Adaptive Optics for Extremely Large Telescopes

Adaptive Optics has become a key technology for the largest ground-based telescopes currently under or close to begin of construction. Adaptive optics is an indispensable component and has basically only one task, that is to operate the telescope at its maximum angular resolution, without optical degradations resulting from atmospheric seeing. Based on three decades of experience using adaptive optics usually as an add-on component, all extremely large telescopes and their instrumentation are designed for diffraction limited observations from the very beginning. This review illuminates the various approaches of the Extremely Large Telescope, the Giant Magellan Telescope, and the Thirty-Meter Telescope, to fully integrate adaptive optics in their designs. The article concludes with a brief look into the requirements that high-contrast imaging poses on adaptive optics.


Introduction
Over the past 30 years, ground-based astronomical observations have drawn level with their space-based counterparts. Thanks to adaptive optics (AO), the seeing limit caused by optical turbulence of the Earth atmosphere is decreasingly a real limitation for existing large optical telescopes. For the three extremely large telescopes currently under construction or about to begin construction, adaptive optics is an integral part of telescope design, allowing to achieve the highest angular resolution on sky: the telescope's diffraction limit.

Basic principle
The technique behind AO was first formulated by American astronomer Horace W. Babcock (1953) and can be summarized as cancelling optical aberrations caused by the atmosphere in real-time. Babcock's proposal to compensate astronomical seeing was to measure wavefront distortions (using a rotating knife-edge) and feedback that information to a wavefront correction element (Eidophor mirror). In modern applications a fast real-time controller (RTC) sits in between a wavefront sensor (WFS) and a wavefront corrector (e.g. a deformable mirror (DM)) such that these three elements build a closed-loop system, i.e. controlling wavefronts before they reach the telescope instrument (e.g. the scientific camera). However, it took until the early 1970s before Babcock's idea could be tested experimentally (Hardy et al., 1977).

Early years of AO in astronomy using single natural and laser guide stars
The usage of AO in routine astronomical observations on telescopes of the 3-4 m class can be dated back to the end of the 1980s when the instrument Come-On (Rousset et al., 1990) delivered diffraction limited images in the near-infrared spectral range on a 1.5 m telescope. Upgrades from Come-On to Come-On-Plus (Gendron et al., 1991) and eventually to ADONIS (Beuzit et al., 1994) allowed diffraction limited imaging on a 3-4 m class telescope in the 1-5 µm spectral range.

FIRST ON-SKY RESULTS
First Argos commissioning results obtained on the right side of LBT have been presented in 4. We repeat some of the discussion made there and complement it with new comparisons of Argos performances and simulations, as well as a larger set of Argos on-sky measurements both with left and right sides of LBT. Those results confirms the satisfactory performances of Argos providing an improvement in PSF FWHM of a factor around 2 depending on the wavelength, relative strength of the ground-layer, etc.. As a GLAO system, the improvement brought by Argos are the increased sensitivity and the uniform improved resolution over a 4 ×4 field-of-view. In the following sections, we therefore discuss those aspects.

Uniformity
Uniformity of the PSF allows to more easily derive the completeness limits of deep imaging surveys and ease accurate photometry and astrometry. It is also an important parameter in efficiently carrying multi-object spectrograph observations.
Argos typically correct 150 modes with tip-tilt being corrected by one NGS and all other modes by performing a tomographic ground-layer reconstruction. The single tip-tilt star is picked in a field-of-view of 2 ×3 within the Luci FoV, and is typically taken as close to the center of the field as possible. From this concept, the main potential source of non-uniformity due to the AO system is the tip-tilt anisoplanatism. But since the GLAO correction is modest, the sensitivity to anisoplanatism is also smaller.
The measurements done so far with Argos and Luci show no sign of anisoplanatism. Of course this also depends on the location of the tip-tilt star which has been preferentially chosen in the center of the Luci FoV. This is illustrated in Figure 2, for Maffei1 observed on the left LBT side, showing the constant PSF FWHM as a function of radius for the three NIR bands in closed loop.

Point source sensitivity
An important outcome of improving the resolution is to increase the point source sensitivy. We reproduce here a discussion presented in 4. 2016) show that a seeing improvement by a factor of 2 can be achieved under good conditions, down to 0.2" at 1.6µm over a 4 arcmin wide field of view.
GLAO systems using multiple sodium laser guide stars are currently installed and commissioned on Paranal La Penna et al., 2016). First results obtained with the adaptive optics facility (AOF) and the instrument MUSE -good to recognize in ( fig. 2) -show the potential of this technique in particular for wide-field GLAO spectroscopic observations. During the MUSE science verification in 2017, observations under one arcsecond seeing conditions were improved with GLAO down to ∼0.6 arcsec across the full 1 arcmin field of view (Leibundgut et al., 2017).
The last flavor of AO type to mention before looking at the AO related hardware designs for the upcoming extremely large telescopes, is the multi-object adaptive optics (MOAO) system. The goal of MOAO is to compensate atmospheric turbulence over a wide field of view up to 5-10 arcmin, using individual deformable mirrors for each object to be observed. Using a small number of either natural or laser guide stars, similar to MCAO, the wavefront is reconstructed using tomographic methods Vidal et al., 2010). Each deformable mirror is optimally shaped to correct the turbulence in its viewing direction, like a multi-SCAO system operated in open-loop. MOAO demonstrator systems used in the laboratory Ammons et al., 2008) and on sky (Gendron et al., 2011;Lardière et al., 2014), have shown that MOAO performs close to conventional SCAO systems.
August 9, 2018 0:19 ao18_nocomments 4 Hippler Fig. 2. Exemplary ground layer adaptive optics performance at the very large telescope (VLT) with sodium laser guide stars. VLT MUSE+AOF view of the planetary nebula NGC 6563 without GLAO correction (left) and with GLAO correction using 4 sodium laser guide stars (right). The field of view is 1.01 x 1.03 arcmin. Both images are color composites observed in the optical with various band filters ranging from 500 nm to 879 nm. (ESO, 2017). The first MOAO system installed on a 8 m class telescope was the RAVEN Lardière et al., 2014;Davidge et al., 2015;Ono et al., 2016;Lamb et al., 2017) demonstrator, which was in operation at the Subaru telescope between 2014 and 2015. Developed as a pathfinder for the 30 m TMT, results obtained with this instrument are encouraging, with better performance than GLAO ( Fig. 3) but not reaching SCAO performance. RAVEN can use either three natural guide stars or a combination of one sodium laser guide star and two natural guide stars. Two deformable mirrors with 145 actuators each, allow AO corrections on two science fields with a maximum separation of about 3.5 arcmin.
Conceptual drawings of the various AO concepts, i.e. SCAO, GLAO, MCAO, LTAO, and MOAO, including drawings of conventional wavefront sensors and photos of deformable mirrors, can be found in the review article of Davies & Kasper (2012).
It goes without saying that the current, often already the 2nd generation of AO systems on 8-10 m class telescopes has generated a wealth of experience. These experiences and lessons learned (see for example Lloyd-Hart et al. (2003); Wizinowich (2012); Fusco et al. (2014); Lozi et al. (2017)) are of course incorporated into the design and simulations of the AO facilities to be built. Since it would go beyond the scope of this review, here are just two current examples. The SPHERE XAO system installed at the Paranal observatory discovered a so named low wind effect Milli et al. (2018) (sometimes also named island effect), which basically distorts the wavefronts in the pupil of large telescopes obstructed by spiders. The effect occurs mainly when wind speeds above the telescope are close to zero. The SPHERE XAO system cannot measure these piston aberrations, thus they propagate uncorrected to the science channel and degrade the image quality. The solutions proposed and tested will influence the design of possible upgrades to the SPHERE XAO system as well as future AO systems.
The second example concerns a general problem of most AO systems: telescope vibrations and their impact on AO performance. Experiences with the SCExAO system at the Subaru telescope Lozi et al. (2016) show weaknesses of the control system in handling vibrations, which may be mitigated by the use of predictive-control (Males & Guyon, 2018).  (blue), the adopted asterism (green dashed), and the two science targets (red). The two black circles correspond to 60 and 120 arcsec. This image was taken from the DSS survey archive (http://archive.eso.org/dss/dss). custom 145 actuator ALPAO DMs, with 11 actuators spanning the pupil. MOAO features open-loop control as opposed to the closedloop control of a classic AO system (otherwise known as SCAO), meaning the system will command the DM using WFS data of the full turbulence as opposed to the residual turbulence. In other words, the WFS does not see the DM correction because the DM is located after the WFS. However, RAVEN is also capable of employing SCAO and ground-layer adaptive optics (GLAO). GLAO is a technique that corrects for ground-layer turbulence only, and is employed by averaging the measurements of all three WFSs to drive the DMs. GLAO can be particularly useful when the RAVEN SLODAR showed the majority of the turbulence to be in the ground layer. Fig. 1 shows a typical asterism, defined by the green dashed boundary, as well as two potential science targets within the asterism shown in red.
RAVEN re-images the two channels on to the slit of the IRCS on the Subaru Telescope. The IRCS includes two 1024 2 ALADDIN III arrays side by side with a wavelength coverage of 0.9-5.6 µm. With one target in each pick-off arm, then both stars are directed to opposite sides of the IRCS slit, such that two spectra are gathered simultaneously. All observations were obtained using the echelle mode of the IRCS with the 0.14 arcsec slit, yielding R ∼ 20 000 spectra. The H-band wavelength region was chosen (H+ filter 3 ), with intermittent coverage from 15 000 to 17 000 Å. This region was selected to take advantage of the line list and techniques for analysing this region developed by the SDSS-III APOGEE project (García Pérez et al. 2013;Smith et al. 2013;Shetrone et al. 2015). 3 www.naoj.org/Observing/Instruments/IRCS/echelle/orders.html

Performance and science observations
Three engineering runs took place between 2014 May and 2015 June at the Subaru Telescope, during which several AO tests were carried out, followed by the observations of our metal-poor star candidates. The details of the observations are summarized in Table 1. Over the four nights when the science observations were taken, the mean seeing of no-AO, SCAO, GLAO, and MOAO was 0.47, 0.09, 0.26, and 0.26 arcsec, respectively; this shows a general improvement of at least 0.20 arcsec in full width at half-maximum (FWHM) and demonstrates the AO performance of the instrument. As a result, 2-3 times more flux passed through the slit of our observations, depending on the night and observing mode (MOAO or GLAO). Fig. 2 demonstrates the quality of the point spread function (PSF) for MA8 (prior to incidence on the slit) with no AO correction, and with MOAO and GLAO corrections.

Galactic Centre targets
Three red giant branch (RGB) stars were selected for their low metallicity and proximity to the Galactic Centre, based on both SkyMapper photometry and low-resolution spectroscopy (from the EMBLA survey; Howes et al. 2014Howes et al. , 2016. When targeting metalpoor stars in the Galactic Centre, pre-selection is necessary considering that the vast majority of the stars in this region are metal-rich (i.e. Hill et al. 2011). Furthermore, observations of these targets in the IR with AO exploits the issues involved with a crowded and dust-ridden field. These pre-selected targets were observed with RAVEN during two engineering runs (in 2014 August and 2015 June). Each pointing involved a simultaneous observation of an additional star of similar brightness to demonstrate the multiplexing capabilities of RAVEN; these additional stars were found to be relatively metal-rich and will be presented in a companion paper. The targets were observed in different AO modes (i.e. MOAO, GLAO, and SCAO) depending on the turbulence profile of the evening (further described in Section 2.4.3).

M22 targets
Two metal-poor stars in the globular cluster M22 were selected for our programme. This cluster is directly in front of the Galactic bulge but still subject to crowding, with reddening variations of 0.10 mag in E(B−V) across the face of the cluster (Marino et al. 2011), thus an ideal target for our MOAO demonstrations. This cluster also has evidence for a spread in iron, a characteristic found in only a few globular clusters within the MW (e.g. M54 and Omega Cen). It has been argued that M22 is host to multiple stellar populations, with one of the key indicating factors being a significant spread in iron, calcium, A(C+N+O), and s-process elements between stars (e.g. Marino et al. 2009Marino et al. , 2011Alves-Brito et al. 2012). However, there is also evidence claiming the contrary (Mucciarelli et al. 2015), where it is argued the Fe spread can be explained by the large systematics involved with deriving surface gravities spectroscopically. We have selected two targets in M22 from Lane et al. (2011), who determined Fe abundances for stars in M22 from low-resolution spectra (part of the RAVE survey). Our targets include a metal-poor and a metal-rich RGB star from this sample; the spread in Fe between the two stars is ∼ 0.3 dex. In addition, we required the stars exist within a 3.5 arcmin vicinity of each other. Our goal is to determine detailed abundances of the two stars and search for element abundance differences. This demonstration can show the advantage Notes. a For the rest of this paper, we adopt this alternate name when referencing these target stars. b Measured at ∼12-16 different featureless regions across the entire H-band spectrum and reported c The median seeing taken for the entire night. d This mode was used solely to demonstrate the SCAO capabilities of RAVEN and was otherwise modes is in the text. of the multiplexing capabilities of RAVEN and the AO correction available in a crowded field, provide homogeneous observations mitigating potential systematic differences in derived abundances, and utilize the IR to overcome extinction from scattering by dust. An image of the M22 field used in this work, along with a footprint of the asterism, is shown in Fig. 1.

Standard star
One standard RGB star in M15 with known low metallicity ([Fe/H] ∼ −2.3; Mészáros et al. 2015) was observed to track the precision and accuracy of our methods. This star is well described in the literature (Carretta et al. 2009;Sobeck et al. 2011;Mészáros et al. 2015), providing a comparison sample for all of the light-element abundances derived in this work. Furthermore, the comparison sample is derived from both optical and IR spectroscopic methods, providing an excellent framework to determine the validity of our methods. The comparison is shown in Section 4.1.  The Thirty Meter Telescope (TMT), which will be located either on Mauna Kea, Hawaii or Canary Islands, Spain, is developed by a consortium of universities, foundations, and national observatories in the United States, Canada, China, India, and Japan. It has a Ritchey-Chrétien optical design and no adaptive mirror built-in. The TMT will probably start operation with a conventional 3 m class "active" secondary mirror and a facility AO device including deformable mirrors and wavefront sensors. This AO facility, called NFIRAOS (see sect. 2.2), includes deformable mirrors with about 8000 actuators in total, and will feed three instruments.

TMT first-generation instruments
First-generation instruments of the TMT using adaptive optics are IRIS  and IRMS (Mobasher et al., 2010).
IRIS (infrared imager and spectrograph) is a first-generation near-infrared (0.84-2.4 µm) instrument being designed to sample the diffraction limit of the TMT. IRIS will include an integral field spectrograph (R=4000-10000) and imaging camera (34"x34"). Both the spectrograph and imager will take advantage of the high spatial resolution achieved with the Narrow-Field Infrared Adaptive Optics System (NFIRAOS) at four spatial scales (0.004", 0.009", 0.025", 0.05"). The features of the design will enable a vast range of science goals covering numerous astrophysical domains including: solar system science, extrasolar planet August 9, 2018 0:19 ao18_nocomments 6 Hippler studies, star formation processes, the physics super-massive black-holes and the composition and formation of galaxies, from our local neighborhood to high-redshift galaxies. Description for IRIS taken from TMT (2017b) and OIR-Laboratory-UCSD (2017).
The IRMS (infrared multi-slit spectrograph) is a near diffraction-limited multi-slit near-infrared spectrometer and imager. It will be fed by NFIRAOS. IRMS will provide near-infrared imaging and multi-object spectroscopy in the spectral range 0.97-2.45 µm. The science case for IRMS includes planetary astronomy (exoplanets), stellar astronomy (massive stars, young star clusters), active galactic nuclei (co-evolution of black holes and galaxies), inter-galactic medium (interaction with galaxies at high-z), and the high-redshift universe (population III stars, cosmic re-ionization, nature of the high-z galaxies). Description for IRMS taken from TMT-UCR (2014). The data cubes are then aligned to a common center calculated using the four satellite spots (Wang et al. 2014). The satellite spots are copies of the occulted central star, generated by the use of a regular square grid printed on the apodizer in the pupil plane (Marois et al. 2006;Sivaramakrishnan & Oppenheimer 2006;Macintosh et al. 2014). The satellite spots also help convert the photometry from contrast units to flux units. No background subtraction was performed since the following steps of high-pass filtering and point-spread function (PSF) subtraction efficiently remove this low-frequency component. Further steps to remove quasi-static speckles and large-scale structures were executed outside the DRP. Each data cube was filtered using an unsharp mask with a box width of 11 pixels. The four satellite spots were then extracted from each wavelength slice and averaged over time to obtain templates of the star PSF. The Linear Optimized Combination of Images algorithm (LOCI; Lafrenière et al. 2007) was used to suppress the speckle field in each frame using a combination of aggressive parameters: dr = 5 px, N A = 200 PSF FWHM, g = 0.5, and N 0.5 0.75 = d -FWHM for the three data sets, where dr is the radial width of the optimization zone, N A is the number of PSF FWHM that can be included in the zone, g is the ratio of the azimuthal and radial widths of the optimization zone, and N δ defines the maximum separation of a potential astrophysical source in FWHM between the target and the reference PSF. The residual image of each wavelength slice was built from a trimmed (10%) temporal average of the sequence. Final K1 and K2 broadband images were created using a weighted mean of the residual wavelength frames according to the spectrum of the planet, examples of which can be found in Figure 1. These broadband images were used to extract the astrometry of the planet in each data set thanks to a higher signal-to-noise ratio (S/N) than in individual frames. To do so, a negative template PSF was injected into the raw data at the estimated position and flux of the planet before applying LOCI and reduced using the same matrix coefficients as the original reduction (Marois et al. 2010). The process was iterated over these three parameters (x position, y position, flux) with the amoeba simplex optimization (Nelder & Mead 1965) until the integration squared pixel noise in a wedge of 2×2 FWHM was minimized. The best-fit position was then used to extract the contrast of the planet in each data set. The same procedure was executed in the noncollapsed wavelength residual images, varying only the flux of the negative template PSF and keeping the position fixed to prevent the algorithm from catching nearby brighter residual speckles in the lower S/N spectral slices. To measure uncertainties, we injected the template PSF with the measured planet contrast into each data cube at the same separation and 20 different position angles. We measured the fake signal with the same extraction procedure. The contrasts measured in the 2015 November 06 and 2016 January 28 K1 data sets agreed within the uncertainties, the latter having significantly better S/N, and were combined with the weighted mean to provide the final planet contrasts.
Notes. a Macintosh et al. (2015). b This work.   Adaptive optics for extremely large telescopes 7 TELESCOPE 6-12 is the only existing telescope combining multiple large-segments in a common mount. The next generation of extremely large telescopes will all have segmented primary mirrors.

ELT first-generation instruments
E-ELT and TMT use small segments similar in size to Keck, requiring 500 and 800 segments for TMT and E-ELT, respectively. The GMT has taken a different approach; its primary mirror is an array of the largest practical (8.4.m) segments manufactured at the University of Arizona Steward Observatory Mirror Lab. A central mirror surrounded by six off-axis segments is the obvious geometry for a compact telescope structure. The GMT project considered omitting the central segment in the interest of cost reduction and operational simplicity. However, the center segment was retained because the six-mirror configuration falls short of providing the desired collecting area and has two other significant drawbacks relative to seven segments:  The image point spread function is significantly broadened with just the outer six segments.
 Phasing the primary mirror segments would be much more difficult without the center segment to anchor the array.
The seven segments share a common parent optical surface to provide 25 m diffraction limited imaging with adaptive optics and a wide field of view for natural seeing applications. The segments provide large subapertures of well-corrected wavefront so that phasing is not required for natural seeing operation or GLAO. The large gaps between segments add to the difficulty of  Cosmology. e G-CLEF: GMT consortium large earth finder, GMTIFS: GMT integral-field spectrograph, GMACS: GMT multi-object astronomical and cosmological spectrograph, GMTNIRS: GMT near-infrared spectrograph. f NFIRAOS: narrow-field infrared adaptive optics system, TMT-AGE: TMT analyzer for galaxies in the early universe, WFOS: wide-field optical spectrometer.
HARMONI, the high angular resolution monolithic optical and near-infrared integral field spectrograph, "will be used to explore galaxies in the early Universe, study the constituents of the local Universe and and characterise exoplanets in great detail." METIS, the mid-infrared imager and spectrograph, will "focus on five scientific goals: exoplanets, protoplanetary disks, Solar System bodies, active galactic nuclei, and high-redshift infrared galaxies." August 9, 2018 0:19 ao18_nocomments 8 Hippler MICADO, the multi-adaptive optics imaging camera for deep observations, "is the first dedicated imaging camera for the ELT. MICADO's sensitivity "will be comparable to the James Webb Space Telescope, but with six times the spatial resolution." MAORY, the multi-conjugate adaptive optics relay for the ELT, "is designed to work with the imaging camera MICADO and with a second future instrument ... MAORY will use at least two deformable mirrors, including the deformable mirror of the telescope. It measures the light from a configuration of six sodium laser guide stars, arranged in a circle on the sky, to obtain a kind of three-dimensional mapping of the turbulence. The laser guide stars are projected from around the circumference of the telescope's primary mirror."

GMT first-generation instruments
For the GMT, the suite of first-light instruments (Jacoby et al., 2016) using AO consists of a fiber-fed, cross-dispersed echelle spectrograph able to deliver precision radial velocities to detect low-mass exoplanets around solar-type stars, hence the name of the instrument GMT-consortium large earth finder G-CLEF . Both, the wide-field optical multi-object spectrograph GMACS (DePoy et al., 2012) and G-CLEF will benefit from the GMT's ground-layer AO observing mode . The near-infrared integral field spectrograph and imager GMTIFS (Sharp et al., 2016) and the near-infrared high-resolution spectrograph GMTNIRS Jacoby et al., 2016) will use the GMT's laser tomography AO system van Dam et al., 2016).
An overview of all three telescopes, their adaptive optics type, and their planned first suite of instrumentation is given in table 1.

Natural and artificial laser guide stars
Wavefront sensors analyze light passing through the Earth's atmosphere. The ideal light source for wavefront sensing, called natural guide star, is a stellar point like object, as close as possible to the scientific target, and sufficiently bright to minimize measurement errors. Artificial light sources created high above the telescope -for example in the ∼ 100 km high sodium layer of the mesosphere -, called laser guide stars, can be positioned as close as required to the scientific target. Their brightness depends on the used technology. Laser guide stars enormously increase the usability of AO systems. Laser guide stars as well as the technique of laser tomography AO are mentioned in this article but going into details is beyond the scope of this review. For an overview, table 2 summarizes for each telescope the plans for installing laser guide star facilities.

Wavefront sensing architecture
The GMT SCAO and LTAO architecture ( fig. 6) is designed to use reflected light from a tilted instrument entrance window for wavefront sensing, using either light from natural or laser guide stars. The reconstructed wavefront is used to shape the adaptive secondary mirror of the GMT, such delivering corrected wavefronts to the instrument. In the case of laser guide star wavefront sensing additional on-instrument wavefront sensors are required to sense low-order atmospheric aberrations. For GLAO operations, the GMT has an integrated acquisition, guiding, and wavefront sensing system (AGWS) analyzing the light in an annular field of view outside the science instruments field of view. This "technical" field of view with an inner diameter of about 6 arcmin and an outer diameter of about 10 arcmin, is large enough to find suitable natural guide stars for GLAO operations.
In contrast to the GMT, the ELT does all AO related wavefront sensing inside the instruments. An exception from this ELT AO design is MAORY, which is designed to support two instruments. Wavefront compensation is performed using the ELT's internal adaptive mirrors M4 and M5 as shown in fig. 4. Active optics and possible GLAO support is done in a similar way as for the GMT, i.e. using a "technical" field of view.
How does the TMT compare with the GMT or ELT AO architecture? The TMT has a facility MCAO/SCAO system called NFIRAOS (Herriot et al., , 2017, which can feed three instruments: an infrared imaging spectrograph IRIS, an infrared multi-slit spectrometer IRMS, and a third future instrument. From this perspective, NFIRAOS is similar to the ELT's MAORY "relay" system, which feeds MICADO and a second future instrument. In its current design, the TMT itself has no adaptive deformable mirror built-in, instead, the wavefront correction devices are located inside NFIRAOS. A laser guide star facility (LGSF) will provide at least six sodium laser guide stars (Li et al., 2016) feeding NFIRAOS, thus enabling LGS MCAO observations . In September 2017, a design study for an adaptive secondary mirror was launched (TMT, 2017a). In particular, the wide-field optical spectrometer (WFOS) would benefit from such an integrated wavefront correction unit, as it is not connected to the NFIRAOS unit, and the LGSF is designed to provide various asterisms, including a GLAO asterism (TMT, 2018a). Table 3 summarizes the main characteristics of the wavefront sensors (WFS) used with the first light and first-generation respectively instrumentation.

Wavefront correction devices
This brings us to the topic of wavefront correction devices, their characteristics and location within each of the described telescopes and instruments respectively. As shown in fig. 4, the ELT comes with a deformable mirror M4 and a fast steering mirror M5. The 2.4 x 2.5 m sized elliptic mirror M4 is supported by 5316 contactless voice-coil actuators . This mirror can be controlled at frequencies up to 1 kHz. Together with the fast and flat steering mirror M5, which compensates for rather large amplitude atmospheric tip-tilt aberrations (field stabilization), this combination in interplay with the wavefront sensing devices, allows GLAO, SCAO, and LTAO observations. The ELT's multi-conjugate adaptive optics relay MAORY adds another two deformable mirrors for instruments attached to it. Each of them with 500-1000 actuators (actuator number and technology not yet decided). The GMT has two exchangeable secondary mirrors, each 3.25 m in diameter, segmented and shaped similar to the primary mirror ( fig. 5). The fully adaptive secondary mirror (ASM) is supported by 4704 contactless voice-coil actuators. An alternative fast steering secondary mirror (FSM) with 7 rigid circular segments (Lee et al., 2017) will be used during commissioning of the telescope and later on during ASM maintenance or repair.
Wavefront correction at the TMT takes place in it's NFIRAOS facility ( fig. 7) operated at a temperature of -30 degrees. A tip-tilt stage and two deformable mirrors with a total number of 7673 actuators (3125 actuators for the ground layer DM and 4548 actuators for the high-layer DM) support MCAO observations. As mentioned in section 2.2, TMT organization has launched a design study for an adaptive secondary mirror in September 2017. Table 4 summarizes some basic parameters of the planned wavefront correction elements for all 3 telescopes. From today's perspective, piezo actuator technology is still in use, voice-coil actuator technology has become a standard, and bimorph mirror technology has disappeared. Liquid crystal spatial light modulators do not play a role in the context of AO for the first light instrumentation of the 3 telescopes. The role of micro electro-mechanical systems (MEMS) based deformable mirrors is rather small, only the GMT considers a MEMS DM for some of their first light instrumentation (Copeland et al., 2016).
A further increase in actuator density, i.e. for future XAO systems, from 5-10 actuators/m 2 to 12.5-25 actuators/m 2 seems possible but has to be developed (Madec, 2012;Riaud, 2012;Kasper et al., 2013;Kopf et al., 2017). For high-contrast XAO systems, high-density deformable mirrors might be used to remove residual speckles seen in the science focal plane. In combination with the science focal plane detectors behind coronagraphs, a second stage AO system can be set-up in order to calibrate and remove residual speckles seen in the science images (Macintosh et al., 2006b;Thomas et al., 2015).

AO real-time control systems
The last major component of an AO system is the real-time control system (RTC), which receives wavefront measurements from the wavefront sensors, calculates wavefront estimates, and eventually sends control signals to the wavefront correction devices like deformable mirrors and tip-tilt mirrors. For the SCAO case, the largest size of an interaction matrix using numbers from tables 3 & 4 has about 5000 actuators and 10000 wavefront x-y gradients. The time available due to AO latency requirements, i.e. the time to • receive pixel data from the wavefront sensors, • pre-process the raw wavefront sensor data, • calculate the wavefront slopes from calibrated pixel data, • reconstruct the wavefront using for example a matrix-vector multiplication (MVM), • calculate the new shapes and positions of the wavefront corrections devices including tasks like vibrations control, deformable mirror saturation management, etc., • send control commands to the wavefront correction devices, is rather short and of the order of 0.2-2 milliseconds Kerley et al., 2016;Bouchez et al., 2014). This "low latency" requirement defines the control bandwidth of the AO system as well as the stability and robustness of the closed-loop system. Looking at the matrix-vector multiplication, one of the August 9, 2018 0:19 ao18_nocomments 12 Hippler most time consuming operations in this list, year 2016 off-the-shelf hardware is compliant with the latency requirements. Exemplary performance tests using Intel's Xeon Phi (Knights Landing, KNL) CPU with 68 cores already shows satisfactory results as shown in fig. 8.
A schematic and simplified view of the real-time control system of the ELT instrument METIS is shown in fig. 9. The core of METIS SCAO real-time control system (AO RTC) consists of a hard realtime controller (HRTC) and a soft real-time controller (SRTC). While the low-latency HRTC processes WFS data and generates control commands for the control systems of the correction devices, the soft realtime controller receives additional wavefront correction signals from the science focal planes (FPs) and their detector control systems (DCSs), respectively. Such additional wavefront corrections signals can for example compensate non-common path aberrations between the science and WFS light path, through adding slope offsets to the actual measured wavefront slopes. The generated correction commands can be either sent directly to the ELT's M4 and M5 local control systems (LCS) or through the ELT central control system (CCS). Further optical components within the common-path of the METIS science and WFS channel are an optical de-rotator and the pupil stabilizer (pupil stabil.). Monitoring the pupil position can be performed on the SRTC and if necessary correction commands generated and sent to the pupil stabilization controller. The field selector inside the WFS channel is pre-set via the AO observation coordination system (OCS) and the pyramid wavefront sensor modulator has to be synchronized with the SCAO WFS detector control system. Wavefront control taking into account non-common path aberrations (NCPAs) is achieved through a communication channel between the science detectors and the SRTC. The AO function control system (FCS) completes the METIS SCAO control system.
To investigate real hardware options supporting all first-generation ELT instruments, the Green Flash project  aims in building a prototype/demonstrator designed to support all AO systems. After having investigated GPU clusters like the NVIDIA DGX-1 (Bernard et al., 2017), Intel Xeon August 9, 2018 0:19 ao18_nocomments Adaptive optics for extremely large telescopes 13   For the TMT's multi-conjugate AO system NFIRAOS, more real-time tasks, compared to the METIS SCAO RTC, are required to process the data acquired by up to six high-order wavefront sensors. The sequence of the high-order AO data flow starts with pixel processing and wavefront reconstruction (HOP) for each wavefront sensor. A wavefront corrector control server (WCC) combines all measurements including METIS SCAO control schemes for wavefront control, pupil stabilization and NCPA compensation. inside the gray box are part of METIS SCAO. Red arrows indicate measurement signal flow, blue arr ow. Red boxes represent sensors, black boxes represent control instances. The HRTC within AO RT for the core wavefront control loop, data products that rely on WFS telemetry are computed in the SR control tasks, such as registration management, modulation, and field selection will be performed outsid C. For NCPA compensation, data from the Science Detectors will be used as well as AO telemetry data. Fig. 9. Functional diagram of the METIS SCAO control system including non-common path aberrations control . See text for details.
low-order (LO) wavefront errors (full aperture natural guide star tip/tilt as well as laser guide star tip/tilt), and finally sends the correction signals to the high-order deformable mirrors DM0 and DM11 as well as to a tip-tilt stage (TTS), the laser guide star fast steering mirrors (FSM), and the telescope control system. The TMT RTC design allows to run the high-order pixel processing and wavefront reconstruction using a matrix-vector multiplication (MVM) almost in parallel. Estimated start and execution times for the various tasks are show in fig. 10. The GMT's wavefront control system (WFCS) runs as an integral part of the telescope control system . The most demanding real-time requirements are set by the LTAO mode with a latency ≤ 200 µs. The GMT's baseline WFCS design foresees 9 commodity servers, one slope processor for each of the 6 laser tomography wavefront sensors, 1 node for communication with the adaptive secondary mirror, 1 node for communication with the on-instrument deformable mirror, and 1 master node for combining all reconstructed wavefronts.
Whether commercial off-the-shelf components and commodity servers can be used for the required high performance AO controllers is currently under investigation. Table 5 lists hardware components under consideration at the particular telescopes. The table brings into focus the high performance computing (HPC) power only, while other hardware components like low-latency, high bandwidth networks (e.g. 10-100 Gbit-Ethernet, Infiniband, Omni-Path), sufficiently fast input/output channels (e.g. Camera Link, GigE Vision, sFPDP, USB3), and high throughput data storage are considered uncritical as it can be categorized as standard hardware. bandwidth memory) which can be used either directly or as L3 cache. Each core can simultaneously execute up to four threads. Single threaded performance of the KNL Xeon Phi is expected to be lower than single core performance of a Xeon CPU, but for use within a HOP server, a single KNL chip should be able to replace four Xeon E7 CPUs due to the high memory bandwidth of the MCDRAM (greater than 400GB/s).

Figure 2 -Estimated Pipeline Stage Execution Times
The Knights Landing Xeon Phi can use sub-NUMA (non-uniform memory access) clustering to appear to the operating system as a four socket Xeon server. This configuration mode provides the lowest memory latency if the workload is highly NUMA optimized. The RTC software will be optimized for NUMA architectures. To make the HOP software less dependent upon hardware selection, time critical HOP server software will assume a four-way NUMA architecture. The HOP server software will use configuration files to allow control of the number of threads and their assignment to physical cores.
The use of the Knights Landing Xeon Phi is expected to increase system latency since the control matrix must be fetched from the MCDRAM for each frame (it is much too large to fit in the on-core caches), but is expected to significantly reduce system cost, power consumption, and size.

AO performance: requirements and expectations
Usually the scientific program determines the necessary minimum requirements for the measuring instrument. As an example, for the ELT METIS instrument various dedicated AO related top-level requirements influence the AO design. These are among others requirements on Strehl and contrast, for example, METIS shall deliver a minimum Strehl ratio of 60% (goal 80%) in L-band (λ=3.7µm). For the contrast, METIS shall deliver a 5σ contrast in L-band of 3×10 −5 (goal: 1×10 −6 ) at a distance to the central PSF of 5λ/D (goal: 2λ/D). Detailed end-to-end simulations are in the design phase the only way to figure out whether the design meets the requirements. Table 6 lists requirements on the AO for the instruments described in table 1 and, where available, simulated performance numbers.

High-contrast imaging requirements on adaptive optics
High-contrast imaging astronomy aims at uncovering faint structures and substellar objects very close to bright stars (an overview on this topic gives for example Oppenheimer & Hinkley (2009)). Investigating distances close to the diffraction limit, i.e. ∼10-1000 mas, requires sophisticated AO installations, like XAO systems, that are able to spatially separate the light from, for instance, a host star and an orbiting planet.
a L=LTAO, M=MCAO, S=SCAO b Natural guide star spectral band and magnitude or fraction of the sky observable. c METIS requirements and estimates given for on-axis PSF. The METIS total corrected field of view is ∼ 12"x12" (Brandl et al., 2018). d GMTIFS field of view. The larger field of view of 20.4" x 20.4" is for the IFS imager .
A good measure to characterize the quality of an AO system is to look at the long and short exposure Strehl ratios it delivers. Together with the residual image motion, also for long and short exposures, one can use these static and dynamic performance data to estimate the imaging quality and the raw contrast delivered to the science channel. Per definition, the Strehl ratio defines the ratio of the measured AO corrected peak intensity of the point spread function (PSF) to the intensity measured or estimated of a perfectly imaged point spread function. If the delivered long exposure Strehl ratio at a wavelength of 3.7 µm is 95%, there is a fraction of 5% of intensity or energy in the halo of the stellar PSF. As the structure and the dynamic behaviour of this un-controlled halo is unknown, it contributes strongly to the achievable contrast. This becomes clear if we look at the standard technique used in high-contrast imaging, where a August 9, 2018 0:19 ao18_nocomments Adaptive optics for extremely large telescopes 17 stellar reference PSF is subtracted from the actual recorded stellar PSF.
In combination with a coronagraph, contrast values of point sources below 10 −4 at angular separations of 5 λ/D on a D=8 m telescope can be achieved. As shown in fig. 11, using the adaptive optics instrument NACO at the 8 m VLT with an annular groove phase mask (AGPM), different contrast values can be achieved depending on the applied data reduction algorithm. The pure adaptive optics intensity profile of a standard star, which shows structures around the 3rd bright Airy ring and at the edge of the AO control radius, can be reduced by a factor of approx. 20 at an angular separation of 0.5 arcsec (5 λ/D) with λ=3.8 µm. The subtraction of the stellar PSF using the angular differential imaging (ADI, Marois et al. (2006)) method together with a principal components analysis (PCA, Gomez Gonzalez et al. (2016)) results in higher contrast values towards smaller angular separations. The rather flat contrast values at angular separations beyond the AO control radius indicate the highest contrast achievable at the smallest angular separations for this observation set-up. As the green curve in fig. 11 shows, there are still about 2 orders of magnitude contrast improvements possible. For that reason it is important to understand in detail the contributions of the non-corrected aberrations to this regime.
As Mawet et al. (2012) point out, in particular low-order aberrations contribute to the contrast at angular separations between 1 and 4 λ/D. One proposal to further improve the raw contrast behind an AO fed coronagraph is predictive control. Males & Guyon (2018) show that the raw contrast for bright stars can be increased by more than 3 orders of magnitude at an angular separation from the PSF center of 1 λ/D, for λ=800 nm and D=25.4 m (GMT case). This corresponds to a raw contrast improvement of ∼50 at λ=3.8 µm.
ence of the companion ntrast, we took another entative ADI sequence on ndard star (HD123888). test was performed under ns (seeing ~ 0.85 arcsec nefited from our improved he new mode (efficiency s better than during the first r instantaneous attenuation d. To calibrate our detection flux losses induced by PCA, ke companions (at 15s) nd measured their through . We used the derived ap to renormalise the initial a posteriori (Figure 6). This nt detection capabilities A of the AGPM. The final trast presented here (green e), is limited by the small PA ally at small angles. The beyond an offset of 1-arc to the background at L, er for brighter targets and/ rations. d Instrumentation Figure 6. Normalised azimuthally aver aged relative intensity profiles and contrast curve on a log-log scale. The plain red curve shows the intensity profile of a typical saturated NACO L PSF (similar brightness and exposure time). The blue dashed curve shows the AGPM intensity profile before PCA, demonstrating the instantaneous con trast gain provided by the corona graph at all spatial frequencies within the adaptive optics control radius (~ 0.7 arcseconds). The green dashdot curve presents the reduced PCA-ADI 5s detectability limits (40 frames, 800 s, ∆PA ≈ 30˚), taking both the cor onagraph offaxis transmission and the PCA-ADI flux losses into account. . VLT NACO coronagraphic observations of HD 123888 in L'-band (λ=3.8µm). Normalised azimuthally averaged relative intensity profiles and contrast curve on a log-log scale. The plain red curve shows the intensity profile of a typical saturated NACO L' PSF (similar brightness and exposure time). The blue dashed curve shows the AGPM intensity profile before PCA, demonstrating the instantaneous contrast gain provided by the coronagraph at all spatial frequencies within the adaptive optics control radius (∼0.7 arcseconds). The green dash-dot curve presents the reduced PCA-ADI 5σ detectability limits (40 frames, 800 s, position angle difference ≈ 30 • ), taking both the coronagraph off-axis transmission and the PCA-ADI flux losses into account. Figures and caption taken from Mawet et al. (2013). See text for further explanations.
Another way to further reduce low order aberrations is to increase the speed of the AO system in case there are enough photons available from the AO reference source. An exemplary case for the bright (K=4.5 mag) star 51 Eridani is shown in fig. 12. Increasing the AO loop frequency from 1 kHz to 2 kHz improves the coronagraphic contrast by a factor ≈3. Guyon et al. (2012) have shown similar behaviour for a perfect AO system on a 30 m class telescope, simulating wavefront sampling frequencies up to 100 kHz and limiting the closed-loop servo latency to 0.1 ms. As shown in fig. 13, raw H-band PSF contrasts at small separations between 10 to 20 mas of ≈ 1.e-5 can be achieved.
This indicates that the stability of the AO controlled PSF is a critical factor for high-contrast imaging using coronagraphs and further post-processing techniques. This is long known but the impact on contrast is striking. In particular PSF subtraction in combination with ADI requires careful execution as it can subtract extended structures like circumstellar disks (Soummer et al., 2012). Measuring the reference PSF for post-processing PSF subtraction is a critical task in this context and requires that the AO performs as uniform as possible.   . This plot is to demonstrate that rather small changes of the long-exposure Strehl number from 0.978 at 1 kHz AO loop speed to 0.984 at 2 kHz AO loop speed results in a contrast enhancement of a factor ∼3 at at separation of 5λ/D. For comparison, in fig. 11 this separation is at ∼0.5 arcsec. Table 7 provides an overview of the planned high-contrast instruments and the expected years of commissioning. Some of these instruments will also use the technology of high dispersion coronagraphy (HDC) not further explained here. A recent study on how exoplanets can be observed with HDC from the ground as well as with space telescopes can be found in Wang et al. (2017).

Conclusions and outlook
The technique of adaptive optics has matured over the past thirty years. Permanently installed on all future extremely large telescopes currently under design and construction, it will allow for diffractionlimited observations for wavelengths in the near-infrared regime and longer. Expanding the use of AO in the visible spectrum remains challenging -a niche observing mode for the very bests nights as Close et al. (2017) suggest -, but we should keep our eyes on this.
All instruments under design for the ELT, GMT, and TMT can use AO, hence make use of the huge light collecting power and the maximum achievable angular resolution. Laser guide star facilities push the AO sky coverage to levels well above 50%.
Highly specialized instruments will further boost ground-based diffraction limited high-contrast imaging and characterization of exoplanets to unprecedented planet-star contrast ratios. The entire nearby Alpha August 9, 2018 0:19 ao18_nocomments Adaptive optics for extremely large telescopes 19 4.3 Expected contrast The raw PSF contrast is estimated in Figure 8 for a mI=8.5 target. In the 10 to 20 mas angular separation range where most of the exoplanets are imaged, the contrast is limited by time lag in the loop and photon noise, and the other fundamental limits to raw contrast (scintillation and atmospheric chromaticity effects) are much smaller. With a high efficiency wavefront sensor able to take advantage of the telescope's diffraction limit, the expected raw PSF contrast at these small separations is approximately 1e-5, provided that the servo lag is no more than about 0.1 ms. This unusually low servo lag can be achieved with a high WFS sampling frequency (>10 kHz), and/or the use of predictive wavefront control techniques. Figure 8 also shows that a seeing-limited WFS such as the SHWFS is very inefficient at these small angular separations 18 , and would be a poor choice for the system, even if it operates at its photon-noise limit with no loop servo lag other than the one imposed by photon noise. Much better choices include the Pyramid wavefront sensor (with little or no modulation) and the non-linear Curvature WFS 19 , currently under development, and soon to be tested on sky on the Subaru Telescope and the 6.5 m MMT telescope.
The analytical model used to estimate raw contrast was also tested for an 8 m diameter telescope under the same conditions. For a 1 kHz system with a diffraction-limited wavefront sensor on an 8 m telescope, the raw contrast at 0.1" is 3e-4 (limited by servo lag), and it is 3e-5 at 0.5". These numbers are consistent with the goals of the future Extreme-AO systems on such telescopes. The detection contrast limit is more difficult to estimate for this system, as a range of PSF calibration techniques could be used (spectral or polarimetric differentiation for example). For simplicity, it is assumed here that spectral or polarimetric PSF calibration techniques are not used, and that the detection limit is imposed by speckle structure in the long-exposure image and photon noise. It is also assumed that static and slow speckles that are not due to the atmosphere are removed by focal plane wavefront control, a scheme that has already demonstrate control and removal of static coherent speckles at the 3e-9 contrast level in the presence of much stronger dynamic speckles.
The PSF halo consists of rapid atmospheric speckles at the 1e-5 contrast level with a lifetime of no more than one millisecond (speckles of longer duration are suppressed by the AO loop). In a one-hour observation, this fast component can thus average to 5e-9 contrast assuming that the AO system has removed correlation on timescales above 1ms. In addition to these fast speckles, chromatic non-common path errors and scintillation create a speckle halo contribution at Proc. of SPIE Vol. 8447 84471X-10 Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 6/29/2018 Terms of Use: https://www.spiedigitallibrary.org/terms-of-use Fig. 13. PSF raw contrast of a 30 m telescope at λ=1.6µm (H-band) vs. angular separation for wavefront sampling frequencies between 5 and 100 kHz. Wavefront sensing at λ=0.8µm (I-band) using an ideal wavefront sensor. Shown are individual contributions to the contrast due to atmospheric scintillation, scintillation chromaticity, and atmospheric refraction chromaticity at zenith. A Shack-Hartmann wavefront sensor (SHWFS) not taking advantage of the 30 m telescope diffraction limit cannot deliver the performance of an ideal wavefront sensor. See text for further explanations.  ELT METIS, mid-infrared ELT thermal imager and spectrograph 2025 1 st generation instr., Kenworthy et al. (2016).

ELT
Planetary camera and spectrograph (PCS, also called EPICS) 2028 2 nd generation instr., Kasper et al. (2013). Centauri system, including the planet Proxima Centauri b (Kreidberg & Loeb, 2016) orbiting in the habitable zone, has received strong attraction in the science community, accompanied by an intense public attention. The pure imagination to find nearby Earth-like planets in habitable zones has triggered ideas to fly there for further investigation (Popkin, 2017).

Acknowledgments
I would like to thank my colleagues Markus Feldt, Dietrich Lemke, and Kalyan Radhakrishnan for their advice and comments. Special thanks to the anonymous reviewer for his/her very valuable comments and suggestions to improve the quality of this review. I thank the following organizations and journals for allowing me to reproduce the figures listed below