Efficient chip-based optical parametric oscillators from 590 to 1150 nm

Lasers operating at visible and near-infrared (NIR) wavelengths are essential to modern science and technology,^1–4 but affordable systems typically suffer from poor spectral purity and gaps in spectral coverage, while higher-performance options are large and expensive. The latter often rely on bulk nonlinear optics to spectrally translate longer-wavelength lasers to the targeted frequency, employing either sum-frequency or second-harmonic generation in χ⁽²⁾-nonlinear media.^5–7 Their operational complexity and substantial power consumption (they often require liquid cooling systems) render them impractical in many situations. Hence, it is desirable to transition the nonlinear wavelength conversion to a more scalable nonlinear optics platform, e.g., integrated photonics.

One approach is to leverage the wavelength flexibility inherent to optical parametric oscillators using nonlinear microresonators, which possess large optical quality factors (Q) and small mode volumes to intensify circulating light and promote efficient nonlinear interactions.^8,9 Recent studies of microresonator-based optical parametric oscillators (μOPOs) have demonstrated broad spectral separation between pump and generated light,^10–12 low-power operation,^13,14 and visible-wavelength access.¹⁵ In particular, operating in the visible spectrum presents specific challenges: stronger material dispersion, larger scattering losses that reduce quality factors, and shorter evanescent field decay lengths that impede waveguide coupling. While both χ⁽²⁾ and χ⁽³⁾ μOPOs offer some wavelength flexibility, χ⁽³⁾ systems are useful to generate visible light from a NIR pump; moreover, their natural availability in the popular silicon photonics platform¹⁶ can enable their scalable fabrication and integration with other components, including pump lasers.¹⁷ On the other hand, the reported or inferred (e.g., from optical spectra) conversion efficiencies are ≲0.1%,^{10,13,15,18–20} and the available output power is far too low for many applications (e.g., $<$10 μW for previous visible μOPOs¹⁵). Realizing higher conversion efficiencies and output powers would enable a wide range of on-chip applications and broaden the reach of silicon photonics in the visible spectrum.

Here, we demonstrate efficient μOPOs that generate coherent light within the spectral window between 260 and 510 THz (590 and 1150 nm). We measure conversion efficiencies between 3.5% and 14.5% with corresponding on-chip output powers greater than 1 mW (and as high as 5 mW). Our results spring from efficient broadband waveguide-resonator coupling, which we realize with pulley-waveguide geometries designed using coupled-mode simulations. Notably, previous μOPO works have implemented such geometries to minimize threshold power;¹³ in contrast, our goal is to maximize output power using pulleys that operate in a fundamentally different regime. In the rest of this paper, we first introduce the key μOPO physics and specify our experimental procedures. Then, we explain our coupled-mode simulations and present measurements to confirm their accuracy. Next, we present the optical spectra of 16 different μOPOs from which we determine output powers and conversion efficiencies. Finally, we show how parasitic nonlinear processes currently constrain the maximum realizable output power. Our work is an important step forward in the quest for practical, chip-based sources of visible laser light using nonlinear optics.

The μOPOs we consider generate monochromatic signal and idler waves from a continuous-wave (CW) pump laser through resonantly enhanced degenerate four wave mixing (FWM), as depicted in Fig. 1(a).²¹ In experiments, we pump a fundamental transverse-electric (TE0) eigenmode of a silicon nitride microring near 385 THz, and FWM transfers energy to TE0 signal and idler modes. In principle, the range of accessible output frequencies, as constrained only by energy conservation, is DC to 2ω_p, where ω_p is the pump frequency. However, in practice, this range is dictated by the group velocity dispersion (GVD), which must be engineered such that FWM to the targeted signal and idler modes is favored (simultaneously phase- and frequency-matched), but FWM to other modes is suppressed. In the supplementary material, we recount our approach to dispersion engineering that is also described in Ref. 15.

A separate challenge is to ensure that pump power is efficiently converted into output signal or idler power. Hence, we define the on-chip conversion efficiency as

where P_s(i) is the signal (idler) power in the waveguide output and P_in is the pump power in the waveguide input. Recent theoretical work has derived the maximum obtainable CE as

where κ_s(i) and κ_p are the waveguide-resonator coupling rates of the signal (idler) and pump modes, Γ_s(i) and Γ_p are the total loss rates (i.e., loaded linewidths) of the signal (idler) and pump modes, and ω_s(i) and ω_p are the frequencies of the signal (idler) and pump light, respectively.¹⁰ Clearly, obtaining large CE involves engineering κ for both the pump mode and targeted signal/idler modes. Such coupling engineering is a common problem in the nonlinear optics of Kerr microresonators; it arises, for example, in the efficient extraction of octave-spanning Kerr microcombs.²² The problem is that, given a straight waveguide evanescently coupled to a microring resonator, κ decreases exponentially with frequency due to decreasing overlap of the microring mode with the waveguide mode. Hence, when the pump mode is critically coupled, the signal mode is undercoupled, resulting in low CE. Moreover, when ω_s is a visible frequency, the smaller evanescent decay length compared to NIR frequencies exacerbates the challenge. One solution is to utilize so-called pulley waveguides, which increase the physical distance over which the waveguide and microring can exchange energy.^22–24 Figure 1(b) illustrates the physical differences between straight-waveguide couplers (top panel) and pulley-waveguide couplers (bottom panel), and it depicts the three geometric parameters that define such couplers in our study: The waveguide width, WW; the waveguide-resonator gap, G; and the pulley length, L_p (which approaches zero in the straight-waveguide limit). In Fig. 1(c), we present measurements of μOPO spectra extracted from nominally identical microrings using either a straight-waveguide coupler (top panel) or a pulley-waveguide coupler (bottom panel). These measurements are representative of other comparisons between the two coupling schemes. While P_i is roughly the same in each case, P_s is ∼20× greater in the pulley waveguide. To explain the result, we show in the same panels the simulated coupling ratio, defined as K = κ/γ (equivalently, K = Q_i/Q_c, where Q_i and Q_c are the intrinsic and coupling quality factors, respectively), where γ = Γ − κ is the intrinsic loss rate. Near ω_s, K is $≈8×$ higher for the pulley-waveguide coupler.

To design couplers for testing, we simulate κ spectra for a variety of coupling geometries with the goal of achieving K ≈ 1 (i.e., critical coupling) at frequencies between 260 and 510 THz. Notably, achieving high CE only requires that κ be optimized at ω_p and either ω_s or ω_i, depending on which output wave (signal or idler) is to be used. (Moreover, ideally, the unused wave is undercoupled to reduce threshold power.) At the same time, broadband near-critical coupling is preferable, since then a single coupling geometry is robust against design imperfections, and it may be used for many different μOPOs. Our simulations are based on a coupled mode theory for optical waveguides,²⁵ which calculates κ according to the following equation:

where R is the microring outer radius, c is the speed of light, n_g is the group refractive index, and k_t is a coupling coefficient defined as

where ɛ_WG(R) is the dielectric permittivity of the access waveguide (microring), E_WG(R) is the electric field of the waveguide (microring) eigenmode, and ϕ is an accumulated phase accounting for the difference in the waveguide and microring propagation constants. The coordinates r and z are horizontal and vertical coordinates in-plane with the microring/waveguide cross section, as labeled in the inset of Fig. 2(a), and l follows the direction of light propagation. The central integral in Eq. (4) evaluates the evanescent overlap between the microring and waveguide modes at the frequency ω. For more details, see Ref. 22. Figure 2(a) shows simulated κ spectra for a SiN microring with R = 25 μm, ring width RW = 825 nm, and height H = 500 nm, which are chosen to suitably engineer the GVD (see the supplementary material). In addition, we choose WW = 375 nm, G = 150 nm, and L_p from 0 μm (i.e., a straight-waveguide coupler) to 5 μm. Increasing L_p results in larger κ at higher frequencies compared to the straight-waveguide coupler (dark gray line), but it introduces resonances in the κ spectra (i.e., regions where κ → 0) that arise from the coherent energy exchange between the microring and waveguide. These resonances blueshift when L_p is increased. Based on these data, we select L_p = 3 μm for further study because it minimizes variations in κ over the targeted spectral region.

Next, we optimize WW and G. The predominant effect of changing G is to vertically shift the entire κ spectrum; i.e., G has little impact on the spectral location of the coupling resonances. However, the relationship between κ and WW is more complex. Within the range of values considered, larger WW increases κ because the waveguide propagation constant shifts closer to the microring propagation constant, and the evanescent overlap between the microring and waveguide modes is not appreciably changed. At the same time, coupling resonances are redshifted. Figure 2(b) shows κ spectra for G = 150 nm, L_p = 3 μm, and three values of WW. Apparently, the flattest κ spectra are realized for WW between 375 and 400 nm. After choosing WW, G may be chosen to realize critical coupling near ω_p.

To assess the accuracy of our simulations, we fabricate an array of SiN microrings with systematic coupling parameter variations (see the supplementary material for details on our fabrication process), and we measure κ for each device. Specifically, we use either L_p = 0 or L_p = 3 μm, G between 110 and 160 nm, and WW between 350 and 400 nm. We carry out mode spectroscopy using a CW titanium sapphire (TiS) laser, which is tunable from 305 to 415 THz (720–980 nm), and from the microresonator transmission spectra, we calculate K and κ values following Pfieffer et al.²⁶ In addition, we can perform sum frequency generation with the TiS laser and a 154 THz (1950 nm) laser to generate coherent light from 460 to 510 THz (590–650 nm). We find that simulations predict slightly larger κ values than we measure; to compensate, we reduce G by ∼20 nm in experiments. In Fig. 2(a), we present κ measurements (gray circles) of a straight-waveguide-coupled device with WW = 375 nm and G = 125 nm. The measured κ values are slightly lower than the corresponding simulation with G = 150 nm, but both data decrease exponentially with frequency, which is a known characteristic of straight-waveguide couplers.²² Specifically, we measure κ ≈ 800 MHz near 350 THz, but it drops sharply to κ ≈ 40 MHz near 500 THz. In contrast, we observe a more achromatic κ spectrum in a pulley-waveguide-coupled device with WW = 375 nm, G = 135 nm, and L_p = 3 μm. Our measurements (green circles) are shown in Fig. 2(b), and they agree with the corresponding simulation with G = 150 nm. Our measurements indicate that between 300 and 500 THz, κ takes on values over the relatively narrow range (compared to the straight-waveguide coupler) of 180–400 MHz. We also measure γ ≈ 300 MHz that is approximately independent of optical frequency for the wavelengths of interest (we have observed that γ increases at higher frequencies, but this behavior varies between different fabrication runs and requires further examination). Therefore, according to Eq. (2), our best pulley-waveguide-coupled devices should support many different μOPOs with CE > 1%.

To test our prediction, we record μOPO spectra with a calibrated optical spectrum analyzer (OSA) and calculate P_s, P_i, and CE values after accounting for optical losses between the waveguide and OSA (for details, see the supplementary material). To generate μOPOs, we tune ω_p into resonance, starting blue detuned and decreasing ω_p until P_s and P_i are maximized. We repeat this procedure for different P_in values with the goal of maximizing CE. To ensure a variety of ω_s(i) values, we engineer the GVD by systematically varying RW in different devices (see Ref. 15 and the supplementary material). We utilize pulley-waveguide couplers such as those characterized in Fig. 2, and we find that CE is maximized for L_p = 3 μm, WW between 375 and 400 nm, and G between 125 and 135 nm. In Fig. 3(a), we present a compiled set of μOPO spectra that we extract from pulley-waveguide couplers with parameters in the above optimum range. In most cases, both P_s and P_i are greater than 1 mW, and P_in is typically between 30 and 45 mW. However, there are atypical μOPO spectra for which either P_s or P_i is ≪1 mW. Most likely, these result from mode interactions that locally alter the microring GVD and Q.^27,28 Table II of the supplementary material lists the individual pump, signal, and idler frequencies and powers for each μOPO spectrum shown in Fig. 3(a).

To characterize the μOPO performance, we calculate from Fig. 3(a) the largest values of P_s and P_i in spectral bins spanning ∼20 THz each, and we plot the results in Fig. 3(b) along with the corresponding CE values. We find P_i > 1 mW from 264 to 346 THz and P_i > 2 mW from 275 to 346 THz. In the best case, we generate 4 mW of idler power at 315 THz using P_in = 34 mW, which equates to CE ≈ 12%. Meanwhile, P_s > 1.9 mW from 416 to 506 THz, and in the best case, we generate 5 mW of signal power at 454 THz using P_in = 34 mW, which equates to CE ≈ 14.5%. Moreover, as expected from our simulations and measurements of κ and Eq. (2), CE decreases in the spectral wings as a result of smaller κ. Indeed, we can make a useful comparison between our CE measurements and a theoretical prediction (gray dashed line) based on our coupling measurements and Eq. (2). The central portion of data indicates CE values smaller than our theoretical prediction due to competing nonlinear processes that are exacerbated for narrow-band spectra.²⁹ Discrepancies at higher frequencies are primarily due to imperfect frequency mismatch arising from the stronger impact of higher-order dispersion. In between these extremes, we observe good agreement between our measurements and theory.

Finally, we discuss the limits to μOPO output power and analyze an example. The relationship between P_s(i), P_in, and other experimental parameters has been analyzed theoretically.²⁹ Therein, it was predicted that P_s(i) increases with P_in for P_in ≳ P_th, but further increases in P_in lead to saturation or even reduction of P_s(i). The reason is that parasitic FWM processes compete with the targeted μOPO process. The predominant parasitic FWM process that we observe in experiments is mode competition between the targeted signal/idler modes and their spectral neighbors.²⁹ In Fig. 4, we show the μOPO spectrum for a single device as P_in is increased. For P_in = 32 mW, only the pump, signal, and idler modes oscillate, and P_s(i) > 1 mW (top panel). The idler power in this case is marked by the red line in each panel. When P_in = 50 mW (middle panel), ω_s and ω_i shift to higher and lower frequencies, respectively. This behavior was predicted in Ref. 29 and termed “mode switching.” In addition, other modes with frequencies close to ω_s(i) begin to oscillate and steal energy from the μOPO. Hence, P_i decreases to a level below the red line, despite the increase in P_in. When P_in is further increased to 70 mW (bottom panel), P_s(i) remains approximately the same, but the power in the competing modes increases. The behavior demonstrated in this example is ubiquitous within our μOPO devices and explains why P_s(i) cannot be increased arbitrarily by increasing P_in. Still, increasing CE and P_s(i) beyond the levels we demonstrate may be possible using alternate phase- and frequency-matching strategies, such as that reported in Ref. 30. As it stands, the achieved power levels are relevant for some applications, such as spectroscopy of various coherent near-infrared and visible systems.^31,32

In conclusion, we have demonstrated energy-efficient μOPOs with practically relevant output powers in a crucial portion of the visible and near-infrared spectrum. Our results are enabled by broadband pulley-waveguide couplers that we design using coupled-mode simulations. For the most widely separated μOPOs, we observe relatively lower CE values that are consistent with undercoupled signal and idler modes. Hence, an important focus for future work is to extend the spectral bandwidth over which devices are nearly critically coupled, thus broadening the range of frequencies that can be efficiently extracted into a single waveguide. Other possible approaches include using frequency-specific coupling geometries (e.g., one may optimize κ at the μOPO-specific frequencies ω_s and ω_p, while neglecting the idler) or to couple multiple waveguides—each designed for different portions of the μOPO spectrum—to one microring. Moreover, new strategies should be devised to avoid parasitic FWM processes and increase the realizable output power. Nonetheless, our work is a compelling demonstration that μOPOs can help satisfy the demand for compact sources of visible laser light.

The supplementary material describes how the microresonator geometry is controlled to engineer GVD for μOPO, explains our conversion efficiency calculations, and details our device fabrication.

We thank Khoi Hoang and Tahmid Sami Rahman for carefully reading the manuscript and giving useful feedback. This project is funded by the DARPA LUMOS and NIST-on-a-chip programs.

The authors have no conflicts to disclose.

Jordan R. Stone: Conceptualization (lead); Data curation (lead); Formal analysis (lead); Writing – original draft (lead); Writing – review & editing (lead). Xiyuan Lu: Conceptualization (supporting); Writing – review & editing (supporting). Gregory Moille: Software (supporting); Writing – review & editing (supporting). Kartik Srinivasan: Conceptualization (supporting); Formal analysis (supporting); Funding acquisition (lead); Project administration (lead); Writing – review & editing (supporting).

The data that support the findings of this study are available from the corresponding author upon reasonable request.