Optimization of photoluminescence from W centers in silicon-on-insulator

W centers are trigonal defects generated by self-ion implantation in silicon that exhibit photoluminescence at 1.218 $\mu$m. We have shown previously that they can be used in waveguide-integrated all-silicon light-emitting diodes (LEDs). Here we optimize the implant energy, fluence and anneal conditions to maximize the photoluminescence intensity for W centers implanted in silicon-on-insulator, a substrate suitable for waveguide-integrated devices. After optimization, we observe near two orders of magnitude improvement in photoluminescence intensity relative to the conditions with the stopping range of the implanted ions at the center of the silicon device layer. The previously demonstrated waveguide-integrated LED used implant conditions with the stopping range at the center of this layer. We further show that such light sources can be manufactured at the 300-mm scale by demonstrating photoluminescence of similar intensity from 300 mm silicon-on-insulator wafers. The luminescence uniformity across the entire wafer is within the measurement error.


I. INTRODUCTION
demonstrated 12 , and we have previously demonstrated waveguide-coupled W-center LEDs in a cryogenic optical link with superconducting single-photon detectors 20 . The W center emits at a convenient wavelength for silicon photonics. The zero phonon line for the W center is centered at 1.218 µm, which falls below the silicon indirect bandgap while being above the SiO 2 phonon band.
Despite previous studies in a SOI substrate 20,21 , an in-depth study of the optimal conditions for maximizing light emission from ensembles of W centers in SOI has not been performed. SOI is the workhorse for silicon photonics, as the mode confinement provided by the buried oxide layer is necessary for waveguides. Previous waveguide integrated LEDs exhibited an external efficiency 20 of 5×10 −7 , while similar measurements on W-center LEDs in bulk silicon yielded external efficiency 12 of 10 −6 . The efficiency value in Ref. 20 includes losses due to the poor waveguide coupling, the device resistance, and the low quantum efficiency of the waveguide-coupled diode itself. Here, we show that the photoluminescence (PL) intensity can be improved by over two orders of magnitude through optimization of the implant conditions. In particular, we have studied the effect of implantation energy, fluence, and annealing conditions for optimization of PL from SOI wafers. We expect that this will translate into a similar increase in efficiency in electrically injected LEDs. This study may also provide helpful information to those hoping to further understand the properties of W centers in silicon.

II. CHARACTERIZATION AND SETUP
Implants were initially performed using a commercial ion implantation service in three different wafer types: (A) 1 Ω·cm to 10 Ω·cm 76.2 mm p-type (boron doped) silicon, (B) > 10000 Ω·cm 76.2 mm undoped silicon, (C) p-type (boron doped) SOI 76.2 mm with a 220 nm silicon layer on 3 µm buried oxide. Wafers were cleaned in a sulfuric acid solution followed by hydrofluoric acid and a de-ionized water rinse. A 7 nm thermal oxide was subsequently grown on the wafers before implant. All implants were performed at room temperature at 7 • off normal to prevent channeling. Samples were subsequently annealed in N 2 ambient at 250 • C for 30 minutes unless otherwise stated. Initial characterization of the W centers was performed in a continuous-flow cryostat with a minimum temperature of 4.2 K. The samples were pumped using a continuous-wave (CW) HeNe laser at 632 nm through a 0.6 NA 20× objective lens with a 20 mm working distance, unless otherwise noted. The W center PL was collected through the same lens. The PL was then passed through a grating monochrometer and detected on a liquid-nitrogen-cooled linear InGaAs photodiode array. Figure 1 (a) shows a typical spectrum from the p-type bulk silicon (wafer type A) implanted at 80 keV with a fluence of 5×10 13 /cm 2 . The time-dependence of the PL when pumped with a 672 nm pulsed laser with a 1 MHz repetition rate is shown in the inset. A fit to the decay indicates a total (radiative and non-radiative) lifetime of (34.5 ± 0.5) ns (the error is estimated from the standard error of the fit). For the lifetime measurement, light from the sample was fiber-coupled to a superconducting-nanowire single-photon detector and correlated with the pump laser trigger. This lifetime is fast enough for superconducting optoelectronic neuromorphic computing 4 . For other applications that require higher speed, the lifetime may be decreased through engineering of the local density of optical states 22 or by using other silicon-based emitters with faster intrinsic lifetime 23,24 . The W center PL also exhibits a strong temperature dependence, with the PL intensity decreasing sharply around 45 K, as shown in Fig. 1 (b). The data also shows a decrease in PL intensity for the point at 5 K. This drop in PL intensity at low temperatures has been observed previously to be sample dependent 6 , and can be explained by free carriers becoming captured in shallow traps at very low temperatures. For operation below 4 K, as required for SNSPDs, this is a concern that should be addressed with measurements at lower temperatures and with different substrates. Measurements presented later in this work were performed in a cryostat with a minimum temperature of around 20 K as the PL intensity is relatively flat in this region (see supplementary information). The error bars given throughout the paper are two standard deviations (2σ), where the standard deviation is a fixed percentage of the mean calculated from repeated measurements of certain samples. If several chips from the same wafer were measured in one cool down, the standard deviation over that cool down was used to calculate the error bars. A discussion of the sources of error and the error bars throughout the paper is given in the supplementary information.
A known property of emitters such as quantum dots and semiconductor defects is that the PL intensity saturates at high pump power. To compare the PL intensity of different samples, it is important to operate in the non-saturating regime.  (c) (see Supplementary Fig. 2 (b)) are slightly less than 1 for most of the samples. The slope increases with higher fluence, suggesting that lower fluences may start to saturate by 300 µW. We note that the saturation power is high since we are measuring ensembles of W centers. The saturation power is likely much lower for a single W center.

III. DEPTH DEPENDENCE
Ions implanted in a substrate will have an average projected range dependent on the energy of the implant, as shown in Fig. 2 (a). Many other atoms in the silicon lattice are displaced during the implantation. The displaced and recoiled ions can settle as interstitials well beyond the projected range, leaving vacancies in their paths. The post-implant anneal allows these interstitials to migrate and form W centers. The intensity of the PL depends on a competition between the rate of the radiative recombination at the W center with the rates of other non-radiative processes at other defects 12,19,25 . A review of the literature (see Table 1) indicates that while W center formation peaks at the projected range of implanted ions, PL intensity peaks much deeper. The literature survey also indicates that there is a strong fluence and energy dependence on the depth at which the PL intensity is maximized, and further study is needed for optimization in a particular substrate, such as SOI.
The W center is thought to be a particular configuration of n = 3 clusters of silicon interstitials 19 . Many other clusters of interstitials and voids are formed during the implantation process. Therefore, it is not only the total number of W centers that are formed that is important, but the number of W centers relative to other centers. In particular, there is evidence that n > 5 clusters strongly quench the W center PL 19 . The strong dependence of stopping range on implant energy and the complicated dependence of the number of clusters of n silicon interstitials on implant conditions and depth suggests that it is important to investigate the PL intensity dependence on the energy and fluence of the implanted ions for a 220 nm thick device layer, the most common choice for integrated photonics using SOI.
The stopping ranges for silicon in silicon, calculated using the software SRIM 26  The i-region, where the W centers should be located for optimal performance, was 1 µm wide. The p-region, located vertically above the i-region, and through which the W centers were necessarily implanted, was 300 nm wide. This means the center of the W center distribution should be ≈ 800 nm deep for optimal performance. Earlier studies [27][28][29] ( Table I) indicated that under similar implantation and annealing conditions the W centers were mostly located much deeper than the projected range (110 nm), with estimates of up to 500 nm.  Table I refers to where the center of the W center distribution would be located for optimal device characteristics (e.g. the center of the p-i-n junction for an LED). Figure   2 also includes the stopping range profile for silicon implanted with 150 keV energy through a 100 nm capping oxide. This is the implant condition in Ref. 20 where waveguide-coupled LEDs were fabricated in 220 nm device layer SOI (geometry shown in Fig. 2 (b) (ii)). The projected range was designed to be 110 nm, the center of the silicon device layer, when the top oxide layer is removed. However, no etch-back study of the PL was performed in this case, and based on the prior work this is likely not the optimal depth for W center PL, as  [20] p-type Si handle wafer W 300 nm 1 µm 2.8 µm (i) Ref. [12] (c)  will also be demonstrated later in this paper.
Based on the mechanism discussed at the start of this section for W center luminescence 19 , the depth dependence could be explained as follows. Despite the fact that W centers are formed at the highest density near the projected range, the n > 5 silicon interstitial clusters cause the W center PL to be quenched until much deeper. These n > 5 clusters are also formed at the highest density near the projected range, but their numbers fall off more steeply with increasing depth. An alternate explanation involving a competitive process with voids has also been proposed 25 , but the depth dependence is not clearly explained by this process.
To test this depth dependence, 28 Si + ions were implanted at 40 keV, 80 keV and 120 keV in the three different substrates through a screen oxide of around 5 nm (5 nm   measured). Implants were made on full wafers with a fluence of 5×10 13 /cm 2 . The initial results are shown in Fig. 2 (c). We observe that for both the p-type bulk silicon (wafer type A) and intrinsic bulk silicon (wafer type B) there is not a strong dependence on the implant energy. The intrinsic silicon (wafer type B) is brighter by a factor of around 1.5 than the low resistivity p-type silicon (wafer type A). It has been noted previously that boron doping has a quenching effect on the PL 31 , and this is likely the cause of the difference in PL intensity.
However, there is a strong energy dependence in the SOI sample. We observe that SOI C has a maximum PL intensity for the lowest (40 keV) energy.

IV. OPTIMIZATION
The data shown in Fig. 2 and Table I indicates the potential for further optimization of the energy, fluence and anneal conditions for the implants, particularly for SOI wafers where the silicon layer thickness is chosen to be 220 nm for waveguide formation. For this study, a single SOI wafer (wafer type C) was diced into 1 cm die, which were attached to silicon carrier wafers with poly(methyl methacrylate). Implants were then performed on these die.
After implantation, each 1 cm die was removed from the silicon carrier wafer, and diced into 2.5 mm die so the PL from multiple die could be compared in a single cooldown. The 2.5 mm die were annealed in batches for a single experiment. The remainder of the measurements were performed in a custom-built continuous-flow cryostat with optical access. This cryostat operated between 24 K and 32 K. As can be seen from the temperature dependence plot in Fig. 1 (b), this is close to the optimum temperature for PL intensity from W centers.
The optical setup for the remainder of the experiments used a 0.42 NA 50× objective. The measurements are normalized to the bulk p-type silicon sample implanted at an energy of 40 keV and fluence of 5×10 13 /cm 2 . The exact fluence at which the PL intensity peaks depends on the energy of the implant, indicated in Fig. 3 (a). This saturation of intensity with fluence has also been previously observed in the literature 25 but at much higher implant energies of 1 MeV. That study found that the W center PL was only proportional to fluence for fluences from 10 8 /cm 2 to 10 10 /cm 2 , two orders of magnitude lower than in this study. This finding is consistent with the trend in Fig. 3 (a), where it appears that the optimal fluence is lower for higher energy implants. This is likely due to the fact that the number of n-interstitial clusters formed is a non-linear function of fluence, with a larger ratio of high-n to low-n clusters formed at high fluences (for the same implant energy). This leads to the optimal ratio for PL occuring at a larger depth 19 .
Next we consider annealing conditions. It has been reported 29 that an anneal temperature of 265 • C gives the optimal PL intensity. However, the test in Ref. 29 was performed in bulk silicon for an implant energy of 80 keV and a fluence of 5×10 13 /cm 2 . Due to the competitive nature of the PL process, we considered that the anneal conditions might also depend on the fluence or energy of implantation. Therefore we annealed different samples at different temperatures to see if the peak in PL intensity versus anneal temperature depends on implantation fluence or energy shifts for different samples. This is shown in Fig. 4

(a) and (b) for different energies and fluences. Unlike Figs. 2 and 3, in this case the data has been
normalized to the maximum intensity for that implant condition. The relative intensities of the peaks can be seen in the previous figure (Fig. 3). The peaks do not appear to be significantly different, suggesting that there is no strong dependence on fluence or energy for the optimal anneal conditions.
The annealing data can be used to extract the activation and deactivation energy of the W center formation process. The activation and deactivation energy refer to the energy of the rate limiting step in the formation and decay of the photoluminescent W center.
The deactivation energy is most likely the decay energy of the W center cluster, but it could alternatively be the energy of formation of a strong competitive non-radiative center.
Reference 32 found that the activation energy of the W center is 0.95 ± 0.05 eV and the deactivation energy is 1.2 ± 0.05 eV for W centers implanted at a fluence 4×10 12 and energy 1 MeV 32 . A similar study 29 found an activation energy of 0.85 ± 0.05 eV for an implant energy of 80 keV and fluence of 5×10 12 /cm 2 , which also indicates that the activation energy is stable over a wide range of implant energies (deactivation energy unreported). The fact that previous studies have found the activation and deactivation energies to be very similar to each other suggests that it is the energy for the formation and decay of the center itself.
A thorough discussion of the activation and deactivation energies of the W center is found in Ref. 32, although the precise mechanism for the formation/decay process is still unknown.
The (de)activation energy is found from the slope of the fit to an Arrhenius plot 33 of ln(k) , where E D is the deactivation energy, T is the anneal temperature, R is the gas constant, and k is the PL intensity. A is a constant for the W center formation process. We show this fit in Fig. 4 (c) for the deactivation energy, where we obtain an average deactivation energy of (0.9±0.1) eV. The error is calculated from the standard error in the slope. The errorbars do not quite explain the deviation of the data from the fit. A detailed discussion of the calculation of the errorbars is given in the supplementary information. The data points in Fig. 4 (c) are averages over the data points in Fig. 4 (b), but we have also calculated the activation energies from the  of 40 keV and a fluence of 5×10 13 /cm 2 . The maximum calculated deactivation energy was (1.0±0.1) eV for an implant energy of 25 keV and fluence of 5×10 13 /cm 2 . This variation is likely measurement error, as there was no trend with either energy or fluence in the calculated value. We do not have sufficient data for low anneal temperatures to fit the activation energy. We note that the 10 keV sample shows an anomalous anneal curve shape, but it is not clear if this is significant.
To produce electrically injected optical devices, it is necessary to mask off the implants that produce W centers. We therefore studied the masking properties of several resists.
Three additional silicon wafers were implanted at 25 keV and fluence 5×10 12 /cm 2 . The wafers had 600 nm thick electron-beam resist, 1 µm thick photoresist and 3 µm thick photoresist. We found that the electron-beam resist was fully removed with 30 minutes heated N-methyl-2-pyrrolidone followed by 10 minutes soak in a sulfuric acid solution. The photoresists were fully removed with 5 minutes sonication in acetone followed by 2 minutes sonication in isopropyl alcohol and 10 minutes soak in a sulfuric acid and hydrogen peroxide solution. We did not observe PL from the areas of silicon that were masked by any of the resists, which allows us to use any of these as a lithographic masks for the implants. Finally, it was reported 34 that high temperature implants in silicon lead to significantly less lattice damage and consequently fewer nonradiative channels. To test how this affects PL from the W center, a final wafer was implanted at 265 • C (and not subjected to a post-implant anneal), with an energy of 40 keV and fluence 5×10 13 /cm 2 . The hot implanted wafer showed a factor of two increased PL over a wafer annealed post-implant. While this is promising for future improvements, more advanced fabrication is necessary to mask hot implants, as typical photoresists cannot withstand processing at this temperature.

V. W CENTER LIGHT SOURCES WITH 300 MM CMOS-FRIENDLY PROCESSES
To generate low-cost on-chip light sources, these devices must be fabricated in a conventional foundry process. We take the first steps towards this goal by demonstrating that the implantation can be done at a 300-mm-wafer scale. These implants were done at the cleanroom at the State University of New York (SUNY) Polytechnic Institute. In this study, ions were implanted at 40 keV, 80 keV and 120 keV at an angle of zero degrees to normal, with a fluence of 5×10 13 /cm 2 , in 300 mm SOI wafers with buried oxide thickness of 145 nm.
The implants were performed through a thick barrier oxide, with the intention of leaving the high-damage sections in the oxide and removing them with hydrofluoric acid after annealing. The oxide deposition, anneal and oxide removal were all performed on full wafers using standard 300-mm process tools at SUNY Poly. Ten wafers were implanted at three different implant energies (40 keV, 80 keV and 120 keV) through three different oxide thicknesses (120 nm, 150 nm and 200 nm). The fluence was again 5×10 13 /cm 2 . In addition, a single wafer was implanted with a 5 nm screening oxide at 40 keV, still at zero degree tilt. PL measurements were performed in the same method as in the previous section, and normalization was done relative to the same bulk silicon sample. Because the top layer of the wafer contains significant damage after implantation, we hypothesized that a thick barrier oxide could capture that high damage region for later removal. PL from high-fluence-generated W centers has been observed to increase in brightness after etching away of the top layer 27,29,32 . that W centers have also been formed in the handle wafer and are contributing to the PL.
Further etch-back studies could elucidate this effect.
A primary goal of this portion of the study was to test the uniformity of W center creation across the 300 mm wafer. The wafer implanted at 40 keV through a 5 nm barrier oxide was diced into 1 cm squares, and die were selected from equidistant points along the equator and meridian. Each of these 1 cm die was then diced into 2.5 mm die. The measurement was performed with three 2.5 mm die from each 1 cm square. The variation in the measurements from within 1 cm was compared to the variation across the entire wafer, with a total of 15 2.5 mm die compared in each cooldown (Fig. 5 (b)) No significant variation was observed across the wafer (equator standard deviation 4%, meridian standard deviation 11%). Interestingly, die from a previous wafer implanted with the same conditions but annealed at NIST showed consistently higher PL intensity (labeled NIST anneal on Fig. 5 (b)). This discrepancy may be explained by a difference in furnace temperature calibration. We observe approximately 4% variation in PL per degree around the optimal anneal temperature. Therefore this discrepancy could be explained if the anneal chambers were inadvertently operating at slightly different temperatures. The difference between the equator and meridian samples is due to the fact that the error between cooldowns is higher than the error between samples measured in the same cooldown. This is possibly due to differences in the temperature of the cryostat or the optical apparatus.

VI. FURTHER WORK
We have investigated the PL intensity from W centers implanted in the standard SOI substrate used for silicon integrated photonics, with a 220 nm device layer thickness. We find that the PL is strongly dependent on the energy of the implant. We find optimal photoluminescence for the implant conditions of 5×10 12 /cm 2 fluence and 25 keV energy.
Based on the data presented here, it is unlikely that previous attempts to use the W center as a light source have used optimal implant and anneal conditions. It may be possible, using some of the results of this study, to significantly improve the brightness of W center based silicon light sources. While there have been demonstrations of cavity-coupled PL from silicon defects in the past 14,22 , a direct comparison of the brightness of these emitters to the W center has not been made. It is possible that the W center could demonstrate much stronger cavity-coupled luminescence. The one previous known attempt to cavity couple the W center luminescence used suboptimal implant conditions (implanting with 100 keV energy for a membrane thickness of 220 nm) 35 . For applications in superconducting optoelectronic neuromorphic computing 3,4 , total light-production efficiency of 10 −4 for LEDs operating at 4K is sufficient to enable power-efficient, large-scale systems. In that context, light production efficiency greater than 1% provides little advantage 36 , as photon detection with superconducting detectors dominates the energy budget at that point. The previous demonstration with all-silicon waveguide-integrated LEDs showed a system efficiency of 5×10 −7 . If the optimal conditions remain the same in the case of electroluminescence, we expect an efficiency of 5×10 −5 by only changing the implant conditions. Significant further gains may be achieved through a combination of improved electrical injection and improved coupling of the W centers to the optical mode.
It remains to be seen how the implant conditions affect electroluminescence in LEDs. It has been observed that the series resistance in LEDs increases with fluence in the LEDs. Therefore, it is possible that there is a tradeoff in electroluminescence intensity for higher fluence. It is also possible that if the W centers are not uniformly distributed in depth in the device layer, current will preferentially flow through regions with a lower density of W centers due to decreased resistance, leading to a trade-off between coupling to the optical mode and electrical injection that is not present in the PL case.
Beyond W centers, there are numerous other luminescent centers 6 in silicon, and some of these may be brighter or more suitable in other ways for various applications. A systematic study of the relative brightness of these centers in silicon has not yet been performed. It is also likely that the implant conditions for these centers must also be optimized. For example, electroluminescence of the G-center in silicon has been observed 11 , and G-centers have been fabricated via ion implantation of C followed by proton irradiation 37 . Meanwhile, studies on the depth dependence of the G center PL 27 have indicated that the G centers are formed at even larger depths relative to the projected range than the W center. This indicates that for fabrication of G center LEDs in SOI via ion implantation, a similar study to this paper is required.
There has also been renewed interest in defects in silicon as solid-state spin qubits. While solid-state spin qubits are highly stable with extremely long coherence time, coupling of these qubits has remained a challenge. Photonically addressable spin qubits can provide a method for scaling to quantum networks 38 . Isotopically pure silicon is a strong candidate for these networks, if a photonically addressable spin qubit can be found. There has been recent interest in chalcogenide 39 and magnesium 40 defects coupled to cavities for these cavity-QED applications. The spin properties of the W center have not been fully examined 15 , and it remains to be seen if W centers can act as single photon sources. An on-chip electrically injected single photon source that could be easily coupled to silicon photonic integrated circuits would also have a variety of applications in quantum optics.

ACKNOWLEDGMENTS
We thank Dr. Maria Aboy Cebrian at the University of Valladolid and Dr. Jeff Chiles and Dr. Matt Brubaker at NIST for helpful conversations and insights. We thank Mr. Ronald Bourque and his colleagues at TEL Technology Center America for help with processing at certain steps. Contributions to this paper by SUNY Poly co-authors require the statement that this material is based on research sponsored by the Air Force Research Laboratory under agreement number FA8750-1-1-0031. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the Air Force Research Laboratory or the U.S. Government. This is a contribution of National Institute of Standards and Technology (NIST), an agency of the U.S. government, not subject to copyright.