Recent variability of the solar spectral irradiance and its impact on climate modelling

The lack of long and reliable time series of solar spectral irradiance (SSI) measurements makes an accurate quantification of solar contributions to recent climate change difficult. Whereas earlier SSI observations and models provided a qualitatively consistent picture of the SSI variability, recent measurements by the SORCE satellite suggest a significantly stronger variability in the ultraviolet (UV) spectral range and changes in the visible and near-infrared (NIR) bands in anti-phase with the solar cycle. A number of recent chemistry-climate model (CCM) simulations have shown that this might have significant implications on the Earth's atmosphere. Motivated by these results, we summarize here our current knowledge of SSI variability and its impact on Earth's climate. We present a detailed overview of existing SSI measurements and provide thorough comparison of models available to date. SSI changes influence the Earth's atmosphere, both directly, through changes in shortwave (SW) heating and therefore, temperature and ozone distributions in the stratosphere, and indirectly, through dynamical feedbacks. We investigate these direct and indirect effects using several state-of-the art CCM simulations forced with measured and modeled SSI changes. A unique asset of this study is the use of a common comprehensive approach for an issue that is usually addressed separately by different communities. Omissis. Finally, we discuss the reliability of the available data and we propose additional coordinated work, first to build composite SSI datasets out of scattered observations and to refine current SSI models, and second, to run coordinated CCM experiments.

ability, recent measurements by the SORCE satellite suggest a significantly stronger variability in the ultraviolet (UV) spectral range and changes in the visible and near-infrared (NIR) bands in anti-phase with the solar cycle. A number of recent chemistry-climate model (CCM) simulations have 10 shown that this might have significant implications on the Earth's atmosphere. Motivated by these results, we summarize here our current knowledge of SSI variability and its impact on Earth's climate.
We present a detailed overview of existing SSI measure-15 ments and provide thorough comparison of models available to date. SSI changes influence the Earth's atmosphere, both directly, through changes in shortwave (SW) heating and therefore, temperature and ozone distributions in the stratosphere, and indirectly, through dynamical feedbacks. We in-20 vestigate these direct and indirect effects using several stateof-the art CCM simulations forced with measured and modeled SSI changes. A unique asset of this study is the use of a common comprehensive approach for an issue that is usually addressed separately by different communities. 25 We show that the SORCE measurements are difficult to reconcile with earlier observations and with SSI models. Of the five SSI models discussed here, specifically NRLSSI, Correspondence to: I. Ermolli (ilaria.ermolli@oaroma.inaf.it)

Introduction
The question of whether -and to what extent -the Earth's climate is influenced by solar variability remains central to tainty in this figure because several aspects of solar forcing and the different mechanisms by which solar variability influences the Earth's environment are still poorly understood (see e.g. Pap et al., 2004;Calisesi et al., 2007;Haigh, 2007;Gray et al., 2010;Lockwood, 2012, and references therein).
For these reasons, the quantification of solar contribution to climate change remains incomplete. This is further highlighted by some of the most recent investigations of solar spectral irradiance (SSI) variations and estimates of their influence on the Earth's atmosphere based on chemistry-70 climate model (CCM) simulations. Regular space-based measurements of the solar irradiance started in 1978. The total solar irradiance (TSI), i.e. the spectrally integrated radiative power density of the Sun incident at the top of Earth's atmosphere, has been moni-75 tored almost continuously and was found to vary on different time scales (Willson et al., 1981;Fröhlich and Lean, 2004;Fröhlich, 2009). Most noticeable is the ≈ 0.1 % modulation of TSI in phase with the 11-yr solar cycle. Changes 2-3 times larger than this are observed on time scales shorter than few 80 days. Measurements of SSI, however, are not continuous over the satellite era and until recently have concentrated on the ultraviolet (UV) radiation, because of the larger relative variability of SSI below 400 nm (Fig. 1) and the impact of these wavelengths on the terrestrial atmosphere through ra-85 diative heating and ozone photochemistry. SSI variations differ from those observed in the TSI. The variability of visible and NIR bands barely exceeds 0.5 % over a solar cycle; in the near UV and shorter wavelengths variability increases with decreasing wavelength, reaching 90 several percent at 200-250 nm, and several tens of percent, and even more, below about 200 nm (e.g. Floyd et al., 2003, and references therein). These bands are almost completely absorbed in the Earth's middle and upper atmosphere (Fig. 1) and are the primary agent affecting heating, photochemistry, 95 and therefore, the dynamics of the Earth's atmosphere. Variations of the solar UV radiation between 120 and 350 nm lead to changes in stratospheric ozone and heating that amplify the effect of the UV radiation in the Earth's atmosphere, possibly also through indirect mechanisms (e.g. the "top-100 down" mechanism, Gray et al., 2010). Hence, although the UV radiation shortward of 400 nm represents less than 8 % of the TSI, its variability may have a significant impact on climate. In contrast, the visible and IR bands, which have the largest contribution to the TSI, small variations over the 105 solar cycle, and no absorption in the atmosphere but in some well-defined IR bands, directly heat the Earth's surface and the lower atmosphere. The large amount of solar flux at the visible and IR bands implies that small flux differences may induce important terrestrial consequences. The impact of the 110 variability of these bands on the Earth's climate is expected I. Ermolli Woods and Rottman, 2002), UARS/SUSIM (green, only in UV; Morrill et al., 2011a), NRLSSI (black), SATIRE-S (blue), COSI (purple), OAR (light blue), SCIAMACHY (brown), SORCE (red), and SORCE re-analysis (orange, only in UV; Woods, 2012). The exact wavelength ranges used for SUSIM and SCIAMACHY in the UV are 150-400 nm and 240-400 nm, respectively. The possible related corrections are, however, expected to lie within 2-3 %. Note that for the SCIAMACHY-based model, the original values listed by Pagaran et al. (2009) are shown. As discussed by Krivova and Solanki (2013), these values should most likely be corrected by a factor of roughly 1.2.
Early satellite measurements of the solar UV vari-115 ability have shown a qualitatively consistent behaviour (DeLand and Cebula, 2012), which is fairly well reproduced by SSI models (e.g. Lean et al., 1997;Krivova et al., 2006;Unruh et al., 2012;Lean and DeLand, 2012). This situation changed with the launch of the Spectral Irradiance Moni-120 tor instrument (SIM, Harder et al., 2005a) onboard the Solar Radiation and Climate Experiment satellite (SORCE, Rottman, 2005) in 2003, which was shortly after the most recent maximum of solar activity. The SORCE/SIM data showed a four to six times greater decrease of the UV radi-125 ation between 200 and 400 nm over the period 2004-2008(Harder et al., 2009, part of the declining phase of solar cycle 23, compared to earlier measurements and models (Ball et al., 2011;Pagaran et al., 2011a;Unruh et al., 2012; DeLand and Cebula, 2012; Lean and DeLand, 2012). This 130 larger decrease measured in the UV (Fig. 2), which exceeds the TSI decrease over the same period by almost a factor of two, is compensated by an increase in the visible and NIR bands. Variability out-of-phase with solar activity is indeed predicted by some SSI models in the NIR, but with a sig-135 nificantly lower magnitude than found by SORCE/SIM. Details are provided in the following. The inverse variability observed by SIM in a wide integrated band in the visible was, however, unexpected. It can be interpreted as a result of effects induced by the evolution of surface magnetism 140 in the solar atmosphere (e.g. Harder et al., 2009). However, other observations and analyses of existing long-term SSI data show results in contrast with those derived from SORCE/SIM (Wehrli et al., 2012).
When used as solar input to CCM simulations, 145 SORCE/SIM observations lead to significantly larger shortwave (SW) heating rates in the upper stratosphere compared to results obtained by using the commonly utilised NRLSSI model data Lean, 2000), and a decrease of stratospheric ozone above an altitude of 45 km 150 during solar maximum . These changes in radiative heating and ozone photochemistry in the stratosphere also impact the responses of the "top-down" solar UV mechanism in the Earth's atmosphere and at the surface (Kodera and Kuroda, 2002), which may depart from cur-155 4 I. Ermolli et al.: Spectral irradiance and climate rent understanding (Cahalan et al., 2010;Ineson et al., 2011;Merkel et al., 2011;Oberländer et al., 2012;Swartz et al., 2012). Validation of the results of these model simulations with ozone, zonal wind, or temperature measurements is difficult, because the data are sparse and do not cover 160 enough solar cycles. Although some ozone observations seem to agree with model calculations (e.g. Haigh et al., 2010;Merkel et al., 2011), it should be noted that the SIM measurements employed for the analysis covered less than one solar cycle and required extrapolation over a full cycle, 165 and therefore, added uncertainty (Garcia, 2010). Also the transition altitude from in-phase (lower and middle stratosphere) to out-of-phase (upper stratosphere) ozone signals with the solar cycle is not consistent among the different models and requires further investigations. 170 Unfortunately, observations of the full solar spectrum will likely have a multi-year gap before the next generation SSI instrument is launched. Based on data presently available, a thorough understanding of the impact of SSI on climate requires verification and validation of existing SSI measure-175 ments for internal consistency, calculations of middle atmosphere climate models with different reliable scenarios of SSI variations, and comparison of measurements and model results with climate records, i.e. a study involving the coordinated work of various research communities, which is part of 180 the COST Action ES1005 1 to which most of the authors of this paper belong. This paper summarises and compares, for the first time, a large number of SSI observations and models, and discusses the impact of these data on Earth's climate. A review 185 in this area was recently published by Gray et al. (2010), but their focus was more on different possible forcing mechanisms. Although the Sun affects the climate system in numerous ways, we here focus on radiative forcing only, with particular attention given to the role of the SSI rather than 190 that of the TSI, which is still the sole solar input in many climate models. Since space-based observations of the SSI and of the terrestrial atmosphere conditions are sparse or absent before 1980, we restrict our analysis to the data of the last three decades, i.e. roughly the last three solar cycles. Fur-195 thermore, we concentrate on the effects of the UV variability because of its potentially large impact on the terrestrial atmosphere.
The paper is organised as follows: in Sect. 2, measurements of SSI variations are described, and their accuracy 200 on time scales from days to 11-yr solar cycle is discussed. In Sect. 3 we delineate mechanisms responsible for the SSI variations, outline methods of irradiance reconstructions and briefly describe and compare several of the most broadly used models of SSI variations. Section 4 discusses the impact 205 of the current lower and upper boundaries of SSI solar cycle estimated variations on the atmospheric response in several 1 TOSCA -towards a more complete assessment of the impact of solar variability on the Earth's climate, http://www.tosca-cost.eu. current state-of-the art CCMs. A summary and concluding remarks are provided in Sect. 5. 210 We first discuss the available solar irradiance measurements, focussing on the most recent data. These data are input to and, at the same time, the main source of constraints for both the SSI models and CCM simulations presented below. We describe the evolution of measurements carried out from 215 space since 1978 and discuss major instrumental and measurement problems that limit the creation of single composite time series from existing records (Sect. 2.1). Long and accurate SSI time series are critical to obtain reliable estimates of the solar variability impact on the terrestrial atmosphere. 220 We then present instruments and techniques used to derive four SSI data sets that are employed for CCM simulations (Sect. 2.2). Finally, we discuss recent re-assessment of the TSI absolute value (Sect. 2.3,) since it enters some climate and SSI models, and outline the perspectives of future SSI 225 observations (Sect. 2.4).

Observation of solar irradiance
For many centuries, the Sun has been considered as an example of stability and, not surprisingly, the TSI, i.e. the quantity of radiative power density (Wm −2 ) at normal incidence on 230 top of the atmosphere, at a Sun-Earth distance of one astronomical unit, has been called solar constant. This term is still often used. However, regular space-based observations that started in 1978 (Willson et al., 1981) have revealed that the TSI varies over time scales of minutes to decades, and 235 probably even longer (e.g. Fröhlich, 2006Fröhlich, , 2009. The most noticeable variation of the TSI is a 0.1 % modulation in phase with the solar activity or sunspot cycle. Although quantifying such small variations is a major technological challenge, it is strongly motivated by the desire 240 to understand solar variability, and even more so by the importance of the TSI for terrestrial energy budget studies (Trenberth et al., 2009).
The TSI is the spectral integral of SSI over all wavelengths but its weak variability masks the fact that relative SSI varia-245 tions show a strong wavelength dependence (Fig. 1). In particular, the visible and NIR bands are the least variable of the solar spectrum with a relative solar cycle amplitude of the same order as for the TSI (0.1 %), whereas values of 1 to 100 % are observed in the UV variations, and in excess of 250 100 % in the soft X-ray range (below 10 nm). Each individual spectral band has a markedly different impact on the terrestrial atmosphere, which depends on the atmospheric processes affected by the given band, the amount of the spectral flux, and its variation.  The longest records of SSI measurements were provided by SOLSTICE and SUSIM aboard the UARS (Upper Atmosphere Research Satellite) spacecraft . These instruments observed solar UV radiation between 120 and 400 nm from 1991 to 2001 and 2005, respectively. These 270 measurements pointed to the importance of the irradiance variations in the UV Rottman et al., 2001;Floyd et al., 2002Floyd et al., , 2003, although the solar cycle variability of solar radiation above approximately 250 nm remained relatively uncertain due to insufficient long-term 275 stability of the instruments (Woods et al., 1996). In addition to SOLSTICE and SUSIM, there is the long time series of 200-400 nm solar UV observations by several NOAA SBUV instruments, which had underflight calibrations aboard the Space Shuttle . Table 1 280 of DeLand and Cebula (2012) summarises the measurement uncertainties for these instruments.
Merging all UV observations into a single homogeneous composite record is a major challenge (DeLand et al., 2004) that is hampered by several problems. First, the lifetime of most instruments does not exceed a decade. This makes 295 long-term observations covering periods that exceed single instrument lifetimes, of prime interest for climate models, very difficult. A second obstacle is the differing technologies and modes of operation of various space-based instruments. The cross-calibration of individual records is further 300 hampered by the fact that overlapping observations disagree, and the existing data records are spectrally and temporally intermittent. Although missing observations can be filled in by using data regression based on time series of solar proxies such as the Mg II index, which are well correlated with UV 305 variations (DeLand and Cebula, 1993;Viereck et al., 2001;Lean, 1997), none of the existing solar proxies can properly reproduce solar irradiance in a spectral band on all time scales (Dudok de Wit et al., 2009).
The most critical issue for all SSI instruments is the op-310 tical degradation caused by the energetic radiation in the space environment. Two options have been employed to account for instrumental degradations. The first one is to provide redundancy in the instrument design, by using, e.g. a dual spectrometer setup with detectors that ex-315 perience different accumulated exposure time or by planning redundancies in spectral channels and calibration lamps from which degradation corrections can be derived. This approach was used, for instance, for UARS/SUSIM and SORCE/SIM (Brueckner et al., 1993;Harder et al., 2005a).

320
The second option is to use reference data, e.g. solar irradiance measurements for long-term calibrations, as reported for the NOAA SBUV instruments, periodic recalibrations using sounding rockets, or stable external calibration targets like selected stars, as done for UARS/SOLSTICE and 6 I. Ermolli et al.: Spectral irradiance and climate SORCE/SOLSTICE (McClintock et al., 2005). Sensitivity changes and degradation are strongly wavelength-dependent, which makes creating a properly cross-calibrated SSI record very difficult. One attempt for a composite SSI time series in the UV is provided by DeLand and Cebula (2008).

330
Cross-calibration of different SSI records is also limited by the lack of realistic confidence intervals for the existing data. This aspect has been thoroughly investigated for the TSI and sound estimates have been obtained for time scales from days to years. No meaningful estimates, however, exist for time 335 scales exceeding one decade, which are needed for climate studies. The situation is much worse for SSI measurements, whose relative uncertainties often are two to three orders of magnitude larger.
In the following subsections a brief summary of newly 340 available SSI observations obtained during the last decade is given. Also a brief description of recent developments regarding the absolute values of the TSI, and its variability is provided.

SCIAMACHY and GOME
SCIAMACHY (SCanning Imaging Absorption spectroMeter for Atmospheric CHartographY) and GOME (Global Ozone Monitoring Experiment) onboard the ENVISAT and ERS-2 satellites, respectively, are atmospheric sounders that 350 measure terrestrial atmospheric trace gases (Burrows et al., 1999;Bovensmann et al., 1999;Bovensmann et al., 2011). The primary purpose of direct solar measurements by SCIA-MACHY and the two GOMEs (a second GOME is flying on METOP-A since 2006 and a third GOME on METOP-355 B since 2012) is to Sun-normalise the backscattered light from the terrestrial atmosphere, which is then inverted to determine atmospheric trace gas amounts. This normalisation does not require absolute radiometric calibration and cancels out degradation effects. The SCIAMACHY and 360 GOME instruments have been radiometrically calibrated before launch. In order to provide estimates for solar cycle variability from SCIAMACHY measurements (230 nm-2.4 µm) without the need for a detailed degradation correction, a proxy 365 model (hereafter referred to as SCIAMACHY proxy model) was developed by fitting solar proxy time series to observed SCIAMACHY measurements over several 27-day solar rotation periods (Pagaran et al., 2009). The proxies used in this model are the photometric sunspot index (Balmaceda et al.,370 2009) and the Mg II index  for sunspot darkening and facular brightening. Assuming that the fitting parameters linearly scale from solar rotations to an 11-yr solar cycle, one can then use the solar proxies to extrapolate beyond the lifetime of the single instrument (Pagaran et al.,375 2011b). Note, however, that this assumption might not be accurate, and probably results in an underestimate of the mag-nitude of the solar cycle variation. The short-term variability of the SCIAMACHY proxy model over several solar rotations agrees well with direct solar observations from SCIA-380 MACHY, SORCE/SIM, and UARS/SUSIM (Pagaran et al., 2009(Pagaran et al., , 2011a. Over longer periods during the descending phases of solar cycles 21 to 23, larger differences between model and direct observations become apparent. In particular, SORCE/SIM data show UV changes that are 385 about four times larger than the SCIAMACHY proxy model (Pagaran et al., 2011b;Fig. 2).
Very recently an optical degradation model has been developed that uses the various light paths from different combinations of mirrors within SCIAMACHY. The main causes 390 for the optical degradation are believed to be contaminants on the mirror surfaces. This new degradation model with improved calibration corrections should allow in the near future to derive long-term time series of SCIAMACHY measurements without the use of the SCIAMACHY proxy model.

SORCE
The Solar Radiation and Climate Experiment (SORCE; Rottman, 2005), launched in January 2003, has made continuous daily measurements of SSI from 0.1 to 2400 nm (with missing portions of the extreme UV between 35 nm 400 and 115 nm), accounting for about 97 % of the TSI. These measurements are unique compared to SSI existing data. TSI is also measured on SORCE by the Total Irradiance Monitor (TIM; Kopp et al., 2005). The two instruments onboard SORCE pertinent to this study are the So-405 lar Stellar Irradiance Comparison Experiment (SOLSTICE; McClintock et al., 2005) and the Spectral Irradiance Monitor (SIM; Harder et al., 2005a,b). SORCE/SOLSTICE is a grating spectrometer that measures SSI in the UV from 115 nm to 320 nm with a res-410 olution of 0.1 nm and with an absolute calibration uncertainty of approximately 5 % and measurement precision to better than 0.5 % on all time scales (Snow et al. 2005). SORCE/SOLSTICE is a second generation of UARS/SOLSTICE (Rottman et al., 1993) which acquired 415 UV measurements from 1991 to 2001. SOLSTICE uses nighttime observations of stars to track and correct for changes in responsivity. SORCE/SOLSTICE uses two channels to cover the spectral regions 115-180 nm and 170-320 nm. The long-term stability in the latest data version 420 is about 1 % per year (M. Snow, personal communication, 2012). SORCE/SIM (Harder et al., 2005a,b) was developed to replace the longest wavelength channel (280-420 nm) in the original UARS/SOLSTICE and to extend wavelength cover-425 age well out into the NIR. SORCE/SIM employs a single optical element, a Féry prism, for dispersion and to focus light on four detectors in the focal plane. Two photodiode detectors cover the range from 200 nm to 950 nm, another covers the range from 895 nm to 1620 nm, and an electrical sub-430 stitution radiometer (ESR) operates over the spectral range from 258 nm to 2423 nm. This ESR is also used to calibrate the other three. Because SORCE/SIM is a prism spectrometer its resolution varies from less than 1 nm in UV to approximately 40 nm in the NIR. SORCE/SIM began reporting daily 435 SSI results in April 2004. It achieves an absolute calibration uncertainty of approximately 2 % and measurement precision of 0.1 % or better at most wavelengths. Merkel et al. (2011) stated that the SIM long-term uncertainty in the 200-300 nm region is ≈ 0.5-0.1%, from 310-400 nm it is ≈ 0.2-440 0.05%, and in the 400-1600 nm range it is better than 0.05%. However, they also reported the lack of independent observations for direct validation of these estimates. Moreover, the values above are relative to annual changes. Therefore, they can be assumed valid for short-term variations, but on longer 445 time scales, the instrument stability could be considerably lower.
It is worth mentioning that the SORCE public Level 3 data include the SOLSTICE data up to 308 nm and SIM data above 308 nm. However, the SIM spectra do extend 450 down to about 200 nm. Although the SIM data between 200-308 nm are not publicly available yet, they have been made exploitable for several studies through personal communications. The SORCE data used in our study to estimate the atmospheric response to SSI solar cycle variations are speci-455 fied in Table 3. They also include data received by personal communications with instrument team members (J. Harder, personal communication, 2012). Harder et al. (2009) presented multi-year SORCE/SIM trends indicating that UV variability during the declining 460 phase of solar cycle 23 (between 2004 and 2008) was larger than that observed in previous cycles, and was compensated by trends in other bands in the visible and NIR that increased with decreasing solar activity. SORCE/SOLSTICE has also shown enhanced UV variability for the same time 465 period. Solar UV variability measured by both SORCE/SIM and SORCE/SOLSTICE exceed the variability observed by UARS/SOLSTICE and UARS/SUSIM over solar cycle 22 and ascending phase of cycle 23 by a factor of 3-10 depending on wavelength (DeLand and Cebula, 2012;Figs. 2, 4, 470 and 8).
These discrepancies with prior cycle observations and with SSI models have inspired new analyses and collaborations aimed at a better understanding of the potential sources of instrument degradation that might have affected SORCE in-475 struments and previous instruments as well. The studies have been concentrated on SSI instrument observations, capabilities, and estimated spectral irradiance uncertainties, methods of correcting for degradation, and refinement of estimated uncertainties. It has been understood that all detectors and 480 optics suffer some degradation in space, largely due to exposure to solar light, and also due to hydrocarbon contamination that dominates below 400 nm. Accordingly, new models of degradation based on total dose, rather than just exposure time, are being developed for the SORCE and other instru-485 ments. Revised data sets, which e.g. will include SIM data down to 240 nm, are expected out in 2013. Besides, degradation trends have also been analyzed by considering the expected invariance of SSI over the solar cycle minimum. The latter method has been developed by Woods (2012) and ap-490 plied to data during last solar cycle minimum (2008)(2009) to estimate possible degradation trends for SORCE/SIM and SORCE/SOLSTICE. It consists of identifying near-identical solar activity levels on both sides of the minimum to derive corrections for instrument degradation. This analysis showed 495 good agreement of the variability from moderate solar activity level to minimum level from various measurements and models, from 120 nm to 300 nm for solar cycles 21 through 24. However, as the method has about 30% uncertainty in variability due to the assumptions about selecting times of 500 similar irradiance levels, the results may not be as accurate as those derived from analyses based on instrument degradation alone.
The analysis by Woods (2012) reduces the variability of the integrated UV irradiance from 200 nm to 400 nm, rela-505 tive to the measured TSI change, to 110 % (Fig. 2) from the 190 % change reported by Harder et al. (2009). Nevertheless, to be compatible with the observed TSI changes even this lower amplitude of the UV variation over the solar cycle still requires compensation from out-of-phase trends at other 510 wavelengths, in particular above ≈ 400 nm. Other analyses of solar cycle variability suggests that the UV variability in the 200 nm to 400 nm range is about 60 % of the measured TSI change Pagaran et al., 2009;Morrill et al., 2011b).

The
SOLar SPECtrum instrument (SOLSPEC; Thuillier et al., 2009) is composed of three double monochromators (170-390 nm, 380-850 nm, 800-3000 nm) and a set of lamps allowing corrections for aging related to the 520 harsh space environment. The SOLSPEC spectrometer flew several times on the Space Shuttle and its twin instrument, SOSP (SOlar SPectrum), was placed on the EURECA (EUropean Retrieval Carrier) platform for 10 months (April 1993 to January 1994). These missions have provided data 525 to build the ATLAS (ATmospheric Laboratory for Applications and Science) spectra, specifically ATLAS-1 (March 1992) and ATLAS-3 (November 1994) (Thuillier et al., 2004), which are composites using UARS/SUSIM and UAR/SOLSTICE data from Lyman-α at 121 nm to 530 200 nm, and ATLAS/SSBUV, ATLAS/SUSIM, and AT-LAS/SOLSPEC from 200 to 400 nm, ATLAS/SOLSPEC from 400 to 850 nm, and EURECA/SOSP from 800 to 2400 nm. The ATLAS spectra were calibrated to absolute radiometric scale using the blackbody of the Heidelberg 535 Observatory, and tungsten and deuterium lamps calibrated by NIST. SOLSPEC has been up-graded for operations onboard the International Space Station (ISS) by implementing several changes in the electronics, optics, and mechanisms and by 540 adding redundant components in order to generate data with proper degradation correction. ISS/SOLSPEC has been calibrated to an absolute scale at the Physikalisch-Technische Bundesanstalt (PTB) using the BB3200pg blackbody radiator (Sperfeld et al., 1998). Over the whole spectral range, 545 SOLSPEC accuracy is within 2 to 3 %. The SOLSPEC spectrometer has been in operation since February 2008 onboard the ISS along with the SOL-ACES (SOLar Auto-Calibrating EUV/UV Spectrophotometers) instrument measuring below 150 nm. When the ISS orientation allows, ISS/SOLSPEC 550 records the solar spectral irradiance. Presently, data have been obtained during the solar minimum preceding solar cycle 24 and at some specific periods during its rising phase for direct comparisons with SORCE/SIM observations (Thuillier et al., 2012).

Statistical analysis of SSI time series
As mentioned above, the assessment of long-term variations in SSI observations can be done only by stitching together different records that, individually, often do not last for more than a few years and are almost always offset by different 560 calibration scales. The first systematic effort toward building such a single composite data set for the UV was done by DeLand and Cebula (2008), who created a data set with daily spectra covering the wavelength range 120-400 nm for the time period November 1978 to August 2005. However, the 565 instruments frequently differed in their radiometric calibra- tion and in their long-term stability. One is therefore often left with making subjective adjustments that can dramatically alter the interpretation of long-term variations (Lockwood, 2011) and, for example, mimic a solar cycle variation that 570 does not exist. Normalisation to a single reference spectrum, as done by DeLand and Cebula (2008), can only partially compensate such differences in absolute calibration. A more objective approach consists of incorporating these instrumental discrepancies in the reconstruction, and then 575 explicitly using them as a contribution to the overall uncertainty. A Bayesian statistical framework is ideally suited for this analysis.
To illustrate this approach, we concentrate on the solar cycle (i.e. decadal) modulation, which has the advantage of 580 being one of the conspicuous signatures of solar variability in climate records while offering sensitive means for diagnosing the quality of the solar observations.
Our analysis consists first in extending in time all SSI observations, and subsequently in analysing their 11-yr mod-585 ulation. First, for each instrument and for each wavelength, we extrapolate the SSI observations backward and forward in time while preserving their statistical properties with respect to all other observations (cross-correlation, etc.). This is made possible thanks to the empirical evidence for all neigh-590 boring spectral bands to evolve remarkably coherently in I. Ermolli et al.: Spectral irradiance and climate 9 time, on time scales of hours and beyond Amblard et al., 2008). This coherency, which is rooted in the strong magnetic coupling between solar atmospheric layers, allows us to describe all salient features of the variability in the SSI with just a few degrees of freedom. There are typically three of them in the UV. The extrapolation is based on the expectation-maximisation technique (Dudok de Wit, 2011), which is routinely used to fill in missing data in climate records. As a result, we obtain for each wavelength 600 as many records as there are instruments observing it. The overall dispersion of the various observations is then naturally reflected by the dispersion of the reconstructions, and so no offset or trend adjustments are required. We would like to highlight that no proxies are used to constrain the recon-605 struction.
The next step consists in comparing the 11-yr modulation amplitude and phase for each reconstruction and identifying possible changes since 1978. To date, all spectral bands and all solar proxies show in-phase variations with each other (up 610 to a constant pass shift), regardless of the solar cycle. This is a strong, albeit not sufficient, indication that different spectral bands are likely to remain in phase from cycle to cycle. Likewise, their modulation amplitudes are unlikely to change. Any departure from this picture should thus receive 615 special consideration. We estimate the modulation amplitudes and phases by running an 11-yr running mean through the data. Here, the phase reference is chosen to be the Mg II core-to-wing ratio because it is widely used as proxy for the solar UV (Viereck et al., 2001). This choice, however, has 620 no impact on the results that follow. Figure 4 illustrates the extrapolation for the 220-240 nm band, which is important for ozone production (e.g. Rozanov et al., 2002). This example shows that all observations vary in phase with the solar cycle but differ consider-625 ably in their modulation amplitude. The observations from SORCE exhibit a larger modulation amplitude, as already mentioned above. Note that the use of the MgII index as a reference is just to fix the phase reference for comparing the results obtained for different cycles, without affecting the 630 way the SSI is extrapolated back-and forward in time.
Because this approach provides as many UV composites as there are observations, we can test whether the observations from SORCE/SIM and SORCE/SOLSTICE are compatible with those from other instruments, most of which 635 were obtained during preceding solar cycles. The comparison is summarised in Fig. 5, which compares the modulation amplitude and phase of SORCE versus the distribution obtained by the other instruments. We conclude that all instruments agree remarkably well below 200 nm; at longer UV wavelengths, the modulation amplitude inferred from SORCE/SOLSTICE is systematically larger by a factor of two to six at all wavelengths. The simultaneous sharp drop in its phase raises doubts about the consistency between SOLSTICE and the other measurements, and also between 645 SIM and the other data. Assuming that there is no reason for the SSI to be unusual during the last solar cycle only, we conclude that these observations are likely to be affected by instrumental drifts, in agreement with the conclusions from Lean and DeLand (2012). This may also be the case for 650 SORCE/SIM between 308 and 340 nm, although the departure from other observations is much less significant here.

TSI time series
Some climate models still include solar variability only as derived from TSI, thereby fixing the relative spectral contri-655 bution of the solar radiation hitting the Earth. In addition, TSI time series also provide constraints for the empirical and semi-empirical models of SSI variations that are discussed in the following. Therefore, it seems appropriate to discuss here the current knowledge on TSI measurements. In fact, 660 for many years, the canonical value of the average TSI was 1365.4 ± 1.3 Wm −2 . Now, the most accurate, and generally accepted, value is 1361 ± 0.5 Wm −2 (Kopp and Lean, 2011;. This lower value will be used in the data assimilation and meteorological re-analysis project like 665 ERA-CLIM at ECMWF (D. Dee, ECMWF, personal communication 2012, and see http://www.era-clim.eu/). TSI variability was already predicted in the 1920s from ground-based observations (Abbot et al., 1923). Accurate measurement of TSI and detection of its variability requires 670 observations from space. The first report of the variable solar irradiance with correct amplitudes was made by Hickey et al. (1980). Later observations differed markedly in their absolute values but all basically agreed in the relative amplitude of the TSI variations. In Table 1 we list the main space exper-675 iments that have measured TSI, together with their observed variabilities, which for some experiments were subsequently revised as further detailed below. The given numbers are biased by the duration of the experiments and their phase relative to the solar cycle, but the overall result is that TSI varia-680 tions are observed to be on the order of about 0.5 % standard deviation from the mean value. Lee et al. (1995) estimated the absolute accuracy of the NIMBUS7/HF instrument to be 0.5 % and that of ERBS/ERBE 0.2 %. Willson (1979) expected his 685 SMM/ACRIM-I experiment to remain within 0.1 % for at least a year. Fröhlich and Lean (1998) state that the absolute measurements of the early radiometers are uncertain to about 0.4 %, which corresponds to 5.5 Wm −2 . The main reason for this relatively large spread was due to the uncertainty in 690 the aperture area. When the technology for determining aperture area improved the agreements among the measurements improved. However, the SORCE/TIM experiment proved to be a new outlier. Lawrence et al. (2003) claim an uncertainty of 0.5 Wm −2 , i.e. accurate to 350 ppm. Because 695 SORCE/TIM is 4.5 and 5 Wm −2 below SOHO/VIRGO and ACRIM/ACRIM-III, respectively, the uncertainties given by the instrument teams do not overlap (Kopp and Lean, 2011). The PREMOS experiment on the French satellite PICARD, which was launched in July 2010, has con-700 tributed to the understanding of the instrument offsets, by confirming the lower TSI value initially reported by the SORCE/TIM (Kopp et al., 2005). The radiometers of the PICARD/PREMOS experiment have been calibrated in two different and independent ways. The first is a calibration in 705 power response as reported by Schmutz et al. (2009). In addition, the TSI radiometer of PICARD/PREMOS is, so far, the first space instrument that has been calibrated in irradiance in vacuum. This was done at the Total solar irradiance Radiometer Facility (TRF) located at the Laboratory 710 for Atmospheric and Space Physics (LASP) in Boulder, Colorado, USA (Fehlmann et al., 2012). The irradiance calibration is accurate to 330 ppm. PICARD/PREMOS agrees with SORCE/TIM to within 0.4 Wm −2 , with PICARD/PREMOS being lower . Thus, the new experi-715 ments are well within their common uncertainty range and the uncertainty difference between independent measurements of the TSI has decreased by a factor of ten.
The new experiments and the work carried out by the international community, which included realization of new 720 facilities, ground-based tests, and collaborations aimed at identifying, quantifying, and verifying the causes of the discrepancy between the TIM and older TSI instruments, have ultimately led to the understanding of the instrument offsets. In particular, the characterisation of the SORCE/TIM, 725 ACRIM/ACRIM III, and SOHO/VIRGO witness units at TRF, and the calibration of PICARD/PREMOS, led to better characterise instrumental effects and to resolve the source of the discrepancy among TSI observations (Kopp and Lean, 2011;. Instruments such as PMO6 730 and ACRIM type having a view-limiting aperture in front and a smaller precision aperture that defines the irradiance area have a large amount of scattered light within the instrument. This additional light is not fully absorbed by the baffle system and produces scattered light contributing to ex-735 tra power measured by the cavity. Scattered light was one of the potential systematic errors suspected by Butler et al. (2008). Subsequent ground testing involving the different instrument teams verified scattering as the primary cause of the discrepancy between the TIM measurements and the er-740 roneously high values of other TSI instruments. Accordingly, new stray light corrections have recently been assigned to ACRIM/ACRIM-III (based on spare instruments) and its stray light contribution is indeed of the order as the observed differences. For SOHO/VIRGO the scattered light issue was 745 not the reason for its discrepant reading. VIRGO is traceable to the World Radiometric Reference (WRR). New analyses, however, have revealed that the WRR has a systematic offset (Fehlmann et al., 2012). In particular, the WRR offset produced approximately the same systematic shift as the 750 scattering error. Thus, scattered light could be the reason of discrepant readings in both VIRGO and WRR instruments.
It is worth emphasizing that the key issue in irradiance measurements is the traceability of the instrument, which is necessary to meet metrological requirements. The 755 various instruments have undergone increasingly precise post-launch corrections to meet such requirements; PI-CARD/PREMOS was the first to achieve complete traceability. However, there is now a consistent evaluation of the TSI from four instruments: SORCE/TIM, SOHO/VIRGO 760 (corrected), ACRIM/ACRIM-III (corrected), and from PI-CARD/PREMOS. At the same time, TSI changes can be measured to much higher precision than absolute TSI values. On short time scales, relative measurements are accurate to a few ppm on a daily average. On longer time scales of 765 years and tens of years, the stability of the measurements are much more difficult to evaluate. Claims of stabilities of less than 100 ppm over ten years (Fröhlich, 2009) are most likely too optimistic. A more realistic estimate comes from comparing independent composites that have been constructed.

770
Over the time of the last solar cycle these agree to within about 0.2 Wm −2 or about 20 ppmyr −1 . For pre-1996 measurements even higher uncertainties for the systematic drifts have to be adopted. Despite these conservative assessments, the TSI time record is a factor of ten more accurate than any 775 SSI observation.

Discussion of SSI observations
While the analysis of possible degradation trends for SORCE, SCIAMACHY, and other missions may help to resolve some of the differences in existing long-term time-series of measurements, to advance our understanding of SSI variability new and improved, upgraded observations are required in the coming years. To this purpose, the next generation SIM instrument built for the NOAA/NASA Total Solar Irradiance Sensor (TSIS) mission (Cahalan et al., 2012) 785 includes many design improvements for reducing noise and improving in-flight degradation tracking. The TSIS/SIM instrument is currently undergoing laboratory calibrations using radiometric technology similar to that employed to resolve the source of offsets among various TSI instruments 790 (Kopp and Lean, 2011).
The TSIS mission might be launched in 2016. It is highly unlikely that TSIS/SIM and SORCE/SIM observations will overlap in time due to the expected lifetime of the SORCE batteries. Nevertheless, even without any overlap in time, 795 the next generation measurements will help to better understand the performance of previous instruments and SSI solar cycle variation. In addition, several satellite missions dedicated to the observation of the Earth's atmosphere (GOME-2, SBUV/2, OMPS) will also provide information on part of the 800 SSI spectra during the possible gap in observations from the instruments specifically designed to SSI measurements.

Models of SSI variations
Longer, uninterrupted, more stable and reliable observational time series are critical for understanding the origin of the 805 differences between SSI measurements and improving our knowledge of SSI variations. However, the physics of the underlying processes also needs to be understood better, in order to facilitate the construction of more realistic models of SSI variations. Such models are particularly crucial since cli-810 mate studies including stratospheric chemistry urgently need long and reliable SSI data sets for realistic simulations. Acquiring sufficiently long SSI time series is a long process and making them more reliable requires flying multiple new instruments, which will not happen for some time. Even when 815 such time series do become available, they can only be extended into the past or future with the help of suitable models.
Although considerable progress has been made in modeling the TSI variability (e.g. Foukal and Lean, 820 1990;Chapman et al., 1996;Fröhlich and Lean, 1997;Preminger et al., 2002;Ermolli et al., 2003;Krivova et al., 2003;Wenzler et al., 2004Wenzler et al., , 2005Wenzler et al., , 2006Crouch et al., 2008;Bolduc et al., 2012;Ball et al., 2012), modelling the SSI is more difficult, leaving considerable 825 room for improvement. Here we describe recent progress in SSI modelling. We discuss the mechanisms responsible for the irradiance variations (Sect. 3.1), with special emphasis on the possible differences in the spectral response of the irradiance to the various solar magnetic features. We then 830 describe the basic principles and key components of the models and review five current SSI models (Sect. 3.2) that are available for climate studies and that were employed for our estimate of the atmospheric response to SSI solar cycle variations. Finally, we compare the models to each 835 other, confront them with the available observational data presented in the previous section and discuss remaining uncertainties and open issues (Sect. 3.3).

Mechanisms of irradiance variations
Although various mechanisms have been proposed to explain 840 the variation of solar irradiance, it is now accepted that observed variations in TSI (i.e. over the last 3.5 solar cycles) are predominantly caused by magnetic features on the solar surface. We cannot rule out that on longer time scales other mechanisms play a significant role, but this is beyond the 845 scope of this paper.
Empirically it has been known for a long time that magnetic features on the solar surface are generally either dark (sunspots, pores) or bright (magnetic elements forming faculae and the network) when averaged over the solar disk. Two 850 questions arise from this observation: why are some flux tubes (the theoretical concept used to describe faculae and sunspots) bright, while others are dark? What happens to the energy flux blocked by sunspots (or equivalently, where does the excess energy emitted by faculae come from)?

855
The strong magnetic field within both small and large magnetic flux tubes reduces the convective energy flux. The vertical radiative energy flux in the convection zone is comparatively small and cannot compensate a reduction in convective flux. This leads to a cooling of magnetic features.

860
The magnetic features are evacuated due to the large internal magnetic pressure and horizontal balance of total (i.e. gas plus magnetic) pressure. Hence, these evacuated magnetic structures are also heated by radiation flowing in from their dense and generally hot walls. This radiation effi-865 ciently heats features narrower than roughly 250 km, making them brighter than the mainly field-free part of the photosphere, especially when seen near the limb where the bright walls are best visible (Spruit, 1976;Keller et al., 2004). For larger features, the radiation does not penetrate most 870 of their volume (the horizontal photon mean free path is roughly 50-100 km), so that features greater than roughly 400 km in diameter remain dark (pores and sunspots); cf. Grossmann-Doerth et al. (1994).
What happens with the energy that gets blocked by 875 sunspots? According to Spruit (1982), this energy gets redistributed throughout the convection zone and is re-emitted again slowly over its Kelvin-Helmholtz timescale, which exceeds the lifetime of sunspots by orders of magnitude. Similarly, the excess radiation coming from small flux tubes 880 (which act as leaks in the solar surface, since these evacuated features increase the solar surface area from which radiation can escape) is also taken from the heat stored inside the entire convection zone.
length. It is in the nature of integrals that quite different functions of wavelength, i.e. different SSI variations, can lead to the same variation in TSI. The differences in the relative change in irradiance at various wavelengths is given by three effects: (1) The relative sensitivity of the Planck func-890 tion to temperature increases rapidly with decreasing wavelength.
(2) Radiation at the various wavelengths is emitted at different heights in the solar atmosphere. This influences SSI because the contrast of magnetic features relative to their non-magnetic surroundings is height-and hence wavelength-895 dependent.
(3) At very short and at very long wavelengths (EUV and radio), the radiation comes from the upper transition region and corona, where the brightest commonly found sources are complete loops rather than just the loop footpoints (the flux tubes) as at almost all wavelengths in be-900 tween.
Point 3 refers to wavelengths we do not consider here because they interact mainly with the uppermost regions of the Earth's atmosphere (mesosphere and above) and hardly contribute to TSI at all. Point 2, however, is important, since 905 in general the temperature in magnetic features drops more slowly with height in the solar photosphere and increases much more rapidly with height in the chromosphere than the solar average. This means that radiation emitted at higher levels in the solar atmosphere, i.e. in the UV and in the cores 910 of spectral lines, displays larger changes. Together, points 1 and 2 favour the UV to exhibit larger variations than in the visible and NIR. In addition to the fact that the continuum radiation comes from greater heights and the Planck function shows a greater temperature sensitivity, the density of spec-915 tral lines per wavelength interval also increases very rapidly towards shorter wavelengths. Now, can magnetic features be dark in the visible, but bright in the UV? This is in principle possible, if the magnetic feature is cool in the deep atmosphere, but hot in the 920 upper photosphere. For instance, pores qualitatively show such a temperature profile, although there are no simultaneous high-resolution observations in the visible and the UV to decide if the temperature gradient is sufficiently extreme to produce such an effect. But pores are relatively short lived 925 and too few in number. Hence, any long-lasting global dimming in the visible at times of high activity, such as that shown by the SORCE/SIM measurements, can only be produced by the small magnetic elements, specifically those in the network (these are much more numerous and more evenly 930 distributed). However, such small-scale magnetic elements are unlikely candidates to produce a decrease in the visible irradiance along with increased UV irradiance. Firstly, magnetic elements in the network are bright even in the continuum and at disk centre (e.g. Kobel et al., 2011). Secondly, 935 Röhrbein et al. (2011) have shown that the darker than average appearance of some magnetic elements is largely due to spatial smearing of the observations (although there may be some darkening due to the inhibition of convection around  (Shapiro et al., 2010, purple) and SATIRE (Unruh et al., 1999, blue). Note that contrasts depend on the position on the disc. Shown are averages over the entire disc under the assumption of a homogeneous spatial distribution of features. magnetic features). Thirdly, there are also many spectral 940 lines in the visible, which brighten significantly in magnetic elements and counteract any darkening in the continuum. Finally and most importantly, magnetic elements near the limb are always rather bright, so that averaged over the solar disk small-scale magnetic features are expected to lead 945 to a brightening. All this implies that the measurements by SORCE/SIM of a change in the irradiance at visible wavelengths out-of-phase with TSI (Harder et al., 2009) are not compatible with the surface magnetic field as the source of the SSI variations in the visible. However, the SSI variations 950 in the visible are an important contributor to TSI variations (Fig. 2) and the above incompatibility would be strongly contradicted by the result of Ball et al. (2012) that 92 % of TSI variations are reproduced by the evolution of the magnetic field at the solar surface.

Solar irradiance models
It was noticed soon after the beginning of routine monitoring of TSI from space, that changes in TSI were closely related to the evolution of different brightness structures on the visible solar disc (Foukal and Vernazza, 1979;Willson et al., 1981;Oster et al., 1982;Eddy et al., 1982;Foukal and Lean, 1986). These brightness structures (such as sunspots, pores, faculae, plage and network) are manifestations of the solar magnetic field emerging at the Sun's surface. Thus their evolution in a global sense (and without looking too closely 965 at the details of the temporal evolution) can be relatively well represented by different, typically disc-integrated, proxies of solar magnetic activity, such as the sunspot number or area, plage area, the solar radio flux at 10.7 cm (f10.7), the Mg II core-to-wing index, or the Ca II K line. This has widely been used in the oldest (but still widely deployed) irradiance models (e.g, Donnelly et al., 1982;Foukal and Lean, 1990;Chapman et al., 1996Chapman et al., , 2012Fröhlich and Lean, 1997;Fligge et al., 1998;Lean, 2000;Preminger et al., 2002;Jain and Hasan, 2004;Pagaran et al., 975 2009). In these models, the measured irradiance variations are fitted via a set of activity proxies through multiple regressions.
The success and the limitations of the regression methods in accounting for measured TSI variations (on time-scales of 980 days to years) gave a strong impetus to the development of more sophisticated and physics-based models. Such models consider contributions of different brightness structures to the irradiance change separately. Thus the solar energy output is the sum of the fluxes emerging from all the fea-985 tures observed on the solar visible surface (corresponding to the solar photosphere); the number and type of disk features accounted for depends on the model. Usually these models require two prime ingredients: (1) the surface area covered by each photospheric component as a function of time, and

990
(2) the brightness of each component as a function of wavelength and often also of the position on the solar disc.

1005
(2) The brightness of individual photospheric components is calculated from semi-empirical models of the solar atmospheric structure using various radiative transfer codes, such as SRPM (Fontenla et al., 2011, and references therein), ATLAS9 (Kurucz, 1993) in SATIRE (Unruh et al., 1999;1010 Fligge andKrivova et al., 2003) or COSI (Haberreiter et al., 2008;Shapiro et al., 2010). Such computations are significantly complicated by the departures from the local thermodynamical equilibrium (LTE) conditions (Sect. 2.6 of Rutten, 2003) in the solar atmosphere, as well 1015 as by the temporal and spatial bifurcations of the temperature and density in the solar atmosphere (Carlsson and Stein, 1997;Uitenbroek and Criscuoli, 2011). High resolution 3-D models (e.g. Socas-Navarro, 2011) and 3-D MHD simulations (e.g. Vögler, 2005) of the solar atmosphere are grad-1020 ually becoming available. However, current models cannot yet reproduce available observations over the entire spectrum (Afram et al., 2011). Thus, at present semi-empirical 1-D models of the solar atmosphere are a de-facto standard for calculating irradiance variations. Although these models 1025 do not account for the spatial structure and temporal variability of the solar atmosphere, they can be adjusted to calculate the solar spectrum and its variability with high accuracy (Unruh et al., 1999;Fontenla et al., 2011, and references therein). These models do not necessarily catch the 1030 average properties of the inhomogeneous solar atmosphere (Loukitcheva et al., 2004;Uitenbroek and Criscuoli, 2011), but, if validated and constrained by the available measurements, they can be considered as a reliable and convenient semi-empirical tool for modelling SSI variability.
The spectra of the quiet Sun calculated with three different models are shown in Fig. 6 (top panel). All plotted spectra are in reasonable agreement with each other, though models used in COSI and SATIRE are a bit closer to each other than 1045 to SRPM. The bottom panel of Fig. 6 shows the flux differences (i.e. the contrasts) between bright active components and the quiet Sun calculated with different models. The calculations were done assuming that the active regions cover the entire solar disk. To calculate the variability, the con-1050 trasts have to be weighted by the filling factors (surface area coverage), which are specific for every model. Thus, different magnitudes of the contrasts do not necessarily imply that the models produce different variability. At the same time, the spectral profile of the contrasts determines the depen-1055 dence of the variability on the wavelength. Figure 6 shows that the (disc−integrated) spectral contrasts of facular, plage, and network regions calculated with SATIRE and COSI are very similar and positive over the whole spectral range. On the other hand, the SRPM contrasts are much lower in the 1060 visible compared to other models and are negative at some wavelengths. This is not very surprising, since the SRPM contrasts shown in Fig. 6 Fig. 7. Ratio of SSI variability to TSI variability between 200 nm and 1800 nm in bins of 20 nm below 400 nm, 40 nm in the range 400-800 nm, and 100 nm in the IR above 800 nm. Shown are SORCE/SOLSTICE (orange) and SORCE/SIM (red) measurements between 2004 and 2008, as well as the NRLSSI (grey), SATIRE-S (blue) and COSI (purple) models between the maximum and minimum of cycle 23. subsequently revised to account for some observations of the 1065 solar atmosphere, as further detailed below.
The advantage of employing radiances computed from the semi-empirical model atmospheres is that they allow computations of solar irradiance at different wavelength (i.e. spectral irradiance). This is not straightforward with the regression models, since in this case the regression coefficients need to be estimated from observations or alternatively, irradiance changes at individual wavelengths need to be somehow scaled from the TSI changes or changes in the irradiance at some other (known) wavelength. This latter technique is 1075 also partly used in the UV by the NRLSSI and SATIRE models that are described below.

NRLSSI
The Naval Research Laboratory Solar Spectral Irradiance 2 (NRLSSI; Lean et al., 1997;Lean, 2000) uses the photo-1080 spheric sunspot index derived from sunspot area records to describe the evolution of sunspots in time, and Mg II, CaII and f10.7 disk-integrated indices to represent facular brightening.
Below 400 nm, the spectral irradiances are derived from 1085 UARS/SOLSTICE observations through a multiple regression analysis with respect to a (SOLSTICE) reference spectrum. The regression analysis includes a facular brightening and a sunspot blocking component (see Lean et al., 1997;Lean, 2000 for more detail). It is known that the long-term 1090 stability of the UARS spectral instruments (both SOLSTICE and SUSIM) was not sufficient to trace solar cycle variability above roughly 220 nm (Woods et al., 1996). Therefore the coefficients are derived from the rotational variability so as to avoid any long-term instrumental effects. This approach thus 1095 assumes that spectral irradiance changes show the same linear scaling with a given proxy on rotational as well as cyclical time scales. Above 400 nm, the facular and sunspot contrasts are largely based on the contrasts presented in 1100 Solanki and Unruh (1998). They have been scaled to ensure that the overall (wavelength-integrated) solar cycle change due to sunspots (viz faculae) agrees with the bolometric value for the sunspot blocking (viz facular brightening) derived from TSI modelling (see, e.g. Fröhlich and Lean, 1105 2004).
The quiet-Sun spectrum in NRLSSI is effectively a composite of UARS/SOLSTICE observations (below 400 nm), SOLSPEC (Thuillier et al., 1998) and the model by Kurucz (1991). This composite has been scaled so that its integrated 1110 flux corresponds to a TSI value of 1365.5 Wm −2 .
Compared to the SATIRE and COSI models (described below), the NRLSSI model shows lower variations on solarcycle and longer time scales between 250 and 400 nm (Figs. 2 and 7). This is mainly because the regression coef-1115 ficients are derived from rotational variability only. Below 250 nm (Figs. 7 and 8) and in the visible at 400-700 nm (Figs. 2 and 7) the NRLSSI, COSI, and SATIRE models agree well with each other.
In the visible and IR range, the facular contrasts in the 1120 NRLSSI model rely on the facular model atmosphere by Solanki and Unruh (1998), which is an earlier version of the model by Unruh et al. (1999) currently used in SATIRE.
Thus although the NRLSSI model does show a weak outof-phase variability in the IR around 1600 nm (Fig. 7), inte-1125 grated over the range 1000-2430 nm (Fig. 2) the modelled irradiance is in phase with the solar cycle. The SATIRE, COSI and OAR models all display reversed variability in this range. Such a reversed behaviour is also suggested by the SIM variability is much stronger than in the models. Integrated over all wavelengths, the NRLSSI irradiance (i.e. the TSI) is in good agreement with the measurements, although it does not quite reproduce the comparatively low TSI level during the last minimum in 2009 (Kopp and Lean, 1135 2011), as indicated by the PMOD composite and also independently found by Ball et al. (2012) with the SATIRE model. This is because the Mg II index employed to describe the evolution of the bright component with time does not follow exactly the shape of the TSI variation over the last cycle 1140 (although the differences appear to be within the long-term measurement uncertainties; M. Snow, personal communication, 2012).

1145
for the Satellite era) belongs to the class of models using semi-empirical model atmospheres to calculate brightnesses of different surface features and full-disc solar images to describe the surface area coverage by these components at a given time. The full-disc data used in SATIRE are mag-  Kurucz (1991) with an effective temperature of 5777 K, similar but cooler model atmospheres for umbra and penumbra, and model P of Fontenla et al. (1999) slightly modified by Unruh et al. (1999) to achieve better agreement with obser-1160 vations in the visible and near-UV. Since the ATLAS9 code uses the LTE approximation, which is known to fail in the UV, fluxes below 270 nm are re-scaled using UARS/SUSIM observations .
The SATIRE-S modelled variability was found to be 1165 in very good agreement, on both rotational and cyclic time scales, with the PMOD TSI composite Ball et al., 2012), SORCE/TIM TSI measurements (Ball et al., 2011(Ball et al., , 2012, UARS/SUSIM spectral irradiance Unruh et al., 2012), as well 1170 as with UARS/SUSIM and UARS/SOLSTICE Ly-α measurements ( Figs. 2 and 8). On rotational times scales, good agreement is also found between the SSI provided by SATIRE-S and the SORCE/SIM and SORCE/SOLSTICE measurements (Unruh et al., 2008(Unruh et al., , 2012Ball et al., 2011Ball et al., , 1175 see also bottom panel of Fig. 8) and between SATIRE-S and SOHO/VIRGO observations in three spectral channels in the near-UV, visible and near-IR (Krivova et al., 2003

SRPM
The Solar Radiation Physical Modelling 4 (SRPM) is a set of tools used to construct semi-empirical models of the solar atmosphere and to derive solar radiance (or emitted intensity) 1200 and high-resolution irradiance spectra of the Sun. SRPM covers all levels of the solar atmosphere from the photosphere to the corona and takes, for most species, NLTE conditions into account. The calculations include a large number of atomic levels and lines, as well as molecular lines and 1205 molecular photo-dissociation opacities. Seven components have been used in SRPM until recently corresponding to different features that can be identified on the Sun at a mediumresolution of ≈ 2 arcsec: quiet-Sun inter-network, quiet-Sun network lane, enhanced network, plage (that is not facula), 1210 facula (i.e. very bright plage), sunspot umbra and sunspot penumbra. The atmospheric models adopted for the calculations of each of these features are presented in Fontenla et al. (2009). These models go back to the ones of Fontenla et al. (1999, and references therein), which have been revised in 1215 subsequent studies of Fontenla et al. (2004Fontenla et al. ( , 2006Fontenla et al. ( , 2009 to account for observational results and for extension of the atmospheric layers. The observational results include e.g. the negative contrast measurements on facular regions at some wavelengths Topka et al., 1992; 1220 Wang et al., 1998)). Recently, further modifications were introduced into the plage models to take SORCE data into account (see below). In addition, two new components and corresponding model atmospheres, namely dark quiet-Sun internetwork and hot facula, were added but not yet integrated decrease in irradiance at wavelengths below ≈400 nm during the declining phase of solar activity from 2002 to 2009, and a flux increase at longer wavelengths, similarly to the behaviour seen in the SORCE/SIM observations (Harder et al., 2009). The UV variability derived from the model is, 1245 however, significantly weaker than what is measured by SORCE/SOLSTICE. Moreover, the observed changes of TSI on time scales longer than the solar rotation are not captured by the model. Fontenla et al. (2011) argue that a possible reason for this mismatch could be that the current set of the 1250 employed atmospheric models is still missing some components.

COSI
The COde for Solar Irradiance 5 (COSI) calculates synthetic solar spectra for different components of the solar atmo-1255 sphere. COSI returns the most important UV lines including the degree of ionisation of the elements under NLTE. The NLTE Opacity Distribution Functions (Hubeny and Lanz, 1995), implemented in COSI by Haberreiter et al. (2008), indirectly account for the NLTE effects in several millions of 1260 lines. Shapiro et al. (2010) showed that NLTE effects influence the concentration of the negative ion of hydrogen which results in an approximately 10 % change of the continuum level in the visible spectrum. The contrasts between active regions and the quiet Sun are also affected by the NLTE ef-1265 fects. This, however, does not imply that only NLTE codes are capable of reproducing the visible solar spectrum as the NLTE effects can be imitated by a slight readjustment of the atmosphere structure (see, e.g. Rutten and Kostik, 1982;Shchukina and Trujillo Bueno, 2001). The spectrum of the 1270 quiet Sun calculated with COSI is in good agreement with the solar spectrum measured by SOLSTICE (up to 320 nm) and SIM (above 320 nm) onboard the SORCE satellite during the 2008 solar minimum, and with the SOLSPEC measurements during the ATLAS 3 mission in 1994 (Thuillier et al., 2011). 1275 Shapiro et al. (2012a) showed that COSI can accurately reproduce the center-to-limb variation of solar brightness in the Herzberg continuum retrieved from the analysis of solar eclipses observed by LYRA/PROBA2. Solar spectral irradiance variations can be modelled from 1280 COSI spectra by weighting them with filling factors of surface magnetic features (e.g. derived from the magnetograms as in the SATIRE model, cf. Haberreiter et al. (2005), or from the PSPT images as in SRPM). In the current version of the model, this has been done employing sunspot and 10 Be data (Shapiro et al., 2011) and and PSPT images (Shapiro et al., 2012c). The variability so derived with COSI agrees well with SATIRE-S and NRLSSI in the Herzberg continuum spectral range (Figs. 7 and 8) and in the visible (Figs. 2 and 7).
In the near-IR, at 700-1000 nm, COSI shows a weak inverse solar cycle variability (Figs. 2 and 7). This is probably an artifact of the model, as Shapiro et al. (2010) calculated active-region contrasts using the model atmospheres of Fontenla et al. (1999), which do not distinguish between umbra and penumbra. Besides, they employed a single model for both bright plage and plage. These assumptions affect the position of the inversion point; the point where the influence of sunspots starts to outweigh the bright components. To account for this, Shapiro et al. (2012c) decreased plage 1300 and sunspot contrasts with respect to the quiet Sun. A more accurate approach is under development.
Images taken between September 1997 and January 2012 have been processed and segmented to identify different features on the solar disc as described by Ermolli et al. (2010).

1315
The seven classes of atmospheric features proposed by Fontenla et al. (2009) are considered. The brightness spectrum of each feature is computed using the semi-empirical atmospheric models of Fontenla et al. (2009) through the RH synthesis code (Uitenbroek, 2002). Thus the set of compo-1320 nents in the OAR model is essentially the same as used by Fontenla et al. (2011), though their atmospheric structure is based on Fontenla et al. (2009), i.e. before the most recent modifications applied by Fontenla et al. (2011). The

Discussion of SSI model results
There has been steady progress in modelling SSI variations, and a number of models are now available that can be used as input for climate studies. However, the main uncertainty in the models concerns the wavelength range 220-400 nm, 1340 which is of particular interest for climate studies.
Of the five models discussed in this section, specifically NRLSSI, SATIRE-S, COSI, SRPM and OAR, only one (SRPM) shows a behaviour of the UV and visible irradiance qualitatively resembling that of the recent SORCE/SIM mea-1345 surements. However, it should be noted that the integral of the SSI computed with this model over the entire spectral range (i.e. the TSI) does not reproduce the measured cyclical TSI changes. None of the other four models, which are in closer agreement with each other, reproduces the behaviour 1350 of the UV and visible irradiance observed by SORCE/SIM (Figs. 2, 7 and 8). These models are nevertheless all in good or fair agreement with earlier UARS UV observations, TSI measurements and the model based on SCIAMACHY data (e.g. Krivova et al., 2006Krivova et al., , 2009Pagaran et al., 2009;1355Morrill et al., 2011aShapiro et al., 2011;Unruh et al., 2012;Lean and DeLand, 2012;Ball et al., 2012). These models also agree with SORCE data on rotational times scales (e.g. Unruh et al., 2008;Ball et al., 2011;Lean and DeLand, 2012, see also bottom panel of Fig. 8). Note that good 1360 agreement on rotational time scales was also found by DeLand and Cebula (2012) between SORCE and other spectral observations. When comparing various models and data to each other, however, one important issue often remains unacknowl-1365 edged, namely the true uncertainties in the measurements, in particular on time scales of years and longer. As an example, we compare measured and modelled variability at 220-240 nm (Fig. 8). Over the period 2004-2008, the three models shown in this figure (NRLSSI, SATIRE-S and COSI) sug-1370 gest a decrease in the integrated 220-240 nm flux of about 1 %. The fluxes measured by the SORCE/SOLSTICE and SORCE/SIM instruments in this range decreased over the same period by roughly 4 % and 7 %, respectively. In the 220-240 nm region, the accuracy of SORCE/SOLSTICE is 1375 considered to be better than that of SIM . Currently, SOLSTICE long-term stability in this spectral range is estimated to be about 1 %yr −1 (Snow et al., 2005). Thus the observed trend for SORCE/SOLSTICE (4 % over 5 yr) and its difference with the models (3 % over 5 yr), in 1380 fact, lies within the long-term instrumental uncertainty. The difference between the trends for SORCE/SIM and the models (about 6 %) is just outside the value of 5 % over 5 yr, by assuming that the SIM long-term stability is comparable to that of SOLSTICE. Although the SIM uncertainty 1385 stated in Merkel et al. (2011) is lower, this assumption seems to be a reasonable working hypothesis for our study, because it is known that at 220-240 SIM has poorer precision than SORCE/SOLSTICE (J. Harder, personal communica-tion, 2012). It is worth mentioning that new assessments of the SORCE instrument degradation might imply a revision of results derived from several studies, including this one, which however is mainly intended to discuss the data and models available to date for the evaluation of the atmospheric response to SSI variations.
Despite showing much lower variability than SORCE/SIM in the UV, the models also display considerable differences between each other in the range 250-400 nm (up to a factor of three, e.g. between NRLSSI and COSI; Figs. 2 and 7). As a consequence of the low response of UARS/SOLSTICE 1400 to long-term variability above 300 nm and the use of rotational variability to estimate the regression coefficients, it is likely that NRLSSI underestimates the changes in this range (J. Lean, personal communication, 2012), and can thus be considered as the lower limit. All other models rely on semi-1405 empirical model atmospheres in this range, which also need further tests at all wavelengths. The differences between the models in the IR are mainly due to the lack of reliable observations of contrasts of different (solar) atmospheric components and related uncertainties in the corresponding model at-1410 mospheres. In the near-IR (700-1000 nm), all models qualitatively agree, except COSI. However, as discussed earlier, the inverse variability shown by COSI in this range is believed to be an artifact of the atmospheric components and models adopted for the irradiance calculations.

1415
The variability of the spectral irradiance on time scales longer than the solar cycle is beyond the scope of this paper. We note, however, that uncertainties in the longterm SSI reconstructions are essentially the same as the ones discussed in this section convolved with the uncer-1420 tainty in the magnitude of the secular change in irradiance (see, e.g. Krivova and Solanki, 2013;Schmidt et al., 2012;Solanki et al., 2013, and references therein).

Climate impact of SSI measurements
The SSI data derived from recent solar observations and 1425 model calculations presented in the previous sections are employed here to discuss the impact of SSI variations on the Earth's atmosphere and climate. In particular, we present the motivation for simulations of the response of the Earth's atmosphere to variations induced by SSI changes (Sect. 4.1). 1430 We then discuss the solar induced differences in atmospheric heating rates, ozone variability, temperature, and atmospheric circulation (Sect. 4.2). These differences are investigated in recent simulations using atmospheric models with standard solar forcing from the NRLSSI model

Climate modelling
Today, the most advanced tools available for climate simulations are 3-dimensional General Circulation Models (GCMs) 1450 that numerically simulate the general circulation of the atmosphere and/or the ocean based on well-established physical principles. Most climate models that are utilised for future climate predictions in the 4th IPCC report (Solomon et al., 2007) are coupled atmosphere-ocean models reaching up to 1455 the middle stratosphere (32 km), whereas CCMs that are used for future predictions of the stratospheric ozone layer in the WMO report (WMO, 2011) include interactive stratospheric chemistry and reach up to the lower mesosphere or above (80 km). The main external driving force for all cli-1460 mate models is the incoming solar flux at the top of the atmosphere (TOA). Currently there are two mechanisms for solar radiation influence on climate, the so-called "top-down" UV effect (Kodera and Kuroda, 2002) and the "bottom-up" TSI effect (van Loon et al., 2007).

1465
As discussed in the previous sections, even though TSI varies only by about 0.1 % over the solar cycle, larger variations of several percent occur in the UV part of the spectrum, including the ozone absorption bands between 200 and 400 nm that are responsible for the SW heating of the strato-1470 sphere and are important for photochemical processes (e.g. Haigh, 1994). The cyclic variation of the incoming solar irradiance at short wavelengths leads to statistically significant ozone, temperature and zonal wind solar signals in the stratosphere (Austin et al., 2008;Gray et al., 2010). These 1475 solar induced circulation changes in the stratosphere can induce noticeable decadal climate changes in the lower atmosphere and at the surface (e.g. Haigh, 1999;Kodera, 2002;Matthes et al., 2006;Ineson et al., 2011;Matthes, 2011). In order to account for this so-called "top-down" stratospheric 1480 UV mechanism, the radiation code in climate models has to account for spectrally resolved irradiance changes. The first climate models focused on tropospheric climate, thus solar solar forcing was represented by TSI only. As stratospheric changes played a minor role for climate predictions 1485 in the past, climate models did not take into account stratospheric processes and in particular ozone changes due to solar UV absorption. Hence, most of the SW radiation codes developed for use in GCMs did not consider solar irradiance for wavelengths shorter than 250 nm and employed pa-1490 rameterisation using TSI as input. Solar fluxes and heating rates were subsequently calculated in one or two SW absorption bands from the top of the atmosphere to the surface. In contrast, middle atmosphere models use SW radia-tion codes specifically designed for simulations of the upper atmosphere. They, therefore, cover a broader spectral range and include more than two spectral bands in the UV and visible.
Studies on the performance of SW radiation codes with different spectral resolution showed that the observed solar temperature signal in the stratosphere can only be reproduced in models that allow for the effects of spectral variations between solar minimum and maximum (Egorova et al., 2004;Nissen et al., 2007). In a recent paper, Forster et al. (2011) examined in detail the sensitivity of a number of CCM 1505 SW radiation codes to changes in solar irradiance and ozone as well as the ability of the models to reproduce the 11yr radiative solar signal using the NRLSSI data. In their study (SPARC-CCMVal, 2010), the strongest solar temperature signal was found to be in the tropical upper strato-1510 sphere/lower mesosphere, indicating that the direct mechanism of heating by absorption of enhanced UV radiation at solar maximum is well captured by the models that employ spectrally resolved SW radiation schemes. Models that do not account for SSI variations and only consider changes in 1515 spectrally integrated TSI cannot properly simulate solar induced variations in stratospheric temperature (Forster et al., 2011).
Today stratospheric processes are gaining a lot of interest due to their importance for climate. Not only the ef-1520 fect of ozone recovery and its relationship to climate but also stratosphere-troposphere dynamical coupling and its role for predictability from days to decades have been recognised as important issues for future climate studies (e.g. Baldwin et al., 2007;Gerber et al., 2010). Therefore, a better 1525 representation of the stratosphere including improved representation of SW heating processes as well as dynamical coupling with the troposphere in global climate models is critically important.

1530
The uncertainty of SSI variations in recent observations and models has significant influence on simulations of the climate system, since the response of the atmosphere strongly depends on the spectral distribution of the solar irradiance. The effects for middle atmosphere heating, ozone chemistry 1535 and middle atmospheric temperatures are examined in the following.

Effects on atmospheric heating and ozone chemistry
As solar radiation is the primary source of energy that drives 1540 atmospheric as well as oceanic circulation, accurate representation of solar irradiance is of paramount importance for the simulation of the atmospheric temperature, composition and dynamics in climate models. The variability of the solar spectrum in time, and in particular over the solar cycle is necessary for the assessment of solar influence on climate. The amplitude of the simulated solar signal depends on the spectral solar fluxes prescribed at the TOA. Differences in the TOA solar irradiance spectrum result in large changes in the heating rates calculated by SW radiation schemes or radiative 1550 transfer models, as has been shown by Zhong et al. (2008) with differences of up to ≈ 1.1 Kday −1 in mid-latitude summer. Recently, Oberländer et al. (2012) examined the impact of a number of different estimates of prescribed TOA solar fluxes on the solar response in a GCM which includes a radia-1555 tion scheme with enhanced spectral resolution (Nissen et al., 2007) and is therefore able to accurately represent the solar signal induced changes. They used the NRLSSI, the SATIRE and the SCIAMACHY solar flux input data sets, and compared their effects on SW heating rates over the 11-yr so-1560 lar cycle using offline calculations. They also calculated the corresponding temperature response from perpetual January GCM simulations with prescribed ozone concentrations. The comparison revealed clear differences in SW heating rates for the solar minimum of cycle 22 (September 1986).

1565
The simulations forced with the NRLSSI reconstructions show the smallest solar heating rates. The use of the SATIRE reconstructions leads to stronger solar heating of up to 5 % in the middle and upper stratosphere. The SCIAMACHY observations slightly enhance the solar heating in the meso-1570 sphere, with differences arising from the stronger solar fluxes in the Huggins bands of the SATIRE model and enhanced fluxes in the Hartley bands of the SCIAMACHY data set. Using SORCE (SIM and SOLSTICE for wavelengths below 210 nm) measurements over the period May 2004 to November 2007 reveals larger changes in solar heating rates and the resulting temperatures in comparison to the NRLSSI data. The lower irradiance in the visible range at higher solar activity than at minimum activity in the SORCE/SIM data does not lead to a decrease in total radiative heating.
The spectral resolution of the SW radiation scheme of EMAC-FUB has recently been extended towards a more detailed representation of the Chappuis bands, and an update of Oberländer et al. (2012), performed with the extended spectral resolution for solar minimum conditions in November 1585 2007, separated for the UV and VIS spectral changes is presented in Fig. 9. The enhanced UV irradiance in 2004 (compared to 2007) in the SORCE data leads to higher SW heating rates while lower heating rates are simulated in the visible spectral range (Fig. 9a). The SW heating rate change from 1590 2004 to 2007 is stronger when the model is forced with the SORCE than with the NRLSSI data by 0.18 Kday −1 in the global mean (Fig. 9b). An enhanced sensitivity to changes in the Chappuis bands is found, a result which illustrates better the heating rate changes in the UV, visible and NIR spectral 1595 regions from using the SORCE data.
The different magnitude and spectral composition of the SSI changes leads not only to a substantial alteration of heating rates considered above but also affects photolysis rates which drive atmospheric chemistry and regulate the atmo-1600 spheric ozone distribution in the middle atmosphere (e.g. Brasseur and Solomon, 2005). The global ozone abundance is maintained by ozone production, destruction and transport by air motions. However, in the tropical stratosphere above ∼ 30 km the ozone concentration depends primarily 1605 on photochemical processes, with oxygen photolysis playing a crucial role in atmospheric chemistry. Therefore, the spectral composition and the magnitude of the SSI changes are of critical importance, since they define not only the magnitude but also the sign of the direct ozone response 1610 (e.g. Rozanov et al., 2002). The solar induced net effect of ozone in the stratosphere depends on the competition between ozone production due to oxygen photolysis in the Herzberg continuum (185-242 nm) and ozone destruction caused by the ozone photolysis in the Hartley band (between 1615 200 and 300 nm in the UV).

Effects on ozone and temperature from atmospheric model simulations
The effects of different SSI estimates on stratospheric ozone as deduced by CCMs or 2-D radiative-photochemical model 1620 simulations have been recently reported in a number of papers which are discussed and compared to each other in the following. In particular, a number of analyses focused on the differences in atmospheric response between simulations using the SORCE and the NRLSSI model 1625 results. The NRLSSI is the data set widely used for climate modelling purposes, including the SPARC CCMVal (Chemistry-Climate Model Validation) and the CMIP5 simulations for the upcoming IPCC report (Taylor et al., 2012). Cahalan et al. (2010) and Haigh et al. (2010Haigh et al. ( ) pub-1630 lished first the important implications of the SORCE data for middle atmosphere heating and therefore temperatures. Using simulations from a 2-D radiative-photochemical model Haigh et al. (2010) also presented effects on ozone.
To better understand the atmospheric sensitivity to the range of SSI estimates discussed above, these recently published model experiments using the NRLSSI and the SORCE data 1640 are compared to each other in Figs. 10 and 11 with respect to the solar signal in SW heating rates, temperatures and ozone. Additionally, results from three CMIP5 model experiments using the NRLSSI data set are shown. Table 2 presents an overview of the models and their experimental designs.

1645
Details on the spectral solar forcing used in each simulation are given in Table 3. The purpose is to provide the reader with an initial comparison of the impacts of two SSI data sets which represent the lower and upper boundary of SSI variations. One has to keep in mind, though, that all models 1650 used slightly different experimental setups (including slightly different spectral ranges for the SORCE data) and therefore an exact comparison between them as well as to observations awaits common coordinated experiments. Moreover, we note that the solar spectral forcing from the 1655 SORCE satellite used by all the models is derived from a relatively short period, and longer model simulations with full solar cycle forcing would be needed in order to better evaluate the differences in the response.

1660
Response in shortwave heating rates    no Marsh et al. (2012)    for SORCE closer to the HadGEM3 response, as both models are depicting full solar cycle conditions. In contrast to the other models, HadGEM3, the EMAC-FUB did not take into account solar induced ozone variations. EMAC therefore shows the weakest SW response in both experi-1720 ments, followed by the WACCM model. The simulations performed with the CMIP5 models fit nicely to the equilibrium simulations with the same (NRLSSI) forcing. Given the fact that the NRLSSI data provide a lower boundary for SSI variations with the solar cycle, the SW heating rate 1725 effects are likely enhanced in nature as compared to recent chemistry-climate model simulations. The SW heating rate differences described above may be attributed to the difference in the radiation codes of the models, as discussed in the previous section, but are also due to the different model se-1730 tups (for example, in simulations performed with or without interactive ozone, the signal will be enhanced or diminished).

Response in temperatures 1735
Corresponding to the SW heating rate differences, the tropical temperature differences for January are largest near the stratopause; using SORCE, the temperature differences between 2004 and 2007 are three to four times larger than with NRLSSI. Again, the GEOSCCM and the HadGEM 1740 models produce the strongest response with about 2 K and the WACCM and the FUB-EMAC model the weakest response. The response in SOCOL differs slightly from that of the other models by showing the maximum effect around 0.3 hPa, i.e. at a higher altitude than the other temperature response is approximately half of the response obtained using SORCE/SIM, indicating a large sensitivity of the model to the SSI data sets depicting their differences. Although none of the other models has performed a similar set of simulations, it is expected that this sensitivity is typical for all models shown here. Please note that the negative temperature response in SOCOL in the lower and middle stratosphere is not statistically significant. The temperature responses in the CMIP5 models closely agree with the NRLSSI equilibrium simulations. Comparison of the temperature response to observations over a full 11-yr solar cycle (e.g. Austin et al., 2008;Remsberg, 2008) reveals that the signals derived with the use the NRLSSI data generally shows a response peaking around the stratopause, with altitudes of maximum response varying between models according to their model setups as well as the use of solar induced ozone variations. As presented in Chapter 8 of the SPARC-CCMVal Report (2011), where the solar cycle effect on stratospheric temperature from a number of CCM simulations is discussed, the response to the NRLSSI forcing over the full solar cycle is about 0.8K, similar to observations. The response to the SORCE data forcing is 1770 larger than in the observations, but does not scale linearly with larger forcing. It was found to be higher by almost a factor of two in the case of the GEOSCCM simulation which corresponds to a full solar cycle (Swartz et al., 2012). Figure 11 displays the tropical ozone response in percent for month of January for those models that calculate ozone interactively. In the lower stratosphere, ozone changes 1780 induced by the forcing with SORCE data are larger and remain positive as compared to those calculated with NRLSSI data. In the middle to upper stratosphere ozone differences induced by the NRLSSI data are smaller but still positive whereas the ozone response using the SORCE 1785 data is negative. The height of the change from positive to negative ozone response varies from 5 hPa to 2 hPa between the models. Observations from 8 yr of SABER data indicate a transition altitude of about 1 hPa (Merkel et al., 2011) that is higher than for all models. This change of sign in the 1790 ozone response using SORCE data, meaning lower ozone during solar maximum than during solar minimum in the upper stratosphere, is statistically significant in all models. Swartz et al. (2012), who also examined the total ozone response and contributions from heating and photolysis, 1795 show good agreement between observations and simulations with the GEOSCCM model. Haigh et al. (2010) showed that the ozone decrease in the upper stratosphere and mesosphere is related to photochemical processes. Decreased O 3 at solar maximum is consistent with increased HO x and O and also 1800 leads to a self-healing effect with more UV radiation reaching lower levels, enhancing O 2 photolysis and therefore increased O 3 . Additional observational evidence is needed to confirm the sign reversal in upper atmospheric ozone response.

Dynamical changes and impact on the troposphere
So far the discussion focused on the response of stratospheric heating, temperatures and ozone to the forcing with NRLSSI 1810 and SORCE data. The direct solar response in the upper stratosphere induces indirect circulation changes throughout the stratosphere (e.g. Kuroda, 2002) and also affects the troposphere and the surface (Haigh, 1999;Kodera, 2002;Matthes et al., 2006). The impact on the troposphere using 1815 SORCE data has been reported by Ineson et al. (2011), who provide intriguing but provisional results on northern hemispheric winter circulation. Figure 12 displays January multimodel mean signals for three climate models (GEOSCCM, HadGEM3, and WACCM) in zonal mean zonal wind as well 1820 as 500 hPa geopotential height using SORCE data as external forcing. A stronger and statistically significant polar night jet during solar maximum years dominates the stratosphere and reaches down to the troposphere, whereas negative wind anomalies that are statistically significant in 1825 the lower stratosphere and upper troposphere only dominate equatorwards (Fig. 12, top panel). The zonal wind signals correspond to a positive Arctic oscillation (AO) pattern with a statistically significant stronger polar vortex and a partly statistically significant enhanced geopotential heights 1830 in midlatitudes (Fig. 12, bottom panel) in agreement with observations (e.g. Kodera, 2002). The stronger polar vortex in the stratosphere leads to a stronger positive phase of AO and North Atlantic Oscillation (NAO), which means a stronger Iceland low and higher pressure of the Azores 1835 and hence an amplified storm track over Europe. This in turn leads to mild conditions over northern Europe and the eastern US and dry conditions in the Mediterranean. However, the tropospheric AO and NAO already respond to lower UV irradiance variations over the 11-yr cycle, as is shown 1840 in Langematz et al. (2013) who obtained a similar AO/NAO response from a transient EMAC-FUB simulation for the period 1960-2005 using NRLSSI data as TOA input. The signal is seen also in Matthes et al. (2006), but in equilibrium simulations with NRLSSI as TOA input. As shown by 1845 Ineson et al. (2011) these patterns imply that the solar cycle effect on the AO/NAO contributes to a substantial fraction of the typical year-to-year variations and provides therefore a potentially useful source of improved decadal climate predictability for the Northern Hemisphere. Note that the re-1850 sponse is regional and is negligable on the global average. However, a caveat is that the 11-yr solar cycle variability cannot be forecasted into the future (on a daily, yearly or decadal time scale).
Even though the solar variability on time scales longer 1855 than the 11-yr solar cycle is beyond the scope of this paper, we should note here again that the Sun is the fundamental energy source of the climate system. As such, the low so-

Discussion of CCM results
We described the impact of NRLSSI and SORCE data, which 1865 represent the lower and upper boundaries of SSI solar cycle estimated variations, on the atmosphere and climate as depicted in CCM simulations. The NRLSSI reconstructions provide the standard data base for simulations of the recent past and future (e.g. SPARC-CCMVal, 2010; Taylor et al., . The atmospheric response with respect to this standard data set is compared to that derived from a different SSI estimate to understand not only the single model responses, but also to point out the importance and robustness of solar cycle signals for climate simulations. In particular, it is 1875 worth mentioning that an enhanced spectral resolution in the radiation codes leads to enhanced sensitivity in the response. We also described the important role of solar induced ozone changes for the amplification of solar effects on atmospheric composition, circulation and climate. Model simulations using the SORCE measurements as compared to the NRLSSI model show larger (by a factor of two) SW heating and temperature signals. The lower irradiance in the visible range during higher solar activity than at minimum activity in the SORCE measurements does not 1885 affect the increase in total radiative heating. Recent atmospheric model simulations with enhanced spectral resolution, however, point to the importance of the Chappuis bands in lower stratospheric heating. The solar ozone signal derived when the NRLSSI data are used is positive throughout the 1890 stratosphere and mesosphere, whereas the sign is reversed in the upper stratosphere when the SORCE data are used. Observations indicate a reversal of sign in the solar cycle ozone response in the upper atmosphere, however, the transition altitude for the sign change is higher than models suggest. This 1895 result needs to be confirmed by other satellite ozone measurements as observational evidence of the solar ozone signal is still limited. Finally, the tropospheric and surface solar cycle response has been presented in an ensemble mean and it has been highlighted that while these solar induced changes 1900 are of minor importance for globally averaged temperatures there are larger regional responses.

Conclusions
This paper presents an overview of our present knowledge of the impact of solar radiative forcing on the Earth's atmo- 1905 sphere. It covers the observations and the modelling of the solar radiative input as well as the modelling of the Earth's atmospheric response. The focus is on satellite-era (i.e. post-1970) data for which direct solar irradiance observations are routinely available. Special attention is given to the role of 1910 the UV spectral region, whose small contribution to TSI is compensated by a high relative variability with a potentially amplified influence on climate through radiative heating and ozone photochemistry. There is today clear evidence for the impact of solar variability on climate but both its mag-1915 nitude and its confidence level are still subject to considerable debate. One major challenge lies in the extraction of the weak solar signal from the highly variable atmospheric state. Recent progress has been made along two directions. The first one is the assessment of the magnitude of secu-1920 lar trends in solar radiative forcing through the reconstruc-tion of solar activity on centennial and millenial time scales from indirect indices such as cosmogenic isotopes, (e.g. Solanki et al., 2004;Bard and Frank, 2006;Beer et al., 2006;Usoskin, 2008;Schmidt et al., 2012). The second one, which we focus on, deals with shorter time scales, and is about recent solar variability and its impact on the lower and middle atmosphere. The global physical mechanisms that cause the solar irradiance to change in time and eventually impact climate have been well documented, (e.g. Haigh et al., 2005;1930Haigh, 2007Gray et al., 2010;Lean and Woods, 2010, and references therein). More than three decades of SSI observations are now available, but they are highly fragmented and agree poorly because of the difficulties in making radiometrically calibrated and stable measurements from space. Merg-1935 ing these different observations into one single and homogeneous record is a major and ongoing effort. As a consequence, several models have been developed for reproducing the SSI and its variability. The most successful ones are semi-empirical models such as SATIRE, SRPM, and 1940 COSI, which describe the SSI in terms of contributions coming from different solar surface magnetic features such as sunspots and faculae. Most models nowadays reproduce SSI measurements on short term time scales fairly well. However, uncertainties in SSI changes still remain on long term 1945 time scales and in the 220-400 nm band, which is of particular interest because of its impact on stratospheric ozone. These modelled or observed variations in the SSI are today used as inputs to CCM simulations that are capable of properly reproducing most aspects of stratospheric heating and 1950 point to the existence of a significant impact of solar variability on climate. However, major uncertainties remain in their detailed description, in which nonlinear couplings and regional effects can play an important role.
The main topics discussed in this paper and main conclu-1955 sions of our study are: -The spectral and temporal coverage of the SSI measurements gradually improved over time, and with the full operation of the SORCE mission in April 2004, daily observations of the full UV, visible and NIR spectrum 1960 became available. Unfortunately, this situation is likely to end in 2013, when SORCE is anticipated to succumb to battery failure. The lack of SSI observations has led to intensive application of semi-empirical models. There remains a considerable issue in assimilating 1965 SSI observations in such models and in reconstructing the SSI prior to the space age. Reconstructions going back to the early 20th century can be derived from semiempirical models based on historical ground-based solar observations (e.g. Ermolli et al., 2009).
-Recent SSI measurements by SORCE suggest a larger (factor 2-6) variability in the UV (200-400 nm) during solar cycle 23, which is hard to reconcile with earlier SSI and TSI observations (e.g. by UARS/SOLSTICE and UARS/SUSIM) and with SSI models. However, 1975 new estimates of possible calibration corrections suggest that the UV irradiance variation derived from SORCE measurements might be almost a factor of two weaker than reported earlier (Woods, 2012).
-TSI alone does not adequately describe the solar forc-1980 ing on the atmosphere and therefore SSI variations have to be taken into account in climate models. For many years, the canonical value of the average TSI was 1365.4 ± 1.3 Wm −2 whereas now the most accurate, and generally accepted, value is 1361 ± 0.5 Wm −2 1985 (Kopp and Lean, 2011;. -There has been steady progress in modelling SSI variations, and a number of models are now available that can be used as input for climate studies. However, the main uncertainty in the models concerns the wave-1990 length range 220-400 nm, where the magnitude of the variations differs by as much as a factor of three between models. This range is of particular interest for climate studies. Of the five SSI models discussed, specifically NRLSSI, SATIRE-S, COSI, SRPM and OAR, 1995 only one (SRPM) shows a behaviour of the UV and visible irradiance qualitatively resembling that of the recent SORCE/SIM measurements. However, it should be noted that the integral of the SSI computed with SRPM over the entire spectral range (i.e. the TSI) does not re-2000 produce the measured cyclical TSI changes. None of the other four models, which are in closer agreement with each other, reproduces the peculiar behaviour of the UV and visible irradiance observed by SORCE/SIM.
-While there has been major recent progress in better re-2005 producing the SSI changes on short term time scales, there remains now an important issue in the derivation of realistic confidence intervals, both for the observations and for the model results. Further common metrics should be used for comparing them.

2010
-Within the range of recent SSI values from observations and semi-empirical models, the NRLSSI model and SORCE observations represent, respectively the lower and upper limits in the magnitude of the solar cycle variation in the UV.

2015
-Results obtained with CCMs show that the observed increase in UV irradiance at solar activity maximum compared to minimum leads to an increase in atmospheric heating rates and correspondingly an increase in stratospheric temperatures. This direct atmospheric response 2020 is larger for larger UV forcing, i.e. for SORCE as compared to NRLSSI, and is sensitive to resolution in the radiation code. Currently there is insufficient observational evidence to support recent SSI measurements by SORCE on the basis of comparisons between climate 2025 model simulations and atmospheric ozone or temperature observations in the stratosphere. The larger UV forcing also leads to a larger surface response. The surface effect is regional and has little influence on globally averaged temperatures.

2030
-Accurate representation of the spectral nature of the incoming radiation, especially at wavelengths below 320 nm, and therefore ozone photochemical variations in the model simulations, is important since these changes amplify the atmospheric solar signal. A more 2035 accurate representation of the Chappuis absorption band in a radiation scheme revealed enhanced sensitivity in the heating rate response, which again highlighted the importance of a well resolved radiation code.
A unique aspect of this study is the description of the solar 2040 terrestrial connection by an interdisciplinary team of solar and atmospheric physicists. Progress on this hotly debated issue has often been hampered by the fact that limitations on observations or on models are not always properly known outside of a given scientific community. For the first time 2045 a comprehensive comparison and discussion of all relevant SSI measurements and models available for climate studies is presented, as well as a first investigation of their impacts on Earth's climate within a number of different CCMs. These results highlight the importance of taking into account in fu-2050 ture climate studies SSI variations and their effects on the Earths atmosphere. Major efforts, however, are still needed in each of the three research areas covered in this study. Resolving differences in existing long-term time series of solar measurements is a 2055 major challenge; forthcoming initiatives for merging these measurements into single homogenous databases will be of invaluable help (see, e.g. http://projects.pmodwrc.ch/solid/). Overcoming inaccuracies of current SSI models is another important challenge that will benefit from the renewed inter-2060 est in the solar radiative output, following the recent solar minimum. Finally, the realistic evaluation by climate models of the solar impact on the Earths climate is an important issue. For example, we focussed on the effect of the solar forcing without quantifying the impacts on amplification and 2065 feedback mechanisms. This should be done in a coordinated set of CCM experiments where the treatment of SSI inputs to the models are completely specified and results are robustly comparable with each other. Then it will be also possible to investigate the effects of the top-down feedback and for 2070 CCMs with an interactive ocean also the bottom-up feedback mechanism.
the COST Action ES1005.
We make the use of SORCE/SIM v17, SORCE/SOLSTICE v12, TIMED/SEE v10, UARS/SOLSTICE v18, UARS/SUSIM v22, SBUV, and SME data from the LISIRD at LASP University of Colorado and members of the instrument teams. Our special 2085 thanks are to all instrument teams that made all the observations that were used in this study, as well as to the modelling groups that provided published and partly also unpublished data for this study, in particular to J. Haigh We thank the anonymous reviewers, the editor Dr. W. Ward, and all participants in the discussion forum for their comments that helped