Recent advances in ultrafast semiconductor disk lasers

The performance of ultrafast semiconductor disk lasers has rapidly advanced in recent decades. The strong interest from industry for inexpensive, compact, and reliable ultrafast laser sources in the picosecond and femtosecond domains has driven this technology toward commercial products. Frequency metrology and biomedical applications would benefit from sub-200-femtosecond pulse durations with peak powers in the kilowatt range. The aim of this review is to briefly describe the market potential and give an overview of the current status of mode-locked semiconductor disk lasers. Particular focus is placed on the ongoing efforts to achieve shorter pulses with higher peak powers. As small, ultrashort-pulse lasers with high peak powers, semiconductor thin-disk lasers hold promise for spectroscopy, metrology and imaging. Bauke Tilma and co-workers from ETH Zurich review recent progress in developing such lasers for these applications. They report how the latest designs of external-cavity disk lasers that are passively mode locked using a saturable absorber now offer repetition rates of up to 100 gigahertz, pulse durations of shorter than 200 femtoseconds and peak powers in the kilowatt range. The compact design, excellent noise performance and customizable emission wavelength of such sources make them highly attractive for a variety of applications. Future research into the materials providing the optical gain and saturable absorption in such lasers is likely to further improve their efficiency, reliability and thermal stability.


INTRODUCTION
In recent decades, ultrafast lasers have evolved very rapidly toward ever-higher performance. Ultrafast lasers have three key characteristics that enable their application to market areas: First, their short pulse duration allows for high-resolution measurements in the timedomain. In other words, they are a near perfect ultrafast ''flash'' to measure high-speed phenomena. Second, because the laser energy is concentrated in the short pulse, they have very high peak powers, which enables key material interactions, the most important being ''cold ablation'' where the short optical pulse can remove or ablate almost any material without generating significant residual heat in the sample being processed. This technique allows for very precise micromachining of many existing and new types of materials and thin films in use today. It also has potential for use in future products. Additionally, it allows for new types of biomedical and tissue surgery applications. Third, the short temporal pulses have correspondingly large optical bandwidths, and this feature can be exploited for precise measurement diagnostics and metrology. A more detailed overview of these features and numerous other applications is given in several review articles 1,2 and is beyond the scope of this study. The simplicity of semiconductor saturable absorber mirror (SESAM) mode-locked lasers combined with diode-pumped solid-state lasers (DPSSLs), which were developed during the 1990s 3,4 has resulted in many new, practical, and commercially available ultrafast laser systems. These laser systems are being used extensively in many relevant applications, where expensive, power-hungry, maintenance-intensive lasers are being replaced. The recent development of inexpensive, more compact semiconductor disk lasers (SDL) may open new markets, such as compact measurement equipment. This result will ultimately enable ultrafast lasers to access high-volume consumer markets, e.g., light detection and ranging (LIDAR) techniques in the automotive indus-try 5 or natural user interface (NUI) applications in security and interactive media 6 .
For optical frequency comb applications, moving into the gigahertz pulse repetition rate regime has the advantage that the individual lines in the mode-locked frequency comb spectrum are spaced further apart, which increases the accessibility of the individual comb lines. Additionally, the power per comb line for a given average power is increased, which increases the signal-to-noise ratio in frequency comb metrology applications. The current field of gigahertz lasers is still dominated by the Kerr-lens mode-locked (KLM) Ti:sapphire laser, generating short femtosecond pulses with up to 10 GHz repetition rates 7 . Some recent developments in SESAM mode-locked DPSSLs 8,9 may allow further increases in repetition rate beyond the limits of the Ti:sapphire laser. In addition, they can be pumped with relatively inexpensive high-power pump diode arrays and do not have to operate as close to the cavity stability edge as KLM Ti:sapphire lasers.
Ultrafast semiconductor lasers have the potential to be cheaper and more compact in comparison to ultrafast DPSSLs. This type of laser operates extremely well in the gigahertz pulse repetition rate regime, has the advantage of being mass producible on the wafer scale and can be monolithically integrated into more complex optical circuits 10,11 . Furthermore, the emission wavelength can be designed via bandgap engineering, and the repetition rate can easily be increased up to several tens of GHz without Q-switching instabilities.
The field of semiconductor mode-locked lasers can be divided into edge emitters and top emitters. Each of these technologies has its strengths and weaknesses.
Edge-emitting mode-locked lasers are predominantly two-section waveguide lasers. One section acts as an amplifier with an injected current through the forward biased p-i-n diode. The second section is kept in reverse bias to act as a fast saturable absorber. Most of these lasers are designed as a straight Fabry-Perot cavity 10 or in a ring laser configuration 12,13 . The advantage of edge-emitting mode-locked lasers is their compactness with repetition rates often well above 10 GHz. The planar integration method enables the direct integration of these lasers into more complex optical circuits, which makes them very attractive for telecommunication applications 14,15 . In terms of pulse duration, values down to 312 fs for a 92-GHz quantum dot (QD) laser at 1550 nm with 13.2 mW of output power 16 have been reported. Most of these lasers however work in the picosecond pulse duration regime with maximum average output power levels of approximately 200 mW 17 . To reach a higher average power up to 800 mW and a peak power up to 6.5 kW, an external semiconductor amplifier 18 or an external cavity 19 and pulse compression were required, thus increasing the complexity of the system. In addition to the limited laser oscillator output power, edge emitters also have the disadvantage of a long gain interaction length, which increases the noise level 20 and limits their suitability for frequency metrology and optical communication.
Top emitters, also known as SDLs or vertical external-cavity surface-emitting lasers (VECSELs) 21 , combine the advantages of a semiconductor gain material with the benefits of conventional DPSSLs. They also benefit from the wavelength flexibility and mass producibility imparted by semiconductor technology. In mode-locked operation, they share power scaling with SESAM mode-locked diodepumped thin disk lasers 22 , where the mode size can be increased, and the short gain interaction length leads to low noise operation. SDLs have the additional advantage that the SESAM can be integrated into the gain structure to increase the integration level, which is referred to as the mode-locked-integrated external cavity surface emitting laser (MIXSEL) 23,24 . Since the first demonstration of an SESAM mode-locked VECSEL in 2000 25 , the maximum average output power has been increased to 6.4 W 26 , the peak power increased to 4.35 kW 27 and the pulse durations decreased to 107 fs (fundamentally modelocked) 28 or even 60 fs in burst operation 29 . In Figure 1, a timeline is given, including important milestones of the mode-locked SDL technology, such as first demonstrations (blue) and current record achievements (red). In Table 1, an overview of these important achievements is given as well as some publications, which were of major importance at the time of publication. For comparison, we also included some record achievements from SESAM mode-locked electrically pumped VECSELs and mode-locked edge emitters. The recorded average output powers versus pulse lengths are given in Figure 2.
One of the advantages of SDLs is their wavelength flexibility. In continuous wave (cw) operation, SDLs were operated at emission wavelengths ranging from 391 nm 53 to 5.3 mm [54][55][56] . The spectral coverage of cw SDLs has already been reviewed by several authors [57][58][59][60] . Using fourth harmonic generation, even wavelengths down to 244 nm were presented 61 . To date, mode-locked SDLs cover a range of center wavelengths between 665 nm and 2 mm. An overview of important results achieved with fundamentally mode-locked SDLs in different wavelength regions is given in Figure 3 and Table 2. Most of the results were published in the 950-1040 nm wavelength region using InGaAs quantum wells (QWs). The excellent results in this wavelength range were enabled by the relatively straightforward fabrication of thin high reflective aluminium arsenide/gallium arsenide (AlAs/GaAs) distributed Bragg reflectors (DBRs) on GaAs substrates. Due to the large refractive index contrast between the two materials and their wellmatched lattice constants, high quality DBRs can be obtained. Furthermore, the low thermal impedance of the AlAs/GaAs system supports efficient heat removal.
In this study, we placed our focus on the recent developments of ultrafast optically pumped SDLs. These lasers support mode-locked average output powers at the multi-Watt level with low-noise femtosecond or picosecond pulse trains. The high-Q cavity and short interaction length with the semiconductor gain medium minimizes the noise level. The flexible external cavity enables straightforward repetition rate tuning.
This article is organized into three sections. In the following section, we will briefly repeat the principles of the SDL technology. The current technological possibilities, such as repetition rate flexibility and frequency comb stabilization, will be discussed in the second section. In the last section, we present a roadmap toward even shorter modelocked laser pulses with increased peak powers in the kW range.

DEVICE PRINCIPLE
Vertical external-cavity surface-emitting laser (VECSEL) The VECSEL gain chip consists of a monolithically integrated highly reflective semiconductor DBR, a gain region containing several QW or QD layers and an antireflection section. The active region is optically pumped with an external low-brightness pump or electrically with a pi-n configuration. The laser cavity is formed by the gain chip and an external output coupler (OC). A schematic of the VECSEL layer stack and a cw laser cavity is given in Figure 4a and 4b, respectively. With the help of the external OC, the laser can emit a circular beam of high quality, typically a diffraction-limited Gaussian beam (TEM 00 ). The combination of a thin-film gain medium and an external cavity enables simple power scaling by increasing the cavity mode and pump spot sizes on the gain chip while keeping the intensities constant 32 . To accomplish this goal, the radius of curvature (ROC) of the OC and the cavity length needs to match a given pump area. The mode-size power scalability in the semiconductor disk can be attributed mainly to the one-dimensional heat flow within the very thin semiconductor layer toward the heat sink underneath, which has a very high thermal conductivity 32,74,75 . Since the first demonstration of the VECSEL in 1997 21 , cw output powers for optically pumped devices have been increased up to approximately 20 W in fundamental TEM 00 -mode and single frequency operation 76,77 and 106 W in multi-mode operation 78 . For cw operation, the optically pumped VECSEL can be an efficient mode converter from low brightness diode laser arrays to a fundamental Gaussian beam with an optical-to-optical efficiency of more than 40% 77 . Electrically pumped devices were demonstrated with cw output powers up to 500 mW 79 in TEM 00 -mode and were recently pushed to 9 W in highly multi-mode operation 80 . The limited output power from electrically pumped VECSELs compared to that from the optically pumped devices is mainly due to a limitation in the power scaling principle and the increased optical losses from free carrier absorption in the doped layers of the VECSEL chip 81,82 .  Table 1.
Advances in ultrafast semiconductor disk lasers BW Tilma et al 2 Semiconductor saturable absorber mirror (SESAM) Passive modelocking with an SESAM 3,83 is the state-of-the-art modelocking technique for VECSELs and was first demonstrated in 2000 25 . The first review of SESAM mode-locked VECSELs was given in 2006 58 . In the simplest configuration, an SESAM-mode-locked VECSEL is realized in a V-shaped cavity as shown in Figure 5b. In Figure 5a, a schematic of an SESAM layer stack is given. The SESAM consists of a highly reflective semiconductor DBR, one or more QW or QD layers, and some top layers to control the absorption behavior and dispersion 84 . The QW or QD layer(s) act as a saturable absorber. A typical nonlinear reflectivity curve, experimentally accessible with the technique described by Maas et al. 85 , is shown in Figure 6a. The reflectivity of the SESAM at very low probe pulse fluences (F p 5 E p /A, where F p is the pulse fluence, E p the pulse energy and A the cavity mode area) is defined as R lin. The absorbed part consists of nonsaturable losses DR ns and a saturated part DR, also referred to as the modulation depth. The saturation fluence F sat is defined as the fluence at, which the reflectivity is increased by DR/e. F sat is strongly dependent on the design and intrinsic properties of the SESAM and can be optimized for different lasers. For mode-locked VECSELs the saturation fluence of the SESAM should be kept low to start and stabilize passive modelocking 86 . At pulse fluences far above the saturation fluence, inverse saturable absorption effects, represented by F 2 , are observed 87 ; the most prominent of these effects is two-photon absorption. F 2 leads to a reduction in reflectivity proportional to exp(-F p /F 2 ), which is referred to as the rollover of the SESAM.
The mode-locked VECSELs benefit from a fast self-amplitude modulation of the SESAM. This can be obtained by band filling  58 . The related SESAM damage threshold has been studied by Saraceno et al. 98 in 2011. The absorber reaches transparency above certain pulse fluence as long as non-saturable losses can be minimized. The dynamic reflectivity change of the SESAM in combination with the dynamic saturation of the semiconductor gain governs the pulse formation mechanism of mode-locked VECSELs. These dynamics are schematically represented in Figure 7. The incoming pulse (black) reduces the losses in the absorber (red) at lower energies than the gain saturation occurs (green). Therefore, a net gain window opens up in which the pulse can be amplified. In addition to the saturation fluence and modulation depth of the SESAM, these recovery dynamics play a crucial role in the pulse formation process 86,99 . The induced Starkshift was reported to exhibit very fast recovery dynamics on the order of a few hundred femtoseconds 38 . The recovery dynamics of bandfilled saturable absorbers depends on intraband, interband and trap relaxation processes and typically involve a combination of recovery steps 88 , which can be modeled with a bitemporal response function via a simplified approach. In Figure 6b, a typical pump-probe measurement 100 of the recovery time of an SESAM is depicted. The fast recovery component with an exponential decay time t fast and amplitude A fast is induced by the intraband thermalization process and has a recovery time constant of a few hundred fs. The slow recovery component with an exponential decay time t slow and an amplitude 1-A fast is induced by interband recombination processes, mid-gap trap capture processes or interface/surface recombination processes. The combination of the fast and slow components and their relative amplitudes determine the total bitemporal recovery time of the SESAM. A fast recovery of the SESAM is one of the crucial parameters for achieving sub-picosecond pulse durations from SESAM mode-locked VECSELs 43,50 .
Other VECSEL modelocking mechanisms: graphene saturable absorber and self-modelocking In addition to the standard SESAM modelocking techniques, several other VECSEL modelocking techniques have been published. In 2013, Zaugg et al. 101 presented a VECSEL mode-locked with a single layer graphene saturable absorber mirror (GSAM). Despite the encouraging start with the demonstration of a very broad tunable operation bandwidth of 46 nm and a very fast absorber recovery time, the novel GSAMs have not advanced beyond the proof-of-principle status so far. Currently, the GSAM mode-locked VECSELs suffer from a limited average power in the milliwatt regime, which is caused by excessive nonsaturable losses and the low damage threshold. Higher output power levels of up to 10 W 102 from GSAM-mode-locked VECSELs were claimed recently; however, evidence for clean modelocking has not been provided to date.
Another modelocking mechanism reported recently is the so-called self-mode-locked VECSEL or the KLM-VECSEL [103][104][105][106] . No apparent saturable absorber is used in these laser cavities, and a hard aperture is inserted in front of the OC or highly reflective end mirror. In analogy to solid-state lasers, the modelocking mechanism was initially explained by the Kerr lensing effect in the gain region, and the effect can be enhanced with an extra Kerr medium inside the cavity 107 . The reported results were promising, and a more detailed set of measurements 108 was presented only recently 109 . However, it remains to be observed whether these lasers are really fully mode-locked. We have observed in our labs that more sophisticated diagnostics are required for ultrafast SDLs to confirm stable modelocking. Furthermore, the physical origin of the nonlinear effects in the gain medium and more detailed noise characterization will require more research efforts.
Mode-locked integrated external-cavity surface-emitting laser (MIXSEL) A logical following step toward more compact high-power semiconductor mode-locked lasers is the MIXSEL. The MIXSEL combines the active part of a VECSEL with the saturable absorber of an SESAM in a single semiconductor chip. In this way, stable and self-starting  A schematic design of a MIXSEL layer stack is given in Figure 8. The MIXSEL layer stack includes two DBRs underneath the active area. The first mirror is designed to reflect residual pump light and be transparent to the laser light. This pump DBR inhibits pre-saturation of the absorber by residual pump light. The mirror underneath is designed to be highly reflective for the lasing wavelength. The saturable absorber is placed in between the two DBRs in an antinode of the standing wave.
The first MIXSELs had a resonant design in which the electric field enhancement at the absorber position was increased compared to the field in the gain section. This enhancement was implemented to reduce the saturation fluence of the absorber, and therefore, saturate the absorber at lower pulse energies than the gain 110 . However, the resonant design is very sensitive to growth errors, and the group delay dispersion (GDD) fluctuates very strongly (610 000 fs 2 ) around the lasing wavelength. With the development of novel QD absorbers, the saturation fluence was decreased, whereas the modulation depth was kept in the desired range 111 . These QDs were used in the first antiresonant MIXSEL structure, with which a record high average output power of 6.4 W in 28-ps pulses at 2.5-GHz repetition rate was obtained 26 . The trade-off of those QD absorbers is the relatively long slow recovery time, which sets a lower limit to the achievable pulse length. Newly developed QW-based saturable absorbers have overcome this issue. Low temperature grown QWs embedded in AlAs yield fast recovery dynamics and low saturation fluences despite the long annealing process during the subsequent epitaxy. These fast recovery dynamics and low saturation fluences are necessary for a very fast MIXSEL 112 . This resulted in femtosecond pulses from a MIXSEL for the first time 50 , generating 620-fs pulses with a 4.8-GHz repetition rate and an average output power of 101 mW. With the same MIXSEL chip, it was possible to stepwise increase the pulse repetition rate from <5 GHz up to 100 GHz 51 by reducing the cavity length from <30 mm down to 1.5 mm and changing the OC to maintain stable cavity configurations. At the record high 100-GHz pulse repetition rate, we obtained pulses with a duration of only 570 fs and an average output power of 127 mW.

STATE-OF-THE-ART ULTRAFAST SEMICONDUCTOR DISK LASERS (SDLS)
The performance of ultrafast SDLs rapidly improved over the last decade. Cw SDLs already have been successfully commercialized, for Table 2 Overview of important results achieved with fundamentally mode-locked SDLs in different wavelength regions.
Material system (QW/substrate) Number of QW/QDs and material used l c t p P av f rep Author 1 reference
Pulse repetition-rate scaling and tuning Different applications require different repetition rates. For example, biomedical imaging in the multi-photon regime requires lower pulse repetition rates to reduce the average power while maintaining a high peak power, whereas high repetition rates are beneficial for telecommunication, data communication, LIDAR applications and optically sampled analog-to-digital converters (ADCs). Fundamentally mode-locked SDLs have been demonstrated at repetition rates starting from 100 MHz 46,47 up to 100 GHz 51 . The highest performance with respect to peak power, average power, and pulse duration was achieved at repetition rates between 1 and 10 GHz (Figure 9). In contrast to DPSSLs, SDLs do not suffer from Q-switching instabilities 114 in the gigahertz regime. Achieving low repetition rates to increase the peak power is very challenging for semiconductor mode-locked lasers because they exhibit a carrier lifetime that is several orders of magnitude shorter than for typical ion-doped glass or crystalline gain materials, which denotes that the semiconductor gain material is able to store energy for only a limited time of a few nanoseconds. At lower repetition rates, where the separation between subsequent pulses is longer, splitting the energy into two or more pulses represents a gain advantage for the laser, which introduces modelocking instabilities, such as multipulsing or harmonic modelocking. This effect has been studied in more detail by Saarinen et al. 115 In the high pulse repetition rate regime, scaling the repetition rates above 10 GHz becomes more challenging due to the short external cavity length. Nevertheless, repetition rates up to 50 GHz have been obtained with a fundamentally mode-locked VECSEL 35 . Recently, the maximum repetition rate has been increased up to 100 GHz with the MIXSEL 51 , which was possible due to the simple linear cavity. The cavity length was only 1.5 mm in this case. To further reduce the cavity length and increase the repetition rate, implementing an external OC to an attached flat 116  Advances in ultrafast semiconductor disk lasers BW Tilma et al 6 highest peak power of any mode-locked laser technology, as shown in Figure 10 51 .
The SDL technology supports very simple repetition rate tuning by changing the cavity length with the OC or the SESAM on a linear translation stage. A repetition rate-tuning range of several GHz has been demonstrated with an SESAM-mode-locked VECSEL in the 1-10 GHz repetition rate range without changing any cavity components 86,118,119 . With the MIXSEL, a scaling range from 5 to 100 GHz has been achieved using the same MIXSEL semiconductor chip but different OC 51 . The transmission and ROC of the OC had to be adapted to ensure a stable cavity configuration. Smaller repetition rate changes could be achieved without changing the OC. This relatively    simple change in repetition rate is an advantage in case the repetition rate of the laser has to be stabilized to an external reference source.

Frequency comb stabilization
The low timing jitter and amplitude noise of mode-locked DPSSLs 120 have enabled several records for frequency comb generation 121 and data transmission rates 122 . Comparable low-noise performance has been achieved with SESAM-mode-locked VECSELs [123][124][125][126] and MIXSELs 127 , both in free-running operation as well as when the repetition rate is stabilized to an external reference source. In Figure 11, the timing phase noise from a selection of published VECSELs 125,126 and MIXSEL 127 is shown both in free-running operation and stabilized to an external reference source. For comparison, the free-running timing phase noise of a GHz SESAM-mode-locked Yb:CALGO laser is added 128 . In Table 3, the integrated rms timing jitter and amplitude noise are given. In contrast to edge-emitting semiconductor lasers, SDLs have a very short interaction length of the optical pulse with the gain layers and the cavity losses are very low. The high-Q cavity intrinsically imposes a lower noise level 129,130 than semiconductor edge emitters [131][132][133] . The low-noise behavior and the possibility of stabilizing the repetition rate to an external reference source make ultrafast VECSELs and MIXSELs attractive laser sources for compact and cost-efficient frequency comb metrology applications. For those applications, the carrier envelope offset frequency (f CEO ) 134 and pulse repetition rate have to be stabilized 135,136 . In the spectral domain, the pulse repetition frequency f rep defines the distance between the individual frequency comb lines and f CEO , the offset to DC. The mode-locked frequency comb is then fully described by f CEO 1 m f rep , where m is an integer. With the stabilization of both frequencies, the complete comb can be locked, which has already been demonstrated for Ti:sapphire lasers 137 , fiber lasers 138 , thin disk lasers 139 , and DPSSLs 121 . The detection of the CEO-frequency f CEO is very challenging for ultrafast SDLs. The most widespread method relies on the generation of a coherent octavespanning supercontinuum (SC), which is launched into an f-to-2f interferometer to detect the CEO beat notes 134 . The SC is typically generated in a highly nonlinear photonic crystal fiber (PCF) 140 . Recent work on SDLs has increased the peak power up to the level necessary for SC generation. Nevertheless, an octave-spanning SC has not yet been achieved by directly launching the output of a mode-locked SDL into a PCF 27 . Additional passive compression to 150-fs pulses with an average output power of 270 mW corresponding to a 1015-W peak power (or 220-fs pulses at 520 mW and 1333 W peak power) was not sufficient to generate the octave spanning SC 141 due to the limited average output power from the laser. An obvious intermediate step is pulse amplification in an active fiber amplifier, which has already been demonstrated in the 1.57-mm wavelength region 142 to generate a wide spectrum for broadband WDM telecommunication technology 143 . In this experiment, 14.4-ps pulses with an average output power of 100 mW from the mode-locked VECSEL were amplified to 4.5 W average powers in 15.5-ps pulses. These pulses where launched into the highly nonlinear fiber to generate a broad (most likely incoherent) continuum from 1320 nm to 2000 nm. Furthermore, this range does    Table 3 Integrated rms timing jitter and amplitude noise of the results depicted in Figure 11. Advances in ultrafast semiconductor disk lasers BW Tilma et al 8 not fully cover an octave, a requirement imposed by the f-to-2f detection scheme to observe the f CEO beat notes.
Recently, we amplified 231-fs pulses with an average output power of 100 mW from a 1040-nm mode-locked VECSEL up to several Watts and were able to compress the pulses to 85.4 fs with an average output power of 2.2 W 48 . A fraction of the available power was launched into a PCF to generate a coherent, octave-spanning SC from ,680 nm to .1360 nm Figure 12a. The dispersive wave centered at 680 nm and the Raman soliton centered at 1360 nm were used to generate a CEO beat note in a standard f-to-2f interferometer, as shown in Figure 12b. This f CEO beat note at approximately 733 MHz can be used to stabilize one degree of freedom of the frequency comb generated by the modelocked VECSEL. The CEO beat note shifts when modulating the pump current of the VECSEL, which can be used for a future stabilization of the CEO frequency. In addition to the CEO stabilization control via the pump current to the diode pump lasers, the CEO frequency can also be stabilized by optically pumping the SESAM 144 . Whether this stabilization mechanism is suitable for SESAM mode-locked VECSELs still has to be investigated. For MIXSELs, this result will be even more challenging.
The proof-of-principle CEO-frequency detection from a semiconductor laser is the first step toward a completely stabilized frequency comb from an SDL. The amplification and compression stages used by Zaugg et al. 48 have to be omitted to provide an advantage over conventional oscillators used for frequency combs. Therefore, we target ultrafast VECSELs, which provide 100-150-fs pulses with average output powers of several hundreds of mW directly from the oscillator. The rapid progress in the field of SESAM-mode-locked VECSELs 27,28,45 and MIXSELs 50,51 indicates that the desired performance is within reach.

ROADMAP TOWARD SUB-200-FS PULSES AND KW PEAK POWERS
Many applications, such as frequency comb metrology and biomedical imaging techniques, require ultrashort pulse durations, ideally less than 200 fs, in combination with high peak powers in the kW range. This applies especially for applications where nonlinear effects play a key role, such as coherent SC generation in PCFs 145,146 . Therefore, significant effort is focused on reducing the pulse duration and increasing the peak power of mode-locked SDLs. The determination of the peak power of a mode-locked SDL requires special care because of the strong tendency for multiple pulsing by over-saturated absorbers, which has been well documented in DPSSLs since as early as 1997 147 and reviewed by Keller in 2007 148 . Thus, additional experimental evidence is required for stable fundamental modelocking (e.g., sampling oscilloscope and high-contrast autocorrelators) to confirm that no cw components, very long pulse pedestals or additional satellite pulses are present. A discussion about proving clean modelocking was recently published by Wilcox et al 108 . An overview of SDL peak powers versus pulse duration is given in Figure 13. So far, there have been only three published results with peak powers in the kW range; however, none of these reports have pulse durations in the sub-200 fs regime. In the following section, we intend to discuss the main design and fabrication challenges to achieve ultrashort pulses in combination with high peak powers.
Toward kW peak power Output powers of SDLs (average and peak) depend on a delicate balance and optimization of epitaxial design, growth conditions and  operating conditions, such as heat removal. Several groups have been working on the optimization of cw average output powers from SDLs 149,150 . Table 4 gives an overview of some of the best reported output power levels, optical-to-optical and slope efficiencies of cw SDLs with output power levels above 10 W and slope efficiencies above 30%. No clear relationships between structural design, output power, and efficiency seem to exist. In addition to optimized epitaxial growth, one of the important design issues to achieve high performance SDLs is the ability to remove heat efficiently from the active region. The quantum defect related to the barrier-pumping and non-radiative recombination processes inherently introduce heat in the active region, which leads to performance degradation and a bandgap shift. Heat removal from the active region can be optimized as follows: using high thermally conductive semiconductors, e.g., favoring the use of GaAs and AlAs above the ternary material AlGaAs, which in most cases, has a lower thermal conductivity; using highly thermal conductive heat-sinks, such as chemical vapor deposited (CVD) diamonds; minimizing the distance between the active region and highly thermally conductive mount(s); and minimizing thermal resistance in the bond between the chip and the mount. In addition, the optical-to-optical efficiency can be improved by reducing the reflectivity of the pump light at the semiconductor interface 153 .
We note a large difference in efficiency when we compare the reported cw output powers and efficiencies from Table 4 with the reported mode-locked output powers and efficiencies in Table 5. A significant part of this decrease can be related to the short carrier lifetimes of a few ns in semiconductors compared to ms or ms lifetimes for ion-doped laser crystals. The carrier ''leakage'' between subsequent pulses due to radiative and non-radiative recombination is significant and reduces the overall efficiency. Therefore, it is important to minimize the impurities and defects in the crystal lattice to minimize deep-level traps and increase the carrier lifetime 154 . Furthermore, the intracavity optical peak intensities are significantly higher than in cw lasers. This factor inherently means that nonlinear absorption effects, such as two-photon-absorption, are increased and will reduce the efficiency. Therefore, optimizing growth to reduce defects and doping concentrations is crucial. Reducing the field enhancement and, consequently, the intensity in the semiconductor structure is also possible, which needs to be compensated for to obtain sufficient net gain.
The design of the active region plays a key role. Strain compensation in the active region decreases the number of defects in the structure and increases the lifetime of the device. The active layers (QWs or QDs) can be placed exactly in the antinodes of the standing wave pattern (also known as a resonant periodic gain (RPG)), which is normally performed in cw VECSELs to achieve the highest gain. The QW position can also be altered with respect to the antinodes to reduce the field enhancement and increase the gain saturation fluences 155 , which is beneficial for achieving high peak intensities in mode-locked operation. Therefore, a larger number of (or more efficient) active layers are needed to compensate for the overall reduced gain when the field enhancement of the active layers is reduced. Additionally, the QWs can be placed inhomogeneously within the pump-absorbing layer to distribute the excited carriers more equally over the QWs. This result can even be further enforced by introducing carrier barriers in the active region 156 .

Toward sub-200 fs pulses
The generation of ultrashort pulses in the femtosecond domain depends on a complex combination of parameters in the gain and absorber layers. First, the 1/e recovery time constant of the absorber has to be on the order of a few ps or even sub-ps 50,58 to enable only a very short net gain window for the interplay between the dynamic gain and absorber saturation ( Figure 7). As-grown QDs can exhibit a fast recovery 43 ; however, a long annealing time (e.g., during MBE growth after the absorber growth in an MIXSEL) led to very slow recovery times longer than 100 ps with the same absorber 49 . Newly developed fast QW absorbers using AlAs barriers seem to be more resistant to long annealing times. The saturation fluence can be low 112 , and the  Advances in ultrafast semiconductor disk lasers BW Tilma et al 10 recovery time can be decreased by reducing the growth temperature 50 . However, the interplay between the absorber and gain saturation is more tricky and more temperature sensitive. Thus, we believe that in the long run, improved QD saturable absorbers may still provide better saturable absorbers in an MIXSEL.
The modelocking mechanism of SDLs relies on a balanced interplay between the saturation effects of the gain and absorber in conjunction with the GDD of the structure 99 . Sieber et al. explored the influence of those effects on the pulse formation process in mode-locked SDLs with an experimentally validated numerical pulse formation simulation based on macroscopically measureable parameters 86 . The simulations revealed that the shortest pulses could be generated with a broadband gain material in combination with a total intracavity GDD close to zero (Figure 14a). Furthermore, the gain saturation fluence should be as high as possible (.200 mJ/cm 2 ), whereas the absorber saturation fluence must be kept low (,5 mJ/cm 2 ) to generate sub-200 fs pulses with significant output power (i.e., .500 mW average output power) (Figure 14b and 14c). The gain saturation fluences of ultrafast VECSELs have been measured and found to be in the range between 30 and 80 mJ/cm 2 . 155 Increasing the gain saturation by reducing the field enhancement decreases the small-signal gain, as discussed in the previous subsection, which makes it technologically more challenging to combine sub-200 fs pulses with a high average output power. Recently, the effect of gain saturation has been theoretically studied by Kilen et al. 157 with a detailed microscopic model. The authors clearly demonstrate the effects of gain saturation and kinetic hole burning in mode-locked lasers. This publication also emphasizes the instabilities in mode-locked operation, such as (unwanted) multipulse formation at high inversion levels.

CONCLUSIONS
Recent ultrafast SDLs reached a level of maturity that could potentially enable widespread scientific and industrial applications. Pulse durations have been reduced to sub-200 fs, and peak power levels have been reported in the kW-range. The enormous wavelength flexibility and compact design is a large advantage over other types of ultrafast lasers. Furthermore, excellent noise performance, comparable to conventional ion-doped solid-state lasers, allows for exploring the field of ultrastable frequency comb generation. Therefore, an SDL-based ultrastable gigahertz frequency comb could be interesting for several optical measurement techniques, including high-precision optical frequency metrology and spectroscopy, biomedical imaging, optical coherence tomography, and high-speed asynchronous optical sampling.
Current research is focused on increasing the peak power into the kW range while maintaining clean sub-200-fs pulse durations. In addition to a well-balanced interplay between the gain material, and the saturable absorber and a well-designed dispersion control, research has to focus on the performance of the gain and absorber material. Future improvements in ultrafast SDL designs should increase the efficiency, reduce the thermal side effects and improve the reliability of these devices.