An optical frequency comb (OFC) provides equally spaced optical frequency, which has potential applications in broadband dense wavelength-division multiplexing (DWDM) data transmission systems [1,2]. Especially for photonic integrated circuits (PICs), highly efficient multiwavelength light sources can effectively reduce power consumption, device size, and cost compared to the traditional mono-wavelength distributed feedback laser (DFB) arrays, which enables the capability of large bandwidth, low latency chip-to-chip optical I/O for artificial intelligence (AI) and machine learning (ML) [3–5]. Over the past few decades, researchers have proposed several methods for generating OFC, such as electro-optic modulation [6,7], Kerr microresonators [8–10], and mode-locked lasers (MLLs) [11–18]. In general, semiconductor mode-locked lasers (SMLLs) have great advantages in large-scale PICs due to their properties of small dimensions and high conversion efficiency [19]. In recent years, rapid progress has been made in quantum dot mode-locked lasers (QD-MLLs) as comb sources based on unique physical properties of small linewidth enhancement factor, fast carrier dynamics, and temperature insensitivity [20–23]. This makes the QD-MLL an ideal multiwavelength laser for integrated photonic I/O applications.

SMLLs have been extensively studied for producing optical pulses with a wide range of repetition rates from megahertz to hundreds of gigahertz, primarily by changing the length of the laser cavity. Typically, two-section MLLs, which have a saturable absorber (SA) at the end of the cavity can provide a repetition rate of 100 GHz at most, as the cavity length is already very short, which significantly limits the optical gain [13,14,20]. For high-speed modulation WDM systems, a repetition rate beyond 100 GHz is required to avoid modulation-induced cross talk. Generally, the repetition rate defines the comb channel spacing, which limits the highest modulation bandwidth due to cross talk between adjacent modes [24]. Therefore, for high-speed modulation ($>100\mathrm{Gb}/\mathrm{s}$), larger channel spacing is always preferred to avoid sidemode cross talk. However, increasing the frequency spacing leads to a short cavity length, which also significantly reduces the optical output power. Typically, to generate a 200 GHz repetition rate, an ultrashort cavity length of 198 μm is required, which means that it could hardly achieve sufficient optical gain to lase or operate at low output power. To increase the mode spacing beyond 100 GHz and at the same time retain sufficient output power and electro-optical conversion efficiency, the colliding pulse mode-locked laser (CPML) has been implemented, with significant progress made over the past few years [25–28]. For multiwavelength laser sources, sufficient optical output power is the major issue to be considered to overcome the insertion loss of modulators. All types of comb lasers require additional optical amplifications at the moment, which can significantly increase the overall power consumption. In addition, the presence of optical amplifiers also amplifies the noise level and reduces the signal-to-noise ratio (SNR) of individual channels, which further limits the upper limit of modulation speed. Meanwhile, the integrated optical amplifier also increases the cost and footprint, which is not conducive to the high-density packaging [29–32]. For future chip-to-chip optical interconnection, low power consumption multiwavelength lasers with large mode spacing and the absence of optical amplification are desired.

This study introduces a compact O-band quantum dot colliding pulse mode-locked laser (QD-CPML), enabling the generation of OFCs with 200 GHz frequency spacing that has no requirement of an additional optical amplifier in data transmission. Four comb channels centered at 1310 nm were selected, with an average optical linewidth of 1.23 MHz. Each of these channels exhibited an output power of nearly 10 dBm and remarkable 90 Gbaud/s PAM-8 modulation capacity, facilitated by thin-film ${\mathrm{LiNbO}}_3$ (TFLN) Mach–Zehnder modulator. Maximum comb lines of 12 channels within 3 dB optical bandwidth are also achieved by adjusting laser operating conditions. This advanced high-order QD-CPML stands as an ideal multiwavelength laser for power-efficient optical interconnects for future high-performance computing.

The epitaxial structure of this 200 GHz QD-CMPL is identical to those reported in our previous work [15], which is displayed in Fig. 1(a). The epitaxy process involves molecular beam epitaxy (MBE) on n-type GaAs (100) substrate, comprising eight stacks of self-assembled InAs quantum dot (QD) layers. To enhance high temperature stability, the active region of these dots is p-doped [21,22]. For balanced optical properties, a symmetric cladding layer design is adopted, using p-type and n-type ${\mathrm{Al}}_{0.4}{\mathrm{Ga}}_{0.6}\mathrm{As}$ layers, each 1500 nm thick, for the upper and lower cladding layers, respectively. Figure 1(b) demonstrates the cross-sectional SEM image after all fabrication processes, which shows as-cleaved facet.

As previously discussed, to generate OFCs with 200 GHz spacing, the cavity length of the fundamental mode is calculated at approximately 198 μm by giving an effective index of 3.8. However, this confronts substantial gain reduction challenges that restrict the maximum output power of individual comb lines. To tackle this, a fourth-order CPML design is introduced here to boost the optical gain by implementing longer cavity length but retaining the repetition rate. Figure 1(c) illustrates the schematic of the fourth-order QD-CPML. The total length of the laser cavity is 4 times a two-section MLL producing optical pulses at the same repetition rate, which is 792 μm. This configuration permits four colliding pulses within the cavity to produce fourth-order harmonic pulses at 200 GHz repetition rate, which is 4 times the fundamental frequency at 50 GHz. As depicted in Fig. 1(c), three saturable absorbers (SAs) divide the entire laser cavity into four gain sections. All SAs and gain sections are electrically interconnected through probe metal layers, respectively, for the convenience of butterfly packaging. While prior designs employed a single SA in the center of the laser cavity [25–28], our approach employs evenly distributed SAs throughout the cavity to yield stabler high-order harmonic mode-locked OFCs. To mitigate losses due to multiple SAs, the cumulative length of the three SAs is set at 10% of the total cavity length, slightly shorter than conventional typical 15%–20% design.

The QD-CPMLs exhibit ridge waveguide and mesa widths of 3 and 25 μm, respectively. The three SAs are evenly distributed along the cavity, each spanning a length of 50 μm. Adjacent to each SA, there is a 10-μm-wide electrical isolation gap. The fabrication of the QD-CPML follows the established procedures for ridge waveguide InAs QD lasers [15]. This includes dry etching, dielectric passivation, and metal contact deposition. Notably, the electrical isolation gaps are defined concurrently with ridge waveguide etching to ease the entire process. This is achieved by carefully controlling the dry-etching depth, which is maintained at 100 nm above the active region. This approach eliminates the requirement for additional isolation trench definition. Comprehensive details about these fabrication processes can be found in our previous work [15,32,33]. The wafer was thinned to 110 μm thick for final facet cleaving. High-reflection (HR) coating is applied here at single facet to boost the output power. The device is then wire-bonded onto the AlN chip-on-carrier (CoC) for heat dissipation and contact pad wiring. To be compatible with commercial laser controllers, a butterfly packaging is implemented here, as shown in Fig. 1(d). The butterfly module contains a QD-CPML chip, a thermal electric cooling (TEC) block, and a thermistor.

In order to further analyze the performance and stability of the QD-CPML, temperature-dependent light-current (L-I) measurements and optical spectral mappings are performed, as shown in Fig. 2.

Figure 2(a) shows the temperature-dependent L-I curves in the absence of reverse bias voltage. The threshold current of the laser at room temperature is approximately 30 mA, while the output power can exceed 100 mW at an injection current of 250 mA. As the temperature increases, the maximum output power of the laser reduces to approximately 20 mW, while the threshold current also increases to 120 mA. In order to further study the influence of reverse bias variation, the device is placed at room temperature for L-I measurements with varied reverse bias ranging from 0 to 5 V, as shown in Fig. 2(b). Under higher reverse bias voltage above 3 V, the L-I curve appears to behave nonlinearly with the injection current. The kinks in L-I curves indicate the locking regime transitions. At certain injection currents, such as 145 mA (5 V) and 225 mA (5 V), the output power appears at a minimum value, which corresponds to the termination of current mode-locked state in Fig. 2(d), and the laser spectrum transits into normal Fabry–Perot (FP) mode. On the other hand, at low reverse bias conditions, the L-I curves appear to be linear, due to relatively weak contribution of SAs within the laser cavity. Different reverse bias conditions can be selected for required central wavelength and bandwidth. For example, in the case that a high single-channel power is required, low reverse bias conditions can be implemented with the sacrifice of optical bandwidth. On the contrary, if more channels are required, an area with high current and high voltage can be selected, as shown in Fig. 2(c). At this operating point, a 3 dB optical bandwidth of 2.2 THz containing 12 comb lines has been reached. However, in this case, the power coupling out of each comb is approximately $-3\mathrm{dBm}$, which is not enough to overcome the coupling loss of the external modulator. Therefore, under such conditions, optical amplification may still be required to overcome the insertion loss. In order to systematically investigate the locking states of such a 200 GHz grid comb laser, optical spectral mapping has been performed across the current sweeping range from 0 to 250 mA and reverse bias voltage sweeping from 0 to 5 V, in order to show the number of channels available within the 3 dB bandwidth under different injection currents and reverse bias voltages, as shown in Fig. 2(d). It is obvious that there are two separate mode-locking regimes, divided by an unlocked region ranging from 170 to 190 mA injection current. As observed, the flap-top comb spectral region with maximum optical bandwidth always appears at the right corner of the spectral mapping, which corresponds to high injection current with high reversed bias voltage. As previously mentioned, in order to implementing the CPML at an operating condition where the single-channel power is maximum, a reduced number of optical channels are required here for nonamplified applications.

The experimental setup for the measurement of this QD-CPML is shown in Fig. 3(a), including frequency-resolved optical grating (FROG) pulse measurement (Setup I), optical linewidth measurement (Setup II), and relative intensity noise (RIN) measurement (Setup III). Taking both single-comb line power and RIN into consideration, the optimum operating condition is selected at $\mathit{I}=165\mathrm{mA}$, ${\mathit{V}}_{\mathit{R}}=1.5\mathrm{V}$, and $\mathit{T}=23°\mathrm{C}$. The optical spectrum of such an operating point is shown in Fig. 3(b) with four comb channels within a 3 dB optical bandwidth. The optical power of a single-comb channel can reach beyond 10 dBm at the laser facet, while reaching at least 2 dBm at the fiber end after considering the entire insertion loss of 10 dB, including the fiber coupling loss (5 dB) and optical bandpass filter (OBPF) loss (5 dB). The FROG pulse checker has been also implemented here to measure the time-resolved spectrum of this working point, which is shown in Fig. 3(c). The top of Fig. 3(c) shows the variation of the second-harmonic spectrum over time, which facilitates the calculation of the recovered pulse width. The bottom curves show the intensity of the autocorrelation function over time, where the period of the pulse train is 5 ps, indicating a repetition rate of 200 GHz and a pulse width of approximately 1.56 ps. To note is that, at low reversed bias conditions, the saturable absorption effect is not strong enough; therefore, the pulse cannot return to the noise floor level, as shown in Fig. 3(c). In contrast, at reversed bias voltages higher than 4 V, the time domain pulses appear to be neat pulse trains.

To measure the individual optical linewidth and RIN of each comb channel, a fiber-coupled OBPF with 60 dB extinction ratio and 5 GHz bandwidth has been applied here, which is shown in Fig. 3(a). The spectrum after filter with average linewidth of 1.23 MHz is shown in Fig. 3(d). The optical linewidth is relatively larger than values in our previously reported 100 GHz QD-CPML, which is in the range of few hundred kilohertz [15]. Here, for high output power applications, a relatively high injection current with low reversed bias voltage is selected, which can potentially lead to less stable frequency locking, which introduces a large optical linewidth. Overall, the lower ${\mathit{V}}_{\mathit{R}}$ of 1.5 V reduces the saturation absorption efficiency of the SA, while broadening the optical linewidth during the mode-locking process.

For QD lasers, in order to achieve a narrow optical linewidth, by implanting self-injection locking technology, the frequency noise can be further reduced by at least 3 orders of magnitude [33]. Beyond the compensation of widened optical linewidth, due to the increase of single-channel output power, the average RIN of this 200 GHz QD-CPML is approximately $-151\mathrm{dB}/\mathrm{Hz}$ in the spanning range from 0 to 20 GHz, which is about 15 dB/Hz lower than that of 100 GHz QD-CPML [15,34], as shown in Fig. 3(e). Such low RIN meets the requirements of achieving high-speed external modulation in PAM format.

As illustrated in Fig. 4(a), the 200 GHz grid comb laser is then coupled into TFLN Mach–Zehnder interferometer (MZI) modulator for external modulation. The TFLN MZI modulator is designed and fabricated by Niobate. Four individual comb channels, specifically at wavelengths of 1305.21, 1306.34, 1307.48, and 1308.62 nm, are meticulously chosen for external modulation. The external modulation capability of this 200 GHz fourth-order CPML is shown through an assessment of the eye diagram performance, which is shown in Fig. 4. A 90 Gbaud/s PAM-8 modulation format has been applied here with outer ER ranging from 4.73 to 5.88 dB. It is worth mentioning that the average power received by the receiver is about $-10\mathrm{dBm}$ under no optical amplification, which indicates that the overall optical loss of the entire system is close to 20 dB, including coupling loss (5 dB), filter loss (5 dB), modulator insertion loss (8 dB), and transmission loss (2 dB). More efficient laser–fiber coupling and wavelength filtering can be implemented in the future to reduce coupling loss and filtering loss, which will further reduce the power consumption [35].

Furthermore, if we can optimize the structural design and improve the optical gain bandwidth of QDs, the device can output more comb channels with sufficient optical power. Then the transmission capacity of a single-comb laser device can reach over Tb/s with no requirement of on-chip optical amplifier. Such a device can significantly reduce power consumption while improving efficiency for optical interconnection.

This work presents the initial demonstration of an O-band QD-CPML capable of generating OFCs at 200 GHz grid, encompassing maximum comb channels of 12 within a 3 dB optical bandwidth. For high-speed transmission experiment, four channels centered around 1310 nm were selected, each exhibiting an average linewidth of 1.23 MHz. Remarkably, these channels attained nearly 10 dBm output power at the facet, with the additional ability to achieve 90 Gbaud/s PAM-8 modulation via a thin-film ${\mathrm{LiNbO}}_3$ MZI modulator without additional optical amplification. With its high-order harmonic configuration, this QD-CPML stands as an excellent comb source for power-efficient optical interconnects and communication applications.

Acknowledgment. The authors would like to thank Dr. Wenqi Wei, Dr. Dong Han, and Mr. Jie Yan for their constructive contributions in device fabrications and measurements. Thanks to Wuhan National Optoelectronics Innovation Center (NOEIC) for providing measurement assistance. Thanks to Niobate Ltd. for providing TFLN modulators.