coupling regime (K<1 for under-coupling, K=1 for critical coupling, and K>1 for over-coupling [59]). Previously, coupling ideality of integrated Si3N4 microresonators has been characterized [58], however, based on devices of Q<4×106. For state-of-the-art Si3N4 microresonators of Q>107, coupling ideality has not been experimentally studied, and whether high coupling ideality is still maintained needs to be answered.

Here we perform extra measurement on over-coupled devices with smaller gap values down to 300 nm. The microresonators are identical, except that the gap varies from sample to sample to provide a varying κex. Figure 3(b) shows the measured coupling ideality for the TE00 and TM00 modes from dozens of 100-GHz-FSR microresonators. The TE00 [top of Fig. 3(b)] and TM00 [bottom of Fig. 3(b)] plots contain 3269 and 959 data points, respectively. The data are sufficient to uncover the global trend even in the presence of κ0 variation due to fabrication and AMXs. Meanwhile, for comparison, we also plot the κ-T curves with I=0.9,0.95,1, calculated using Eq. (4). As can be seen, our microresonators coupled with curved bus waveguides can already provide near-unity I and strong over-coupling (e.g., K>7 for the TM00), which are critical for microring-based phase modulation [60,61], wideband tunable delay lines [62], and the extraction of quantum light states generated in high-Q microresonators [27,63–65].

A key application area of our Si3N4 PIC is Kerr nonlinear photonics, where ultralow optical loss is central as it determines the threshold power Pth for Kerr parametric oscillation [29]. For example, for soliton microcomb generation [20,50,51], Pth scales with the microresonator Q factor as Pth∝Q−2. Therefore, with this quadratic dependence, a high Q enables significant reduction of Pth down to milliwatt level.

The experimental setup to generate single solitons in our Si3N4 chips is shown in Fig. 4(a). We note that, due to the high-Q and low thermal effects of our Si3N4 microresonators, single solitons can be generated via simple laser-piezo frequency tuning [50,66] or on-chip heaters [22,47]. This is in contrast to many soliton generation experiments that require sophisticated techniques to manage thermal effects [67], such as power kicking [68,69], single-sideband suppressed-carrier frequency shifters [70], dual-laser pump [71,72], pump modulation [73,74], pulse pumping [75], or laser self-cooling [76].

When the continuous-wave (CW) pump laser’s frequency scans across a resonance from the blue-detuned side to the red-detuned side, a step feature (i.e., the “soliton step”) is observed in the microresonator transmission spectrum [50], signaling soliton formation. Figure 4(b) shows a typical soliton step with sub-millisecond duration, sufficiently long for accessing the single-soliton state via simple laser-piezo frequency tuning [50,66]. To further confirm the soliton nature and measure the soliton detuning value, a system response measurement using an electro-optic modulator (EOM) and a vector network analyzer (VNA) is performed [66]. As shown in Fig. 4(c), the system response features double resonances corresponding to the cavity resonance of the CW pump (“C-resonance”), and the soliton-induced resonance (“S-resonance”). Physically, the “cold” cavity resonance is probed by the intracavity low-power CW as a background to the soliton pulse pattern; thus, the observed C-resonance frequency indicates the effective laser-cavity detuning (where thermal induced resonance shift is eliminated) [66]. The S-resonance is induced by the soliton pattern that has a high peak-power leading to nonlinear frequency shift of the cavity resonance.

As shown in Fig. 4(d), in a 100-GHz-FSR microresonator, a single soliton of 100.17 GHz mode spacing is generated with 19 mW power in the bus waveguide on-chip. The spectrum sech2 fit shows 3 dB bandwidth of 17.99 nm, corresponding to a Fourier-limited pulse duration of 141.8 fs. Increasing the pump power to 126 mW and the pump laser detuning, the soliton’s 3 dB spectrum bandwidth is increased to 35.68 nm (pulse duration of 71.47 fs). We also observe a strong Raman-induced self-frequency shift [77,78] of 10.4 nm. Moreover, we generate a single soliton in a 20-GHz-FSR microresonator from the same wafer. Figure 4(e) shows the single-soliton spectrum of 19.975 GHz mode spacing with 518 mW power in the bus waveguide. The 3 dB bandwidth of 21.96 nm, corresponding to a pulse duration of 115.4 fs, covers 117 comb lines. This coherent soliton microcomb with a microwave K-band repetition rate is advantageous for applications such as high-spectral-efficiency telecommunications [79,80], photonic microwave generation [81–83], and astronomical spectrometer calibration [84,85]. Previously, among all CMOS-compatible high-index materials, single-soliton microcombs of repetition rates below microwave K-band (<20GHz) have only been realized in Si3N4, using either the 4-inch photonic Damascene process [82], or the 3-inch EBL-written subtractive process [86]. Our work represents the first foundry-based, 6-inch subtractive process with DUV stepper lithography to reach this goal.

In conclusion, we have reported a 6-inch foundry fabrication process of tight-confinement, dispersion-engineered Si3N4 PIC of optical loss down to 2.6 dB/m and near 100% yield. We have demonstrated its application in soliton microcomb generation with low power threshold and dense channel spacings. While currently our process is based on 6-inch wafers due to our dry etcher, essentially our process can be scaled up to an 8-inch (200 mm diameter) process, which can offer even better uniformity and higher throughput. Merging our ultralow-loss Si3N4 process with established heterogeneous integration [87,88] can introduce a variety of active functions to the passive Si3N4 PIC, such as narrow-linewidth lasers in the UV and visible bands [89,90], broadband EOMs [91,92], fast photodetectors [93,94], and programmable MEMS-controlled networks [95,96]. Together, foundry development of heterogeneous, ultralow-loss Si3N4 integrated photonics could revolutionize next-generation applications for frequency metrology [97,98], photonic neural networks [99,100], and quantum information processing [101,102].

Acknowledgment. J. Liu is indebted to Dapeng Yu who provided critical support for this project. We thank Chao Xiang for the fruitful discussion. J. Liu acknowledges support from the National Natural Science Foundation of China, Hetao Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone Project, and the Guangdong Provincial Key Laboratory. Y.-H. L. acknowledges support from the China Postdoctoral Science Foundation. J. Liu and Z. Y. conceived the fabrication process. Z. Y., H. J., and Z. H. developed the fabrication process and fabricated the samples. C. S. took the SEM images and provided support in the fabrication. J. Long, B. S., Y.-H. L., and W. S. built the experiment and characterized the samples, while H. G., J. H., and Z. Y. provided support and suggestions. W. S., J. Long, and J. H. performed the soliton generation experiment. L. G. took the photos of the sample. Y.-H. L, B. S., and J. Liu analyzed the data and prepared the paper with input from others. J. Liu supervised the project and managed the collaboration.