1. Introduction

Integrated photonic microsystems serve a variety of important applications in the areas of communications, sensing and computing. Microwave photonics (MWP) is one such area in which chip-scale implementations enable significant promise for multifunctional receivers, by leveraging the wide bandwidth and low loss of transmitting and processing microwave signals upconverted to optical carrier frequencies [1–3]. While photonics has been demonstrated to enhance the performance of these receivers, a suitable platform is required to incorporate all necessary components including passives, sources, modulators and detectors. In recent years, silicon nitride (Si₃N₄) has emerged as a leading waveguide material for the realization of photonic integrated circuits (PICs) exhibiting low-loss, high power handling, broad transparency and foundry-compatible processing [4–6]. Heterogeneous integration of III-V semiconductors on Si₃N₄ based PICs also presents a promising approach to implementing high-power lasers, optical amplifiers and photodiodes [7–9]. Realization of an optical modulator suitable for MWP applications however, remains a key capability missing from this selection of materials. To this end, lithium niobate (LiNbO₃) remains one of the most preferable thanks to its exceptional electro-optic properties. In particular, incorporating LiNbO₃ directly on Si₃N₄ enables co-integration with existing component libraries and opens up numerous opportunities for chip-scale MWP systems.

Lithium niobate exhibits many favorable properties for photonic applications, including a low-loss wide transparency window spanning visible to mid infrared and a relatively high refractive index. It also has a strong electro-optic coefficient (r₃₃ > 30 pm/V) and a high second order optical nonlinearity [10]. These attributes have made it a highly used optical material since the 1960s. Despite these favorable attributes, traditional integrated devices in LiNbO₃ have been limited by difficulties with material processing. Particularly, it was quite difficult to produce strong mode confinement in integrated waveguides on account of a low refractive index contrast generated by titanium diffusion or proton exchange in bulk LiNbO₃ [11,12]. In recent years, thin film lithium niobate (TFLN) on insulator has become widely commercially available thanks to the refinement of “smart cut” technology [13,14]. This process of crystal ion slicing enables the isolation of thin films prepared from high quality bulk material on SiO₂/Si substrates. The lithium niobate on insulator (LNOI) platform [15,16] has subsequently become very successful and used to demonstrate state-of-the-art performance for wideband modulators [17], frequency converters [18] and other components. Despite these advances, PIC platforms implemented entirely in LNOI remain limited by challenges to etching the material in CMOS foundries and edge coupling between fibers and chips. In response to these challenges, heterogenous integration has become increasingly popular by combining TFLN with waveguides formed in Ta₂O₅ [19], silicon-on-insulator [20,21] and Si₃N₄ [22–25]. Most demonstrations however, have focused on bonding of LNOI chiplets and not considered compatibility with other necessary microwave photonic receiver components in foundry-compatible processing.

In this work, we explore the development of a TFLN-on-Si₃N₄ platform using PICs that are fabricated in a 200-mm silicon CMOS foundry operating at a 90-nm node. First, we investigate the use of surface-to-surface bonding to form waveguides of LiNbO₃ on Si₃N₄ PICs with various interfaces, where the optical mode spans these layers. A cross-section of the intended heterogenous waveguide is shown in Fig. 1. Three separate approaches are evaluated with the use of die-die bonding – namely SiO₂-SiO₂, atomic-layer deposited Al₂O₃-SiO₂ and direct bonding of TFLN-SiO₂ via plasma surface activation. These interfaces are characterized by evaluating optical losses generated in each class of structure through measurements of ring resonators. Bond strength and yield are also evaluated through the use of die shear experiments and acoustic microscopy. Finally, the Al₂O₃ bonding process is adapted for wafer-scale processing where a 100-mm TFLN wafer is bonded to a 200-mm Si₃N₄ PIC wafer in a CMOS-compatible foundry for the first time to the authors’ best knowledge. Optical losses are characterized for wafers of 150-nm- and 200-nm-thick TFLN bonded layers. In addition to modulators and nonlinear photonics, the piezoelectric properties of LiNbO₃ also make this a suitable platform for optomechanics [26–28] and acousto-optics [29,30] among other potential areas.

 

Fig. 1. Cross-section of bonded thin film lithium niobate on silicon nitride for use in electro-optic modulation and corresponding simulated TE fundamental mode of hybrid waveguide.

Download Full Size | PDF

2. Layer transfer processes for thin film lithium niobate

2.1 Si₃N₄ PIC fabrication and substrate preparation

The heterogeneous waveguides demonstrated in this work begin with a standard Si₃N₄ PIC platform on 200-mm wafers fabricated in MIT Lincoln Laboratory’s Microelectronics Laboratory (ML). Lithography for the photonic devices is performed with a 193-nm stepper having 90-nm resolution. Processing begins with the deposition of a 3-µm bottom-cladding layer of plasma enhanced chemical vapor deposition (PECVD) SiO₂ on 200-mm silicon substrates. The oxide is annealed at 1100°C to reduce optical losses generated by hydrogen content. The target thickness of the bottom-cladding is set by the use of chemical-mechanical polishing (CMP) to reduce surface roughness. Following this, low-pressure chemical vapor deposition (LPCVD) is utilized to deposit a layer of stoichiometric Si₃N₄ with 200-nm thickness. This layer is once again annealed, then patterned and etched via SF₆-based plasma-etch chemistry. Following this etch, another 3-µm thick layer of PECVD oxide is deposited to form the top-cladding. We note that the Si₃N₄ waveguide terminates at the chip edge with inverse linear tapers to match the mode-field diameter found in lensed optical fibers for the subsequent testing. To allow for dicing lanes of individual chips, a deep facet etch is performed at the boundaries of the chips extending into the silicon substrate, laterally extending to the end of the inverted taper. This forms the basis of the standard Si₃N₄ waveguide, as seen in Fig. 2(a), along with its corresponding simulated transverse electric (TE) mode profile.

 

Fig. 2. (a) Initial Si₃N₄ photonic integrated circuit cross-section with standard 3-µm PECVD SiO₂ top cladding and chemical mechanical polishing (CMP) of the SiO₂ down to 100-nm in preparation for bonding with corresponding TE fundamental mode profiles. (b) 200-mm wafer map indicating distribution of thinned SiO₂ top cladding thicknesses, nominally ∼100 nm.

Download Full Size | PDF

In preparation for bonding and the formation of a heterogeneous waveguide, the top-cladding of the Si₃N₄ waveguides is thinned down from 3-µm to approximately 100-nm with the use of CMP. Surface roughness is measured to be ∼0.51-nm rms (over a 90-µm x 90-µm area) confirming the preparation is suitable for direct bonding. The cross-section and corresponding simulated TE mode profile are also shown in Fig. 2(a). Reducing the top cladding thickness modifies the guided mode profile, increasing sensitivity to surface imperfections. As a result of the CMP, there is also a distribution of the top cladding thicknesses as shown in Fig. 2(b). Portions of the wafer with a nominal 100-nm top-cladding are used for the die bonding qualification of the bond interface. However, there is a larger variation in the heterogeneous waveguides developed as a result of the full 100-mm TFLN bonding discussed in the following sections.

2.2 Direct bonding approaches

Using Si₃N₄ PICs with thinned top-cladding, three separate bonding interfaces were evaluated by direct bonding of dies. Multiple experiments were enabled by dicing both the 200-mm PIC wafer and a 75-mm TFLN wafer. We use X-cut LiNbO₃ in which the extraordinary axis of the thin film is aligned with the electric field for the TE mode of the hybrid waveguide. TFLN wafers are sourced from NanoLN, with a 150-nm nominal thickness on 1.8-µm SiO₂ under-cladding used in the die bonding experiments. The three bond processes evaluated include (i) SiO₂-SiO₂ covalent bonding [31,32], (ii) atomic layer deposited (ALD) Al₂O₃ interface bonding [33] and (iii) direct bonding via plasma surface activation (PSA) [21,34]. These are illustrated in Fig. 3(a). Preparation of the samples prior to bond formation involves meticulous cleaning that is constant for the different bonding chemistries used. The target chips are first cleaned in solvents for 5 minutes, followed by 15 minutes of Pirhana cleaning at 130°C. This is followed by a rinse in deionized (DI) water and 15 min of cleaning in RCA-1 solution at 90°C. The samples are then rinsed in DI again and dried with antistatic N₂ gas. The three processes diverge at this step. In the SiO₂-SiO₂ covalent bond process, 20-nm PECVD SiO₂ is deposited on the TFLN dies only. Both sets of dies are then exposed to O₂ plasma at 300W for 5 minutes, followed by the immediate bonding of the TFLN die on the Si₃N₄ PIC die. In the ALD process, both sets of dies are placed in the tool for 4-nm of Al₂O₃ deposition on both surfaces at 200°C. After a brief cooling, the TFLN die is flipped on the Si₃N₄ die and bonded directly out of the tool. Finally, in the PSA process both TFLN and Si₃N₄ dies are transferred to a He/O₂ asher tool for 150 seconds at 150W. The dies are then removed and the TFLN chip is bonded onto the surface of Si₃N₄ substrate. Following the onset of van der Waals attraction in all approaches, the samples are placed under pressure and annealed at 200°C for 3 hours with a slow temperature ramp to strengthen the bond interfaces.

 

Fig. 3. (a) Three bonding approaches involving (i) SiO₂-SiO₂ covalent bonding, (ii) atomic layer deposited Al₂O₃ interface bonding, (iii) direct bonding via plasma surface activation (PSA). (b) Optical microscope image indicating transition between Si₃N₄ waveguides with 100-nm SiO₂ top cladding and region of transferred thin-film lithium-niobate (TFLN) forming hybrid waveguides. (c) Cross-sectional SEM image of the thin LiNbO₃ layer on Si₃N₄ waveguide following bonding and removal of the TFLN wafer’s silicon handle with measured dimensions.

Download Full Size | PDF

Following the formation of successful bonds, the silicon substrate from the TFLN dies must be removed in order to isolate the transferred thin film. This is completed in two steps. First, the Si substrate is thinned through physical grinding from its starting thickness of 400-µm down to approximately 50-µm. The remaining Si substrate is then removed through etching in XeF₂ gas, which selectively stops on the buried oxide of the LNOI wafer. For subsequent optical characterization, the SiO₂ under-cladding of the LNOI wafers is left on the samples, serving the purpose of a top-cladding for the heterogeneous waveguides. An optical image of a processed sample, in which waveguides and ring resonators defined in Si₃N₄ is visible after the transfer of the TFLN layer is shown in Fig. 3(b). Here, the bonded thin film section consists of LiNbO₃ as well as the SiO₂ top-cladding. A cross-sectional scanning electron microscope (SEM) image of the heterogeneous waveguides formed with an Al₂O₃ interface, including measured thicknesses for the various layers is shown in Fig. 3(c).

3. Bonding yield and full wafer-wafer bonding

Die-bonded samples in which the top Si substrate was removed, generated high yield in most cases. The area of transferred TFLN was >90% in most dies where the typical dimensions were 7-mm x 18-mm. While it was not possible to characterize the bond strength in these dies, the TFLN films surviving the transfer process were found to be robust to further fabrication steps in the subset of samples that were processed. In addition, acoustic microscopy (Sonoscan D9000) was used to image the bond interfaces in a non-destructive manner prior to Si top-substrate removal. An image of an example bonded die is shown in Fig. 4(a). Here, the bonded region in the center of the PIC die indicates fairly uniform color and a lack of voids. This implies the bond interface is likely to survive the process of film transfer.

 

Fig. 4. (a) Acoustic microscopy image of 5 × 25-mm TFLN die bonded on Si₃N₄ PIC substrate. (b) Acoustic microscopy image of 100-mm TFLN wafer bonded on 200-mm Si₃N₄ PIC wafer. (c) Sample diced from 100-mm to 200-mm bonded wafers and sub-diced into 1 × 1-mm squares for die-shear testing with results of 3 tested rows (colors correspond to tested rows).

Download Full Size | PDF

Building upon the development of the die-bonding approaches, we also adapted the process for the Al₂O₃-based bond for use with full 100-mm TFLN wafers bonded to 200-mm PIC wafers. The wafer-wafer bonding was conducted in Lincoln Laboratory’s 200-mm CMOS compatible fabrication facility with a toolset previously used for wafer-scale 3D integration [35]. For the wafer-scale process, both sets of wafers were first cleaned in Pirhana solution for 10 min at 120°C, followed by an RCA clean (SC-1 and SC-2) in a megasonic bath. 4-nm of ALD Al₂O₃ are then deposited on both the 100-mm TFLN and 200-mm PIC wafers. In order to handle the 100-mm-diameter wafers, a 200-mm-diameter carrier wafer was made with a 100-mm-diameter recess etched 350-µm deep to place the TFLN wafer into. The wafers were then bonded in a 200-mm substrate bonding tool (Suss SB8e) where the TFLN wafer in the carrier wafer is mounted on the bottom, and the 200-mm PIC wafer was mounted on the top (face down). The wafers were bonded with a pressure of 3000 mBar and annealed at 265°C for 3 hours, with a slow heat ramp. Once the wafers were removed from the bonding tool, acoustic microscopy was again used to inspect the interface as shown in Fig. 4(b). Finally, removal of the Si substrate from the bonded TFLN wafer was completed by grinding the handle down to 90-µm and then placing the 200-mm wafer in a 40% tetramethyl ammonium hydroxide (TMAH) bath at 90°C until the 100-mm surface Si was cleared. TMAH was selected at this stage thanks to its availability in our CMOS-compatible fabrication facility, high etch rate of silicon (>1 µm/min) and high selectivity against oxide (previously characterized >4000:1 for the dilution of the bath used). Here we note the PIC wafers already have oxide on the backside (deposited at the time of bottom-cladding deposition for stress compensation), providing protection for the Si substrate of the 200-mm wafer from the TMAH etch.

For both the die-bonding and wafer-scale bonding cases, the mismatch in dimensions for the substrates complicates quantification of the bond strength through a conventional test such as the double cantilever beam method [36]. In lieu of this, we performed a test of shear bond strength in terms of breaking force for diced samples. This establishes a lower limit force number for the bonding approach. A section of bonded wafers was diced and then subdiced into 1-mm x 1-mm pieces, as shown in Fig. 4(c). A row of the subdies above and below the tested row (colored in Fig. 4(c)) was then removed to allow access for the test fixture of the die shear tool (Xyztec Sigma). For the wafer-scale Al₂O₃ bond, the average shear force is ∼5.28 kgf with statistics for the tested rows shown in the table of Fig. 4(c). While this test provides a relative measure of the bond across the wafer, it is notable that the bond is strong enough to withstand sub-dicing without local degradation [37].

4. Optical characterization

4.1 Ring resonator measurements for various die-bonded interfaces

Characterization of optical losses in the heterogeneous waveguides is completed by testing ring resonators with a variety of ring-waveguide gaps, where the ring diameter is fixed at 500-µm. We analyzed the data for various types of die-bonded samples prepared by the three different bonding approaches in addition to un-bonded PIC waveguides corresponding to the cross-sections in Fig. 2(a). Insertion losses are measured in the range of 1550-nm. To confirm the coupling regime for a given ring-waveguide spacing, we also made measurements at a lower wavelength (∼1480-nm). This helps confirm the intrinsic quality factor (Q_o) for a given resonator. Each die has two identical sets of resonators, which were characterized and averaged. The experimental setup consisted of coupling light from a tunable laser though a polarization controller and a lensed fiber to the edge of the chip under test, while transmission data are collected from another lensed fiber fed to a power meter and normalized to remove coupling losses. The optical loss is then extracted via fitting to a linear-regression based model [38] for each resonance dip in the collected spectrum. For the all-pass rings tested here, the model generates two solution roots corresponding to the coupling state being in the under- or over-coupled regime. By tracking the extinction ratio of the resonances as a function of the ring-waveguide gap (for fixed wavelength), the coupling state is determined for each device. As a result, the mean and standard deviation for the loss generated with each bonding type can be determined from the collection of resonances in the transmission spectra.

Figure 5(a) shows loss statistics for all five types of samples (two un-bonded PIC types and three different bond approaches). The lowest losses, having an average value (µ) of 0.15 dB/cm and standard deviation (σ) of 0.02 dB/cm are found in pure Si₃N₄ waveguides with the 3-µm cladding, as the mode is well confined and least susceptible to surface imperfections. As the top cladding is polished to 100 nm, the losses increase by roughly one order of magnitude (µ = 1.19 dB/cm, σ = 0.03 dB/cm), which represents the state of the waveguides immediately prior to the bond formations. LN-bonded samples exhibit a range of losses, based on the interface layers. Losses increase for hybrid rings formed with PECVD oxide bonding. Compared to the CMP layer, the PECVD thickness (20 nm) is fairly significant, indicating scattering that may be generated in the additional oxide. We find the lowest losses (µ = 0.44 dB/cm, σ = 0.06 dB/cm) in the Al₂O₃ based bond (8 nm total interface). A low loss interface is also found in the PSA-assisted direct bond samples (µ = 0.55 dB/cm, σ = 0.08 dB/cm). Figure 5(b) shows transmission data for an Al₂O₃-bonded heterogeneous TFLN-Si₃N₄ resonator. The highlighted resonance is shown in Fig. 5(c), with a corresponding intrinsic Q of 8.19 × 10⁵. This indicates the LN bonding does not significantly modify the performance of hybrid waveguides, when compared to conventional Si₃N₄ waveguides with thick cladding.

 

Fig. 5. (a) Optical losses for Si₃N₄ and TFLN-Si₃N₄ bonded samples with different interface stacks, where each bar shows the mean and standard deviation for 5 resonances at 1550 nm. (b) Optical spectrum for 500-µm diameter ring in Al₂O₃-bonded TFLN to Si₃N₄ sample. (c) Zoomed in over-coupled resonance data with extracted intrinsic Q.

Download Full Size | PDF

4.2 100-mm TFLN wafer to 200-mm Si₃N₄ wafer-scale heterogeneous integration

Full wafer bonding was performed for 100-mm wafers with both 150-nm and 200-nm TFLN thicknesses using the Al₂O₃ process described in Section 3. Using the bonded wafers, 4 dies with the ring resonator layouts previously used in the die-bonding experiments were tested from different locations of the 100-mm transferred film region. Given the two sets of identical resonators with varying ring-waveguide gaps located on each die, this resulted in a large set of tested devices. In these wafer-bonded samples, rings of various radius dimensions ranging from 50-µm to 250-µm were measured to test for bending losses. The resonator data was fit to extract the intrinsic Q of each resonator, which is used to derive unloaded loss (thus excluding the effect of the coupling gaps) as described in the previous section. Average losses as a function of the tested ring radii are shown in Fig. 6.

 

Fig. 6. Average losses in ring resonators as a function of device radius.

Download Full Size | PDF

The loss data as a function of the device radius shows the losses remain relatively constant for ring radii >150-µm in both 150-nm and 200-nm TFLN bonded layers. In the thicker TFLN layer, the optical mode is more concentrated in the LN region, slightly increasing the effective index (n_eff ∼1.75) and effective area of the hybrid mode (1.32-µm² in 150-nm TFLN vs. 1.34-µm² in 200-nm TFLN). As a result of the lower confinement in the Si₃N₄ waveguide, this helps explain the higher losses for lower bending radius seen in the 75-µm and 100-µm radii resonators. In the larger radii rings however, the losses of the 200-nm bonded TFLN are generally lower than in the 150-nm bonded wafer. It should be noted that in these wafer-bonded samples, the losses are in general an order of magnitude larger than those recorded in the die-bonded samples reported in Section 4.1. This can be observed by comparing the results for the 250-µm radius resonators in 150-µm TFLN. Some notable differences of the full-wafer bond as opposed to the die-bond were the use of the ALD tool in which the interfaces were prepared as well as the bonding tool with the 200-mm carrier wafer in the case of the full-wafer processing. It is possible the interface preparation in the full wafer process is not as well-controlled as a result, potentially leading to increased scattering. Process optimization for the full wafer bonding remains a subject of investigation.

An image of a fully bonded 100-mm TFLN wafer following the Si substrate removal is shown in Fig. 7(a). A metalized pattern appears on the 200-mm wafer indicating the location of the separate dies where additional devices share a single reticle with the Si₃N₄ PICs bonded to in this work. This demonstrates process compatibility for multi-project runs with the TFLN-bonding process implemented here. The transferred TFLN layer (appearing in light purple) shows high yield relative to the bounds of the original 100-mm wafer (boundary specified by the dotted line). We also note the distribution of the measured resonator intrinsic quality factors as a function of their location on the 150-nm and 200-nm bonded TFLN wafers. This is shown in Figs. 7(b) and 7(c), respectively. Intrinsic quality factor is plotted in order to again exclude the impact of the coupling gap dimensions on the measured (loaded) quality factor of the resonators. The range of measured quality factors is indicated in the scale bar of the plots. On the surface of the wafer, resonators with closely measured values tend to appear closer to one another. This may be attributed to the previously indicated variation in the top oxide bonding layer as shown in Fig. 2(b), and other scattering features of the interface.

 

Fig. 7. (a) Image of 100-mm-diameter TFLN wafer bonded to full 200-mm-diameter Si₃N₄ PIC wafer following TFLN wafer Si handle removal, showing high yield for transferred film. Intrinsic quality factor plotted for 150-µm and 250-µm radius rings relative to their location on the surface of the 100-mm bonded wafer for (b) 150 nm TFLN layer, and (c) 200 nm TFLN layer.

Download Full Size | PDF

5. Conclusion

A robust set of bonding approaches have been demonstrated for developing heterogeneous waveguides of thin-film lithium-niobate on silicon nitride using a variety of interface chemistries in die-based samples. Lowest losses were obtained in the Al₂O₃ die-bonded samples. The loss characterization indicated bonded samples do not significantly increase the losses from silicon nitride PICs with CMP on the top-cladding. In addition, we have adapted the ALD-Al₂O₃ based interface for bonding 100-mm TFLN wafers to 200-mm Si₃N₄ PIC wafers in a CMOS-compatible foundry process. While the losses recorded in the full-wafer approach were roughly one order of magnitude larger for the same TFLN thickness in 250-µm radius resonators, bonded wafers demonstrated good yield, robustness to dicing/shear forces and consistent quality factors across 100-mm surfaces. Through the use of a 200-mm carrier wafer and a bonding tool, our process can also be implemented for 150-mm TFLN wafers, which is currently the size limit for commercially available wafers. Initial experiments have shown promising results, though process optimization for replicating the low losses recorded in die-bonded samples remains. Samples prepared with both die- and wafer-scale bonding processes have also been successfully post-processed, opening up multiple prospects in acousto- [39] and electro-optics [40] with the hybrid TFLN-Si₃N₄ platform. Finally, compatibility of the wafer-scale bonding process with other foundry-based multi-project runs indicates strong prospects for several applications including microwave [40], quantum [41] and nonlinear [42] integrated photonic subsystems.