Integrated CMOS-compatible Q-switched mode-locked lasers at 1900nm with an on-chip artificial saturable absorber

: We present a CMOS-compatible, Q-switched mode-locked integrated laser operating at 1.9 µm with a compact footprint of 23.6 × 0.6 × 0.78mm. The Q-switching rate is 720 kHz, the mode-locking rate is 1.2 GHz, and the optical bandwidth is 17nm, which is sufficient to support pulses as short as 215 fs. The laser is fabricated using a silicon nitride on silicon dioxide 300-mm wafer platform, with thulium-doped Al 2 O 3 glass as a gain material deposited over the silicon photonics chip. An integrated Kerr-nonlinearity-based artificial saturable absorber is implemented in silicon nitride. A broadband (over 100 nm) dispersion-compensating grating in silicon nitride provides sufficient anomalous dispersion to compensate for the normal dispersion of the other laser components, enabling femtosecond-level pulses. The laser has no off-chip


Introduction
High repetition-rate ultrafast mode-locked lasers (MLL) have unique advantages for applications such as photonic analog-to-digital converters, comb spectroscopy, optical arbitrary waveform generation and low-noise microwave synthesis. Traditionally, repetition rates beyond 1 GHz were achieved by either active modulation techniques [1,2], which restricted the pulse duration to more than a few picoseconds, with nonlinearity-induced optical bistability where multimode noise suppression was necessary for a stable operation [3,4], or by introducing a semiconductor saturable absorber, in which case the pulse duration also remained more than a few picoseconds [5][6][7]. Alternatively, passive mode-locking techniques have been shown to generate femtosecond-level pulses at high repetition rates when used with some form of an external repetition-rate multiplier to bring the system into the GHz-level regime [8,9]. Harmonic mode-locking, where several pulses circulate in the cavity at the same time, has also been used to demonstrate high repetition rates [10]. In all of the above cases the architecture of the laser system itself was based on a fiber laser or a diode laser, with the rest of the cavity constructed with free-space optical components. Such systems could significantly benefit from an on-chip implementation which would pump/signal combiner component (Section 3.1) shown in the upper left in Fig. 1. The gain waveguide consists of three nearly straight, approximately 2cm-long sections where the pump and the signal modes have large overlap with the gain material (Section 3.2). These three gain waveguide sections are connected by compact Si 3 N 4 bends (Section 3.3). Prior to each bend, both the pump and the signal are transferred from a large-mode-area gain waveguide into a silicon nitride layer with a much more confined optical mode. This makes it possible to achieve tight 180-degree bends without significant losses. Following the three gain sections, the 1900nm signal is again transferred to the Si 3 N 4 layer as it enters a Kerr-nonlinearity-based mode-locking element, which we refer to as a nonlinear interferometer (NLI). The NLI acts as a CW reflector for 1900nm at low incident optical power. When the incident optical power is high enough, the NLI produces intensity-dependent reflection that can increase for higher input power (Section 3.5). This produces artificial saturable absorber action and can be used to form optical pulses. The 1900nm signal is then reflected back into the laser cavity, propagates through the three gain sections, and enters the pump/signal combiner element. The pump-signal combiner is a four-port device allows most of the 1614nm pump to couple directly through, while simultaneously coupling the 1900nm signal light across. The crossport of the pump-signal combiner has an integrated Si 3 N 4 -based grating, which serves both as a laser resonator-forming reflecting element and as a dispersion-compensating component (Section 3.4). The laser cavity is formed by the integrated grating on one end, an NLI on the other end, and three gain sections in between, connected with compact bends. The unique design of the bends together with the parameters of the gain waveguide allows for the compact footprint of this laser. All the lasers and test structures described in this work, with the exception of the rareearth-based gain material, were fabricated at Colleges of Nanoscale Science and Engineering (CNSE) at SUNY Polytechnic Institute in Albany, New York, on a 65nm CMOS 300mm wafer platform. The laser gain material used for 1900nm emission was deposited over individually diced chiplets at MIT's Microsystems Technology Laboratories.
The photonic layers for the full fabrication process are shown in Fig. 2. Starting from SOI wafers, photonic layers include three Si 3 N 4 layers, with 400nm, 200nm, and 200nm respective thicknesses. We refer to those layers, from the 400nm layer up, as "BN" for "Bottom Nitride", "FN" for "First Nitride", and "SN" for "Second Nitride". The top two Si 3 N 4 layers are separated by 100nm, allowing for the design of the waveguides where the optical mode at 1900nm will reside in both 200nm Si 3 N 4 layers simultaneously. A large 4µm-deep photonic trench is etched in SiO 2 over the gain region of the lasers, with the "SN" Si 3 N 4 layer serving as the etch stop. The gain material (thulium-doped aluminum oxide glass), deposited on top of the chip, partially fills in the trench to a thickness of 1.1µm.
To characterize the optical properties of the photonic layer materials, separate films of Si 3 N 4 were deposited and the refractive indices were measured using the prism-coupler technique at multiple wavelengths for each nitride layer. Index-vs-wavelength data were fit to the Sellmeier formula which in turn was used in the photonic design process. Linear losses of each silicon nitride layer were measured for each wafer by using ring-resonator test-structures implemented in corresponding layers [26]. Due to various design requirements (such as loss, effective Kerr nonlinearity, and dispersion), different MLL components are implemented using different available photonic layers from Fig. 2. Layer-to-layer transition components are designed to facilitate low loss light coupling between those layers. Table 1 lists the main MLL photonic components and the corresponding fabrication layers they are implemented in.

Pump/signal combiner
The purpose of the pump/signal combiner element is to allow the 1600nm pump light to enter the gain waveguide and to simultaneously guide the 1900nm light coming from the MLL to the chirped grating that serves as the end-reflector of the laser. This component was designed as a Mach-Zehnder interferometer with zero phase delay for 1900nm which consists of two directional couplers, each having 50% transmission at 1900nm. The layout of the pump/signal combiner is shown in Fig. 3(a), with a close-up of a single directional coupler shown as closeup underneath the layout. The devic identical dire accurate meas coupling loss device inserti 0.6dB, which signal couplin signal couplin for the devic dimensions.

Gain ma
The gain ma (Al 2 O 3 :Tm 3+ ) has been show provides optic film is depos plasma and a Two widely u the 3 [27,19], while ransition [28]. and thulium w e temperature the in-band pu umping of the tum-mechanica 00nm photons simplifies the r l chosen to velengths. hieve the inimizing elength is elength is e is 89%, e multiple allowing n/off chip measured port 2) is e 1900nm 5% of the ed values of device ide glass xide glass e thulium The gain ith argon of 415 ᵒ C. umping of e 3 H 6 -3 H 4 ally more s for each photonic design of the MLL. Many MLL components (gain waveguide, bends, photonic layer transitions) must work for both the pump and the signal simultaneously, which is significantly easier to achieve when the pump and the signal wavelengths are close to each other. Therefore, MLLs in this work were designed for 1600nm pumping. Al 2 O 3 :Tm 3+ films were characterized spectroscopically to obtain the upper-state lifetime, absorption cross-sections, background losses, and active ion concentration [29]. These parameters were subsequently fed into a comprehensive laser gain model that was used to design the gain waveguide cross-section and to optimize the length of individual gain sections. Separate gain films with various thicknesses and active ion concentrations were deposited on SOI wafers to be used for spectroscopic characterization. In order to achieve different active ion concentrations, different levels of RF power were applied to thulium targets in the deposition chamber. The Rutherford back-scattering technique was used on each of those films to measure the active ion concentration. Absorption cross-section was extracted from the loss measurements on active films using a Metricon 2010/M prism coupler as a function of wavelength (for both 790nm and 1600nm) using the known active ion concentration values. Film background losses were measured with a prism coupler at 633nm, which is outside of absorption/emission of thulium, and were used as upper bound values. Upper state lifetimes were measured by characterizing the amplitude and phase response of a 1900nm signal coupled through the waveguide with a large mode fraction in the gain material, as a function of the frequency and amplitude of the modulated 1600nm pump light [29]. A summary of the relevant spectroscopic parameters is given in Table 2. Upper state spontaneous emission lifetime, τ u 568 ± 48 µs Absorption cross-section (at 1614nm), σ a 2.75 × 10 −21 cm 2 Absorption cross-section (at 790nm), σ a 6.5 × 10 −21 cm 2 Film background losses, α b 0.1 dB/cm Thulium ion concentration*, N t 2.9 × 10 20 cm −3 *Thulium ion concentration is given for the actual film used in final demonstrated MLLs The cross-section of the gain waveguide is shown schematically in Fig. 4(a). It consists of a 5-piece, 200nm thick segmented silicon nitride waveguide, a 100nm oxide gap layer, and a 1.1µm thick layer of Al 2 O 3 :Tm 3+ . A 4µm-deep trench is etched into the SiO 2 layer, over the entire active area of the chip to allow the gain material to be within 100nm of the Si 3 N 4 layer. The pump and signal modes are guided by the silicon nitride pieces, but the majority of the optical modes reside in the thulium-doped aluminum oxide layer. The waveguide was designed to optimize the overlap of the pump and signal modes with each other and with gain material, and to ensure sufficient pump intensity in the gain material to invert the population of active ions. The Si 3 N 4 waveguide design was chosen to be segmented because this reduces the effective index of the signal and pump modes, and thus increases the mode fraction inside the gain region. The optimized gain waveguide has a 60% pump-signal overlap in the gain material. The pump and the signal intensity modes are shown in Fig. 4(b) and 4(d) respectively. The mode confinement factor in the gain material is 83% for the 1614nm pump, and 79% for the 1900nm signal for the optimized gain film thickness of 1.1µm. To measure the small-signal gain of the gain waveguide described above, a separate gain test structure was added to each chip. This test structure included three sections of the gain waveguide interconnected with compact bends as shown in Fig. 1, but with pump/signal combiners at the input and output of the chip and no mode-locking or dispersioncompensating elements. The pump and a low level of signal were coupled into the gain waveguide and the output signal enhancement was measured as a function of pump power. The insertion losses of the pump/signal combiners, transition elements, and on/off chip coupling losses were calibrated out. The results of the gain measurement are shown in Fig.  4(c).

Compact bends
Compact bends at the end of the gain waveguide sections were designed to bend both the pump and the signal light by 180 degrees. In order to minimize the bend radius, prior to the bend the optical mode is shifted from the main gain waveguide to the bottom silicon nitride layer by an adiabatic taper transition. First, the segmented FN layer Si 3 N 4 pieces composing the gain waveguide of Fig. 4(a) are tapered to form a single, wider Si 3 N 4 piece, which is shown in Fig. 5(a). This is done by slowly moving the four nitride side blocks away from the central one, while increasing the width of the central nitride piece. Next, the width of this central FN layer piece is slowly decreased while the width of the bottom 400nm thick BN layer is slowly increased. The higher effective index of the BN layer pulls the optical mode down until it is mostly confined in that layer, as shown in Fig. 5(b). Insertion loss for this transition component was measured by a variation of the cut-back method, where 20, 30 and 40 such transitions were nested together back-to-back. Measured loss per each such transition is 0.02 dB at 1900nm. Once the 5(c)). The be linearly as a f is the angle, a abrupt transit waveguide cro bend radius a 0.04dB for bo in the accumu

Dispersio
One of the implemented signal and as locking, net M pulses [31-33 therefore, the for the net no mode-locking The key d of bandwidth this bandwidt This integ thickness allo larger anomal achieved by a schematically introduced by width constan necessary to apodization p and allows th of the grating and group del designs with three gratings two gratings 5 Figure 6(d) sho d position with own in Fig. 6 100nm, and the grating with −30,000fs 2 has reflection over 90% across the same wavelength range. Larger anomalous GDD values result from longer gratings, which also provide stronger reflection (smaller insertion loss); therefore, there is a fundamental trade-off between the amount of anomalous dispersion and insertion loss. All three designed gratings have dispersion values which put the net dispersion of the MLL into the anomalous regime and provide more-than-sufficient bandwidth to support pulse durations of under 50fs.

On-chip mode-locking element
The mode-locking element in the laser is an integrated nonlinear interferometer device, which is based on the Kerr nonlinearity of silicon nitride [18,34]. The schematic of the device is shown in Fig. 7(a). The device has an input coupler with a 90/10 splitting ratio at 1900nm between its two arms. Each arm consists of a 9.8mm long section of Si 3 N 4 waveguide and is terminated with a near-perfect reflector, implemented as an integrated loop mirror. The loop mirrors reflect 99.8% of the incident light back into the interferometer arms and the light is recombined at the coupler. At low power the 90% output of the device goes back into the laser while the 10% port is used as the MLL output. The two arms of the interferometer have identical lengths, cross-sections, and loop mirrors. Therefore, at low incident power the nonlinear interferometer acts as a CW reflector forming the right side of the laser resonator cavity. At higher incident power a differential nonlinear phase shift forms between the two arms of the device due to the difference in their powers and Kerr nonlinearity. This nonlinear phase shift is proportional to the effective nonlinearity of the waveguide, the length of the waveguide, and the power difference between the two arms. The output of the nonlinear interferometer in this case has a sinusoidal-like dependence on the input power that, with proper linear bias, can provide higher reflection for higher incident power and act as a fast artificial saturable absorber.
To ensure the proper linear phase bias, integrated titanium-gold micro-heaters were deposited along the two arms after the gain medium deposition. The heaters are 20μm wide, 8mm long, with a thickness of 150nm of titanium followed by 100nm of gold. The average resistance of a heater is 180 Ohms. By applying current to one heater, linear phase bias is introduced to a corresponding device arm due to thermo-optic coefficient of silicon nitride. This device is implemented with a combination of the FN and SN silicon nitride layers, which allows a waveguide cross-section with maximized effective nonlinearity and minimized dispersion. The cross-section of the optimized waveguide, along with the optical mode, are shown in Fig. 7(b). Longer interferometer lengths allow the accumulation of higher Kerr nonlinearity but also decrease the reflected power due to accumulated linear losses of the silicon nitride waveguide. Lower power at the output of the interferometer results in significantly reduced CW laser power in the cavity. Therefore, the device length is optimized to provide sufficient reflection back into MLL cavity, while also providing a sufficiently large self-amplitude-modulation coefficient (the slope of reflection vs peak power) and modulation depth for the expected incident power during laser operation. Simulated transmission and reflection of the device are shown in Fig. 7(c) for two different device lengths -23mm and 9.8mm (one-way length). For 23mm-long devices the simulated modulation depth is 12.3% and self-amplitude-modulation coefficient is 3.7 × 10 −4 W −1 . For 9.8mm-long devices, the modulation depth is 19.2% and the self-amplitude-modulation coefficient is 2.8 × 10 −4 W −1 . The 23mm devices were characterized as separate nonlinear interferometer test structures using femtosecond optical pulses from a commercial optical parametric oscillator centered at 1900nm. The measured modulation depth of the device was 9% (compared with the theoretical value of 12%), and the self-amplitude-modulation coefficient extracted from measured data was 7 × 10 −5 W −1 [16]. The discrepancy between the design and the measured values are likely due to underestimating the waveguide dispersion. Higher normal dispersion will result in greater pulse spreading inside of the nonlinear interferometer, thereby decreasing the peak pulse power and reducing both the effective self-amplitude-modulation coefficient and modulation depth.

Laser mode-locking dynamics
We investigate the steady-state mode-locking characteristics of the laser by using a Nonlinear Schrodinger Equation (NLSE)-based model of the complete MLL. For each integrated photonics sub-component in the MLL cavity, the parameters such as gain/loss, Kerr nonlinearity, dispersion, and effective mode area, are measured and calculated. Those parameters are then put into an NLSE model based on the following equation: where A(z,t) is field amplitude, defined such that |A(z,t)| 2 is the optical power, β 2 is the dispersion in fs 2 /m, γ NL is the effective nonlinearity in (W-m) −1 , l is the linear loss in units of m −1 , g is the energy-dependent electric field gain in units of m −1 , where power gain is defined as g = g 0 /(1 + E sat /E), where E is pulse energy at any given time and g 0 is small signal gain. Table 3 lists all the MLL components used in the model, with their respective parameters. The NLSE is numerically solved in a slowly-varying envelope approximation for each successive MLL component. The result of a stable and well-designed MLL is a steady-state pulse which circulates inside the laser cavity. It should be noted that this analysis does not take into account the laser start-up or laser gain dynamics that would have to be included to describe Q-switching behavior. The flow of the model is shown in Fig. 8(a), with components in order of appearance in the MLL cavity (from Fig. 1), starting from the chirped grating side. The model includes the chirped grating, pump/signal combiner, three near-straight sections of gain waveguide with connecting bends, a small section of BN waveguide connecting the last section of gain waveguide to the NLI, and the NLI itself. The dispersion of the chirped grating used for this model is −38,000fs 2 , which makes the net dispersion of the laser −5,910fs 2 -in the net anomalous regime as shown in Table 3. The NLSE model for this laser converges on a steady state solution. The resulting pulse duration dynamics as a function of position along the laser cavity are shown in Fig. 8(b). Since most of the components in the laser have normal dispersion, the pulse is up-chirped when it is incident on the grating. The grating provides a large amount of anomalous dispersion which over-compensates for the normal dispersion of the rest of the cavity. The pulse exits the grating strongly down-chirped, and as it propagates through the gain material towards the NLI the normal dispersion of the gain waveguide slowly counteracts the negative chirp, shortening the pulse as it propagates towards the NLI. Compact bends, although short compared with the gain waveguide, have a larger amount of normal dispersion per unit length, and this is evident in the sudden slope changes on the pulse duration plot in places that coincide with bend locations. The NLI provides a large amount of normal dispersion, and the pulse is again up-chirped as it exits the nonlinear interferometer. The pulse duration changes by over 100fs as it circulates within the laser, with pulse duration being the shortest when the pulse is incident on the NLI, and longest when the pulse is incident on chirped grating. The output of the MLL is directly after the NLI -not quite at the point of shortest pulse duration. This suggests that a better MLL design should have more uniform dispersion distribution within the laser cavity. Future designs include redistributing the anomalous dispersion by placing gratings at both ends of the MLL cavity -within NLI itself as well as after the pump-signal combiner. This way the pulse duration will be more constant within the laser and the output of the NLI could be designed to have the shortest duration.

Results
The lasers were characterized using a setup shown in Fig. 9(a). The pump laser was an Lband EDFA, seeded with a tunable laser at 1614nm. The pump was delivered to the chip through a fiber-based polarization controller and a lensed SMF28 fiber with a 3µm spot size at 1550nm wavelength. The signal from MLL was collected using SM2000 fiber with a lensed tip with a spot size of 3µm at 2µm wavelength. The output was split between two paths, with one path being directed into an optical spectrum analyzer (OSA), and another part going into the fast photodetector (12.5GHz electrical bandwidth) that was connected to an oscilloscope and an RF spectrum analyzer. A photo of an individual chip is shown in Fig.  9(b). Each chip contains 9 full MLLs and numerous other test structures. Titanium/gold heaters, deposited on top of NLI arms, are visible on top of each MLL.  At about 170mW of pump power, the lasers go into a Q-switched and mode-locked (QSML) regime. Maximum achieved onchip outside-of-the-laser-cavity average MLL power is 9mW, which is limited by available pump power. Figure 10(b) shows the optical spectrum of the laser in the QSML regime. The central wavelength of the laser is around 1880nm, which is a result of both the emission spectrum of thulium and the wavelength-dependent losses in the laser cavity. The fit for a sech 2 pulse is overlaid on top of the spectrum, with a 17nm 3dB bandwidth which could support pulse sweep the lase Figure 11 switching rate one Q-switch envelope. A individual mo 1.2GHz. This

Conclusion
In this work, we have demonstrated Q-switching and mode-locking in an integrated laser operating near 1900nm in a silicon nitride-on-insulator platform, with thulium-doped aluminum glass as a gain material. The laser is fabricated using a CMOS-compatible 300mm wafer platform, with the gain material deposited separately over SiN/SiO 2 layers. The laser, with a footprint of 23.6 × 0.6mm, is completely integrated on a chip, with no off-chip components with the exception of a 1614nm pump laser. Maximum on-chip signal power is 9mW, with the fundamental repetition rate of 1.2 GHz and Q-switching rate of 720kHz. The laser produces optical pulses centered at 1880nm, with 17nm optical bandwidth which is sufficient to support pulses as short as 215fs. This work is a major step towards all-integrated CMOS-compatible CW-mode-locked lasers. With suppression of Q-switching instabilities and on-chip pump integration, this laser architecture, together with electronic-photonic integration, could enable stabilized MLL-based on-chip frequency combs, low phase noise microwave sources, and high speed communications applications, fabricated with a compact footprint using CMOS and CMOS-compatible processes.