Recessed SiNx stressor for enhanced and uniform mechanical strain
To present a wider adaptability of this technique, as a proof of concept, the broadband photodetection and EA modulation were demonstrated on Ge-on-insulator (Ge-OI) and Ge0.99Si0.01 on SOI platforms, respectively. The Ge-OI photodetectors are first discussed. In contrast to the technical maturity of the Ge-on-SOI platform, the Ge-OI platform is recently gaining attention for both advanced electronic41 and photonic42, 43, 44 applications. In this work, the Ge-OI was formed on a 200-mm Si (100) wafer via Ge-on-Si epitaxy, bonding, and layer transfer, using silicon dioxide (SiO2) as the intermediate insulator layer45. The detailed fabrication process is discussed in Supplementary section S1. The bonding and layer transfer approach could facilitate a scalable Ge-OI fabrication to any wafer diameter and Ge thickness, with a superior Ge quality than that from the direct Ge-on-Si epitaxy46. Its low process thermal budget (~300 ºC) enables a back-end-of-line electronic-photonic integration. It is without doubt that both the photodetectors and modulators can be demonstrated interchangeably on the other platform. The Ge-OI platform was also used to study the effect of the recessed SiNx stressor on its induced Ge mechanical strain (Fig. 2). A tensile stressor is placed in trenches at both sides of w-Ge-OI and in contact with Ge via the waveguide sidewalls (Fig. 2a); while deeper trenches are created into the underlying SiO2 as the recessed configuration to accommodate the stressor (Fig. 2b). The modelling was performed using a finite element method, with a detailed description in Supplementary section S2. The results are shown in Fig. 2c, d. It was found that the transverse tensile strain () in Ge is enhanced with the use of the recessed SiNx stressor, especially at the bottom portion of Ge closer to the buried SiO2. The strain uniformity is thus improved simultaneously. This increased uniformity is due to the recessed stressor configuration that allows the tensile stress to be applied on both the sidewalls of SiO2 and Ge to pull on both materials. The constraint to the sidewall stressor from the underlying SiO2 and Si bulk is therefore alleviated, enabling a uniform Ge strain profile without selective substrate removal to form suspended Ge structures38, 39, which facilitates a more foundry-compatible strained-Ge integration. The longitudinal strain in Ge ( ~0.17%) reveals a negligible change irrespective of the use of the SiNx stressor (Supplementary Fig. S2). For a quantitative illustration, the vertical (from O to A) and transverse (both O and A to the Ge/SiNx interface) values (in Fig. 2c, d) were extracted and plotted in Fig. 2e, f, respectively. It can be clearly seen that the is enhanced by ~2× near the Ge/SiO2 interface along the points O and A, with the 1-GPa recessed tensile SiNx stressor (Fig. 2e). Additionally, variations along both the transverse (x-) and vertical (z-) directions are significantly reduced with the use of the recessed SiNx stressor, again indicating a more homogeneous strain distribution. Similar effects can be expected for Ge-on-SOI and other material platforms with the use of both tensile and compressive recessed stressors.
To experimentally demonstrate the effect of the recessed SiNx stressor, strained w-Ge-OI structures similar to that in the model (Fig. 2c, d) were fabricated, with the detailed process depicted in Supplementary Fig. S5. The SiNx was deposited in a Cello Aegis-20 plasma-enhanced chemical vapor deposition (PECVD) system. Optimizing the deposition parameters results in a tensile film stress of ~600 MPa. The details on the stress characterization and optimization are discussed in Supplementary section S4. The SiNx on the top of w-Ge-OI was removed via a second electron-beam lithography (EBL) followed by reactive-ion etching (RIE), as it induces undesired compressive strain to the Ge underneath (Supplementary Fig. S8). Chemical mechanical polishing (CMP) can be an alternative process for the top SiNx removal at wafer scale. Cross-sectional scanning electron microscope (SEM) images of the fabricated structures are shown in Supplementary Fig. S6. The induced strain in Ge was characterized via commercial micro-Raman spectrometers at the longitudinal optical (LO) mode. More information about the measurement tools and calibration can be found in Methods. Two laser excitation wavelengths of 532 and 785 nm were used and lead to distinct photon penetration depths of ~9 and ~89 nm47, respectively, into Ge, facilitating Raman spectra revealing depth-dependent strain information. For this reason, the w-Ge was thinned down to ~100 nm, allowing a full Ge thickness coverage for the strain acquisition at the laser excitation of 785 nm.
The micro-Raman spectra on 0.5 µm-wide w-Ge-OIs are shown in Fig. 2h. The laser spot diameter of ~1 µm at both 532 and 785 nm ensures a complete transverse coverage of the waveguide structures. It can be observed that the peak LO phonon frequencies of all spectra are red-shifted compared to that of bulk Ge (~300.8 cm-1), indicating that a tensile strain is induced in Ge as expected. It is important to mention that the widths of the w-Ge-OI sidewall trenches are within 10 µm to accommodate the SiNx stressor, and the corresponding Raman spectra exhibit a consistent Ge tensile strain independent of the trench width (Supplementary Fig. S10). The narrow trench width with a significant strain effect could envision a compact device integration. The Raman spectra reasonably overlap at both 532 and 785 nm for w-Ge-OIs with the recessed stressor, while only a partial overlap below ~300 cm-1 was seen for that with the non-recessed stressor. The broader full-width-at-half-maximum (FWHM, ~12.90 cm-1) of the spectrum for the non-recessed w-Ge-OI at 785 nm, compared to the recessed structure (~9.04 cm-1), is attributed to the strain non-uniformity rather than strain-induced crystal quality deterioration48. This conclusion is made based on the gradual blue shift of the peak LO phonon frequencies with time, observed from some early-batch tensile-strained w-Ge-OIs (Supplementary Fig. S9), where the SEM inspection reveals delaminated stressors from the w-Ge-OI sidewalls. The matching Raman spectra with respect to an un-strained w-Ge-OI (~300 cm-1), after the stressor delamination, verify that the stressor has not caused any Ge crystal quality degradation. Therefore, the phenomena in Fig. 2h suggests an improved strain uniformity in Ge vertically along the z-axis as seen in Fig. 2d. Furthermore, it can be inferred that the Ge is compressively strained at the bottom part of Ge closer to the Ge/SiO2 interface, since the deeper-penetrating Raman spectrum at 785 nm (Fig. 2h, upper panel) exhibits a shoulder at ~305 cm-1 while the spectrum at 532 nm does not. The compressive strain might be due to both the substrate constraint from the SiO2-on-Si bulk and the tensile strain at the top portion of w-Ge, which collectively exerts a compressing force to the bottom part of w-Ge. The reason for no compressive strain shown in the finite element simulation pertains probably to the thinner substrate thickness (100 µm) employed in the simulation than a 200-mm Si wafer (~725 µm), which alleviates the substrate constraint. The compressive Raman shoulder was suppressed with the adoption of the recessed stressor (Fig. 2h, lower panel), suggesting a strain enhancement at the bottom part of Ge. This agrees well with the mechanism of the recessed stressor from the simulation (Fig. 2c, d). As the tensile stressor at both sidewalls essentially pulls the sandwiched Ge in the middle for a tensile strain, the adhesion of the stressor to the sidewalls becomes critical. It is noteworthy that oxygen (O2) plasma treatment significantly enhances the duration of the stressor adhesion (Supplementary Fig. S9). Additional micro-Raman measurements on 2 µm-wide w-Ge-OIs find that the compressive strain is accumulated close to sidewalls at the bottom part of w-Ge-OI, which was turned into tensile strain with the use of the recessed stressor (Supplementary Fig. S7), thus improving the Ge strain uniformity with an enhanced magnitude. The mechanism can be similarly applied to the recessed compressive SiNx stressor for the magnitude and uniformity enhancement of the compressive strain.
Broadband strained Ge-OI MSM photodetectors
To investigate the effect of the recessed SiNx stressor on the performance of Ge photo-detection, metal-semiconductor-metal (MSM) photodetectors were developed based on the strained w-Ge-OI structures, with a schematic shown in Fig. 3a. As a reference, MSM detectors on w-Ge-OIs without stressor were also fabricated. All w-Ge-OIs have identical Ge width of 1 µm and thickness of 400 nm. The detailed fabrication for all the devices is illustrated in Supplementary section S8. It is noteworthy that Al2O3 is employed in the fabrication, serving both as an interlayer (~1 nm) at the metal/Ge interface and an etch-stop layer (~20 nm) between the SiNx and Ge. The former unpins the Fermi level of Ge from its valence band edge49 for an elevated Schottky barrier height (Supplementary Fig. S12) and consequently a suppressed dark current, while the latter avoids an undesired etching of Ge during the SiNx removal at its top50. A top-view SEM image of a recessed SiNx-strained device is shown in Fig. 3b, and the cross-sectional FIB-SEM images of the devices without and with the use of the recessed stressor are shown at the left of Fig. 3e, f, respectively. There is no observable delamination of SiNx stressors from the Ge sidewalls, indicating a good device fabrication. The current-voltage (I-V) characteristics of the recessed detector, both with and without normal-incidence illumination of 20 mW laser power at 1,550 nm, were measured and the results are displayed in Fig. 3c. An explanation of the measurement set-up can be found in Methods. Dark currents of ~9.86 and ~41.70 µA were observed at biases of 1 and 2 V, respectively, while the illumination leads to prominent photocurrents of ~1.18 and ~2.01 mA accordingly. Considering the Ge effective absorption area under the optical fiber tip (tip diameter ~9 µm) without metal shielding, an optical responsivity of ~1.6 A/W was obtained at 2 V, corresponding to an internal quantum efficiency (IQE) of ~129% (Fig. 3c). Compared to that of the reference detectors without stressor, the IQE is significantly enhanced (Supplementary Fig. S14). As the only difference in fabrication between the two devices lies in the introduction of SiNx, the IQE enhancement can be attributed to the defect and trap states in the SiNx, due to its nonstoichiometric (nitrogen-rich) nature from both the ellipsometry measurement (Supplementary Fig. S4) and the low SiH4/NH3 ratio used in the deposition (Supplementary Table S2). Moreover, PECVD-deposited SiNx films are commonly reported with heavy hydrogen (H-) incorporation51. The trap and defect states induce image carriers accumulated at the Ti/Al2O3/Ge metal-insulator-semiconductor (MIS) interfaces, forming interfacial dipoles and fixed charge. This can be explained by the Schottky barrier height lowering of the MIS contact with the use of the SiNx stressor (Supplementary Fig. S12), which had been correlated to the formation of interfacial dipoles52 and fixed charge53 at MIS contacts in earlier studies. Meanwhile, these dipoles and charge enhance the electric field across the Al2O3 interlayer, contributing to an easier transport of photon-generated carriers across the interlayer and consequently a higher IQE. A supporting evidence to this claim comes from an earlier onset of Poole-Frenkel emission in the SiNx-strained detector (Supplementary Fig. S15). The IQE >100% indicates the existence of a photocurrent gain. The gain is unlikely originated from the interfacial charge as previously reported for MSM photodetectors54 that induces an injection of extra image carriers into the device for a charge neutrality. This argument is based on the observation from the electric field simulation (Supplementary Fig. S16) indicating that the carrier transport in Ge had reached its saturation velocity (~6×106 cm/s) at a low bias of 0.3 V, implying a full collection of photon-generated carriers and a saturated IQE. However, this is inconsistent with the increasing, and < 100%, IQE with an increasing bias observed in Fig. 3c and Supplementary Fig. S14. This suggests that the photocurrent is hindered by the Al2O3 barrier and the barrier does not provide a gain mechanism to the photocurrent. Further observation finds that the gain can be attributed to the avalanche multiplication in Ge, since the electric field at 2 V (~180 kV/cm), compared to that in literature55, is sufficiently high to impose avalanche amplification (Supplementary Fig. S16).
Fig. 3d shows the photocurrent spectra of the MSM photodetectors from 1,500 to 1,630 nm, covering the C- and L-bands (1,530–1,625 nm). In contrast to the photocurrent roll-off at ~1,540 nm for the Ge detector without stressor, a prominent drop of the photocurrent appears at longer wavelengths of 1,606 and 1,612 nm for SiNx-strained detectors without and with the use of the recessed-type stressor, respectively. The roll-off points of the photocurrents have been identified as the Ge direct bandgap edges at the G-valley that determine the efficiency of direct-band absorption56. The ~70 nm roll-off extension, along with the relatively flat spectra before the roll-off, indicates an enhanced detector quantum efficiency similar to that at the C-band covering most of the L-band. A further observation finds that there are additional distinguishable roll-off points at ~1,580 and ~1,615 nm for the non-recessed SiNx-strained detector. This again demonstrates the strain non-uniformity in the Ge photodetectors without the use of the recessed SiNx stressor. To investigate the correlation between the absorption coverage extension and the strain-induced Ge bandgap shrinkage, deformation potential theory57 was utilized to calculate the Ge G-valley-light-hole (G-LH) and G-valley-heavy-hole (G-HH) bandgap edges as a function of the transverse strain , shown as the blue solid and red dotted lines, respectively, in Fig. 3g. The detailed calculation procedure is discussed in Supplementary section S11. A constant longitudinal strain of 0.17% was employed based on the earlier studies (Supplementary Fig. S2). Additionally, to estimate the Ge tensile strain in the SiNx-strained detectors, the finite element simulation as described in Supplementary section S2 was performed by constructing the material and geometrical structures identical to that fabricated (Fig. 3e, f (both at left)). The tensile stress in SiNx (600 MPa) is also identical. The simulated profiles are shown in Fig. 3e, f (both at right). As expected, a uniform of ~0.42% is obtained for the recessed SiNx-strained w-Ge-OI, while the non-recessed stressor results in a two-step profile of ~0.35% at the upper part of Ge and ~0.23% at the lower part. Combining these results, the calculated Ge bandgap edges at the respective simulated (Fig. 3g) reasonably match with that observed from the photocurrent spectra in Fig. 3d. Subsequently, the simulated was used to calculate the absorption coefficient (a) enhancement of Ge, based on the photocurrent spectra. The detailed calculation is discussed in Supplementary section S10. Compared to the a (2,265 cm-1) of the w-Ge-OI without stressor, the a (7,474 cm-1) is enhanced by ~3.3× for the recessed SiNx-strained w-Ge-OI at 1,612 nm, which is comparable with that of In0.53Ga0.47As material at the same wavelength58. This suggests Ge photodetectors could reach a similar optoelectronic performance as InGaAs detectors and extend its application into prevailing III-V-dominated fields such as short-wave infrared imaging. In addition, the 3.3× increase on a implies a ~70% reduction on the photon penetration depth (∝ 1/a) in Ge. The detector could consequently become more compact (a ~70% reduction in length and capacitance) with a commensurate photon absorption, alleviating the common responsivity-bandwidth trade-off. A maximum 3.3× enhancement of the 3-dB bandwidth can be expected, provided that the device is limited by RC delay. It is also worthwhile to note that the equivalent a (6,000 cm-1) for the w-Ge-OI without stressor at 1,500 nm is extended by 120 nm towards 1,620 nm for the recessed SiNx-strained w-Ge-OI, covering almost the entire C- and L-bands. This indicates that the performance of a Ge photodetector at 1,620 nm with the use of the recessed SiNx stressor can be comparable to that at 1,500 nm without stressor. More channels at the L-band could thus be incorporated into the PIC design using the WDM scheme for a higher data capacity.
Broadband strained Ge0.99Si0.01 EA modulator arrays
The ultra-broadband EA modulation is demonstrated on a Ge0.99Si0.01 on SOI platform via the design and fabrication of SiNx-strained waveguide-integrated Ge0.99Si0.01 modulator arrays. The reason for the ~1% Si incorporation in Ge is to position the broadband modulation with both tensile (~600 MPa) and compressive (~1 GPa) SiNx stressors within the available tunable laser spectrum (1,510–1,610 nm) of our measurement system. Fig. 4a shows a schematic of an individual modulator and its working principle utilizing the Franz-Keldysh (FK) effect. The inset shows an optical microscope image of a fabricated device. Waveguide-shaped Ge0.99Si0.01 mesas (~300 nm-thick, 50 µm-long) were placed on top of Si -OI waveguides (~250 nm-thick) with tapers on both ends for efficient optical coupling into and out of the modulators (Fig. 4b). The Si waveguides are cross-sectionally in rib shape (Fig. 4c (right)) with a slab thickness of 80 nm to allow for a bottom contact of the modulators and form the recessed trenches for the SiNx stressor for enhanced and uniform (Fig. 2d–f, h) transverse strain across the Ge0.99Si0.01 waveguide mesas. To facilitate the broadband modulation, the waveguide mesa width is varied (0.4, 0.7, 2.0, and 4.0 µm) to engineer the Ge0.99Si0.01 strain and, consequently, its bandgap and operating wavelength coverage in response to an identical SiNx stressor. As seen in Fig. 4c (left panel), a similar finite element simulation (Supplementary section S2) shows distinct transverse tensile strain () profiles ranging from ~0.20 to ~0.50% with decreasing Ge0.99Si0.01 widths from 4.0 to 0.4 µm, at an identical recessed tensile SiNx stressor of 600 MPa. In the device fabrication, the SiNx stressors were deposited via multi-frequency PECVD under controlled conditions (Supplementary section S4). A detailed description of the fabrication process can be found in Supplementary section S12. The dark current-voltage characteristics and a cross-sectional SEM image of a fabricated modulator are shown in Supplementary Fig. S17. To evaluate the modulator performance, optical transmission measurements (see Methods and Supplementary Fig. S18) were performed as a function of DC voltage bias. The lower maximum transmission at a wider Ge0.99Si0.01 width is due to the higher insertion loss (IL, ~5 dB) from the corresponding wider Ge0.99Si0.01 mesa (Fig. 4d). The extinction ratio (ER) spectra were then generated based on the transmission contrast of the modulators between 0 (on-state) and a reverse bias of -4 V (off-state) (Fig. 4e). A significant red-shift of the ER spectra is observed for the tensile-strained Ge0.99Si0.01 modulators with a decreasing Ge0.99Si0.01 width. The spectra ranges are consistent with the Ge0.99Si0.01 bandgap edge wavelengths from ~1,520 to ~1,570 nm at the corresponding widths, obtained from both the fitting of the optical transmission spectra (Fig. 4d) using the generalized FK model59, and the deformation potential theory calculation (Supplementary section S11) according to the respective profiles (Fig. 4c (left panel)). Similarly, an increasing compressive can be expected with a decreasing Ge0.99Si0.01 width, which explains the observable blue-shift of the ER spectra for the compressive-strained modulators. The spectra are not completely visible due to the limited measurement window.
The operating wavelength range of individual modulators is important for the arrayed broadband modulation. Here, the ER/IL ratio, as figure-of-merit (FOM), of 1.3 is considered to determine the modulator operating coverage. This is because, besides ER, IL also increases with decreasing wavelength due to the increased Ge-based direct-band absorption. The value of 1.3 is reasonable among the values (0.8–1.7) reported in prior studies24, 25, 26, 60 and foundry PDK standards for high-speed optical interconnect applications61. Using this criterion, the operating coverage of a tensile Ge0.99Si0.01 modulator (50-µm long, 0.7-µm wide) extends from ~1,580 to ~1,610 nm, where the IL is <6.8 dB. The operating optical bandwidth of individual modulators generally spans ~27.5 ± 3.5 nm at -6 V within the measurement window and is independent of the strain. The operating wavelength coverages for modulators beyond the measurement window were estimated based on the corresponding bandgap edge wavelengths obtained from the optical transmission measurements with the span of the determined average optical bandwidth. The individual spectra overlap and form an ultra-broadband wavelength coverage ranging from ~1,460 to ~1,620 nm at -6 V, where the compressive and tensile modulators cover from ~1,460 to ~1,530 nm and ~1,525 to ~1,620 nm, respectively (Fig. 4f). The modulator insertion loss can be reduced to ~2 dB via the optimization of top Si electrodes and taper coupling design62, which could further increase the ER/IL performance and reduce the operating voltage and modulation dynamic power of the modulator.
Broadband Ge-based modulator-detector co-integration
To investigate the feasibility of the Ge-based modulator-detector co-integration in a single step of Ge-based epitaxy (Fig. 1), a 1-GPa recessed compressive SiNx stressor was introduced to an array of Ge waveguide-mesas on Si-OI waveguides similar to the design in Fig. 4. Similar optical transmission measurements were performed. As expected, the transmission started to increase at a shorter wavelength with decreasing mesa width, indicating an increase of the Ge bandgap due to the higher compressive strain in a narrower mesa (Fig. 5a). The FK model fitting (Fig. 5a, solid curves) agrees well with the transmission spectra (circled dots) and extracts the strain and bandgap information in Ge (Supplementary Table 2). The ER spectra were then calculated (see Methods), as shown in Fig. 5b. The predicted operating coverage for the Ge modulator array ranges from ~1,520 to 1,630 nm. Overlaying with the absorption spectrum (Fig. 3d) of the tensile-strained Ge photodetector provides a broad wavelength window of over 100 nm for the Ge modulator-detector co-integration. The small footprint of the SiNx stressor (Supplementary section S7) provides the convenience to separately use tensile and compressive stressors on the photodetectors and modulator arrays, respectively, for the co-integration. Additionally, with an increasing Si incorporation, the bandgap of GeSi shrinks, which will shift the broad co-integration bandwidth to shorter wavelengths. Bandgap (see Methods) and FK model calculation finds that 1% Si incorporation will shift the co-integrated optical bandwidth range to ~1,460–1,575 nm, covering E- and C-bands (Fig. 5c). A 4% Si incorporation further shifts the co-integrated optical wavelength range to O- and E-bands (1,300–1,450 nm), where the FK effect remains significant since the direct-band transition dominates at the Γ-valley.