Regionalized dynamic climate series for ecological climate impact research in modern controlled environment facilities

Abstract Modern controlled environment facilities (CEFs) enable the simulation of dynamic microclimates in controlled ecological experiments through their technical ability to precisely control multiple environmental parameters. However, few CEF studies exploit the technical possibilities of their facilities, as climate change treatments are frequently applied by static manipulation of an inadequate number of climate change drivers, ignoring intra‐annual variability and covariation of multiple meteorological variables. We present a method for generating regionalized climate series in high temporal resolution that was developed to force the TUMmesa Model EcoSystem Analyzer with dynamic climate simulations. The climate series represent annual cycles for a reference period (1987–2016) and the climate change scenarios RCP2.6 and RCP8.5 (2071–2100) regionalized for a climate station situated in a forested region of the German Spessart mountains. Based on the EURO‐CORDEX and ReKliEs‐DE model ensembles, typical annual courses of daily resolved climatologies for the reference period and the RCP scenarios were calculated from multimodel means of temperature (ta), relative humidity (rh), global radiation (Rg), air pressure (P), and ground‐level ozone and complemented by CO2. To account for intra‐annual variation and the covariability of multiple climate variables, daily values were substituted by hourly resolved data resampled from the historical record. The resulting present climate Test Reference Year (TRY) well represented a possible annual cycle within the reference period, and expected shifts in future mean values (e.g., higher ta) were reproduced within the RCP TRYs. The TRYs were executed in eight climate chambers of the TUMmesa facility and—accounting for the technical boundaries of the facility—reproduced with high precision. Especially, as an alternative to CEF simulations that reproduce mere day/night cycles and static manipulations of climate change drivers, the method presented here proved well suited for simulating regionalized and highly dynamic annual cycles for ecological CEF studies.


| INTRODUC TI ON
Climate change alters ecosystem functioning worldwide with severe consequences for biodiversity, ecosystem services and for the well-being of humankind (IPBES, 2019;IPCC, 2013). In order to understand the impact of climate change on terrestrial ecosystems and to develop sustainable management strategies, scientific experiments with controlled modification of environmental drivers are essential. Previous studies in this field have largely been limited to the experimental manipulation of one or a limited set of meteorological drivers. Thus far, particular focus has been on the modification of atmospheric CO 2 , air temperature (t a ), as well as precipitation and tropospheric ozone (O 3 ) concentration (reviewed by Ainsworth & Long, 2005;Ainsworth et al., 2012;Beier et al., 2012;Lin et al., 2010). Such experiments are helpful in identifying general mechanisms of physiological and ecological responses (De Boeck et al., 2015), but ignore the covariation of multiple, physically interdependent variables. Responses by ecological systems to the simultaneous manipulation of multiple environmental drivers are unique and cannot be directly extrapolated from the response to each of the drivers manipulated individually (Ogle et al., 2021;Suzuki et al., 2014). For example, monthly mean surface temperatures and precipitation are tightly linked (Trenberth & Shea, 2005), and periods of reduced soil moisture availability covary with temperature and high light intensity (Suzuki et al., 2014).
Even large-scale field studies only allow the rough manipulation of a limited number of environmental variables (e.gBurkart et al., 2009;Eastburn et al., 2010) and should be ideally embedded in a framework including experiments in controlled environment facilities (CEFs) and computational modeling (Hanson & Walker, 2020;Roy et al., 2020a). In this context, modern CEFs offer the possibility to precisely and dynamically regulate many environmental conditions. In addition to the common control of t a , relative humidity (rh), and CO 2 , modern LED lighting provides a multispectral, dynamic light regulation. Some facilities accurately dose and monitor O 3 , NO X , or stable isotopes ( 13 C, 18 O). Nonmeteorological parameters such as soil moisture and nutrient supply are also commonly manipulated in lysimeter planters and automatically controlled. Given the high quantity of meteorological and ecological parameters that can be independently controlled, these CEFs are commonly referred to as "ecotrons" (Roy et al., 2020a).
Nevertheless, only a limited number of CEF studies exploit the technical possibilities of their facilities. Leisner et al. (2018) showed that in 80% of 57 reviewed agricultural CEF studies, static day and night temperatures were applied. For climate change scenarios, both temperature and CO 2 were typically increased in discrete steps, more or less based on previous experience with thresholds and tipping points, rather than according to model predictions (Leisner et al., 2018). This simplified representation of future climatic conditions may substantially limit the validity of controlled growth experiments with respect to the regional impacts of climate change on ecosystems.
In order to overcome this shortcoming, attempts have been made to incorporate seasonal and diurnal variability in present and future climate (FC) scenarios employed in CEF experiments. Thompson et al. (2013) summarized those attempts as increment studies, extreme event studies, and down-scaled climate studies.
According to the authors, increment studies aim to preserve diurnal meteorological dynamics by imposing a fixed incremental increase on a natural meteorological quantity. This approach assumes a uniform shift over the entire diurnal and seasonal cycle (e.gGhirardo et al., 2020;Hayes et al., 2019), for example, a model predicted mean increase in t a . However, the general consensus maintains that FC will not only be characterized by a shift in mean meteorological quantities, but will also see a significant increase in the variability, intensity, frequency, and duration of extreme meteorological events such as droughts or heavy rainfall (Jentsch et al., 2007). There is evidence that extreme weather events coupled with gradual climate trends may push ecosystems beyond their tipping-points (Harris et al., 2018). For CEF studies, this implies extreme intensities of a meteorological driver to be temporally superimposed on an existing climate series (Roy et al., 2016;Thompson et al., 2013). Both increment and extreme event studies preserve the natural temporal variability of a meteorological quantity, without considering interdependency and co-variability of multiple environmental quantities.
In recent years, it has been attempted to reproduce climatic conditions of present climate (PC) and FC scenarios in CEF experiments as realistically as possible, taking into account the co-variability of multivariate drivers. This requires global climate model (GCM) output to be adapted for application in controlled environments. GCMs are usually available at relatively coarse temporal resolution. However, in order to capture not only large-scale seasonal, but also regional diurnal dynamics, CEF studies require climate series with high temporal (hours) and spatial (a few kilometers) resolution. This can be achieved by statistical or dynamic downscaling of GCMs to the regional scale (Giorgi & Gutowski, 2015;Wilby et al., 1998). Thompson et al. (2013) developed one of the first ecological applications of statistical downscaling for CEF experiments. To generate valid temperature series for the year 2100, the authors combined the MIROC GCM with statistical information from real observations, and then employed a stochastic weather generator to obtain data at an hourly resolution. Similarly, Roy et al. (2016) used global information from the ARPEGEv4 GCM to simulate realistic 2040-2060 climate forcing at the Montpellier CNRS Ecotron facility. To regionalize the GCM output, the authors used the multivariate statistical downscaling method developed by Boé et al. (2006). The authors generated regionalized climate series through conditional resampling using data from the historical record to match statistical properties of a GCM output. By resampling, the natural variability and the covariance within multivariate climate series was preserved. Recently, Vanderkelen et al. (2020) presented an elegant method for forcing the UHasselt Ecotron units directly with the output of a single well-defined combination of a GCM with a regional climate model (RCM). Their method required the sophisticated identification of the best-performing GCM:RCM simulation for the mid-21st century from the Coordinated Downscaling Experiment-European Domain (EURO-CORDEX) model ensemble for the site of the ecotron. Their method not only accounts for the co-variation between climatic variables and their projection in variability, but also represents extreme weather events. In contrast, Leisner et al. (2018) (Field et al., 2012). In contrast, multimodel averages allow a robust simulation of the mean meteorological conditions of a specific climate scenario. Therefore, they are particularly useful for comparing the mean responses of ecological systems among various FC scenarios. In this study, we present a methodological approach that allows the generation of robust, high-resolution dynamic climate series for application in ecological CEF studies based on GCM:RCM ensemble means. For this purpose, we selected GCM:RCM combinations from the EURO-CORDEX (Jacob et al., 2014) and ReKliEs-DE (Regional Climate change Ensemble simulations for Germany, Hübener et al., 2017) ensembles. For each of the nine selected GCM:RCM combinations, simulations of a reference period  and a future period  were selected, the latter forced by the opposing representative concentration pathway scenarios RCP2.6 (van  and RCP8.5 . Ensemble means were formed for each scenario and averaged over nine 0.11°-grid-points covering the study area. Subsequently, homogenous weather segments were identified in each time series and replaced with historical records of t a , rh, global radiation (R g ), and O 3 in one-hour-resolution using a resampling method. The generated annual Test Reference Years (TRYs) were complemented by CO 2 series and used to force the eight climate chambers of the Model Ecosystem Analyzer TUMmesa at the Technical University of Munich for six consecutive months each year from 2019 to 2021 within the "valORTree" project. First, we describe the methodological approach for generating the TRYs and then evaluate the applicability of the TRYs for ecological climate impact research in the TUMmesa CEFs.

| TUMmesa
TUMmesa is an interdepartmental research institution of the Technical University of Munich and one of the most modern publicly operated CEFs in Germany. The facility features eight identical experimental walk-in chambers ( Figure 1) engineered by regineering GmbH (Pollenfeld, Germany) that allow the generation of a range of ecological conditions (e.g., Yang et al., 2019;Zytynska et al., 2020), with precise control of t a , rh, light, CO 2 , and O 3 , as well as manipulation of soil moisture, soil temperature and nutrient supply in various planters and lysimeters. Each climate chamber provides an experimental space of 2.4 × 3.2 × 2.2 m (W × D × H) and is equipped with the following features: air inlet/outlet/ circulation unit, cooling and heating register, steam generator and humidifier unit, CO 2 and O 3 fumigation, LED lighting, 13 CO 2 labeling system as well as automatic irrigation/fertigation coupled with an automatic weighing system for planters and lysimeters.
Preconditioned air is uniformly directed into the experimental space across the fully perforated sidewalls, producing an average wind speed of <0.1 m s −1 . All aggregates, supply lines, and control cabinets are installed in immediate vicinity to the chambers.
The LED system (Vossloh Schwabe, Urbach, Germany) comprises 10 individually controllable sub-systems that allow near-realistic sunlight and PAR simulation within a spectral range spanning UV-B to far-red. The LED system provides a maximum PPFD of >1500 µmol m −2 s −1 in one meter distance from the panels. By default, LEDs are operated with 55% of maximum capacity, providing a PPFD of >800 µmol m −2 s −1 . An overview of the technical specifications and limitations of TUMmesa is presented in Table A1 (see Appendix S1) and by Roy et al. (2020).

| "valORTree" project and study area
This study is part of the "valORTree" project. In this project, O 3 dose-response functions for two economically important tree species European beech (Fagus sylvatica L.) and Norway spruce (Picea abies (L.) H. Karst.) were established under controlled conditions using a gradient approach. Furthermore, the future O 3 risk potential was evaluated for the RCP2.6 and RCP8.5 climate change scenarios and the parametrization of the FO3REST ozone deposition model was updated (see Bender et al. (2015) for model description).
Briefly, 10 beech and 10 spruce trees were arranged in each of the eight TUMmesa climate chambers, (Figure 1d). Trees, including a 20-L soil monolith, were harvested from a naturally regenerating forest. Tree age ranged 5-10 years with an average tree height of 110 ± 18 cm for beech and 76 ± 11 cm for spruce. Generating robust dose-response functions required continuous measurements of stomatal O 3 uptake. In order to allow realistic diurnal dynamics of stomatal regulation throughout the experiment, external meteorological variables were simulated as dynamically as possible while considering interdependencies between multiple variables. Therefore, it was crucial to generate robust, self-consistent multivariate climate series that reproduced the mean seasonal and diurnal dynamics of PC and FC scenarios at a representative study site.
The study site was selected based on three criteria: (i) availability of long-term meteorological records in hourly resolution, including accurate measurements of tropospheric O 3 concentration; (ii) location of the corresponding climate station in a forested area characterized by beech and spruce stands; (iii) free exposure and no immediate influence of anthropogenic combustion processes. Based on these criteria, the atmospheric measurement station "Spessart" (Jossgrund-Lettgenbrunn, Germany, Code DEHE026, 497 m asl, 50°09′52.0″N 9°23′58.0″E) was selected ( Figure 2). The station is located in the German Spessart mountain range and is operated by the Hessian National Office for Environment and Geology since 1986. Meteorological data, along with air pollutants, are recorded at the station at 3.5 m above ground level. For further processing, t a , rh, P, R g , and O 3 records of the 1987-2016 reference period were selected. Hereafter, the term reference climate (RC) refers to the 30-years average, daily resolved annual course of the parameters measured for the reference period at the climate station.
As CO 2 concentrations were not measured at the "Spessart" station, data from the nearest CO 2 monitoring station "Schauinsland" (operated by the German Federal Environment Agency (UBA), 1205 m.a.s.l., 47°54′49.7″N 7°54′27.9″E) were used. Quality control and gap filling were performed to obtain continuous hourly time series from 1987 to 2016. Precipitation was not considered because the objectives of the valORTree project required adequate soil moisture availability. However, to account for possible reduced precipitation levels and the resulting limited soil moisture availability, FC TRYs were replicated with a period of reduced soil moisture.

| Generation of regionalized TRY for PC, RCP2.6 and RCP8.5
The methodology to generating TRYs that represent annual courses of PC  and FC (2071-2100) is based on DWD (2017) and was further developed and implemented by MeteoSolutions GmbH (Darmstadt, Germany) (Gelhardt et al., 2021). The procedure includes the computation of climate signals (CSs) for the FC scenarios RCP2.6 and RCP8.5, the generation of reference climatologies and resampling from the historical record ( Figure 3).

| Data processing
The EURO-CORDEX initiative provides ensembles of regional climate simulations for the European domain (Jacob et al., 2014). The F I G U R E 2 Locations of the meteorological stations "Spessart" and "Schauinsland" in Germany (a). Views of the "Spessart" station (b) situated in a forested area of the Spessart mountains (c) and the "Schauinsland" station (d). Courtesy of the Hessian National Office for Environment and Geology (HLNUG) (b) and of C. Zinsius, German Environment Agency (UBA) (d) Schematic representation of the workflow for generating the TRY. Ensembles of regional climate simulations were obtained for PC and the FC scenarios RCP2.6 and RCP8.5 from GCM:RCM combinations. Next, CSs were calculated by subtracting PC variables from FC. The RC for PC was calculated as the average annual cycle of climate records at the station "Spessart." Reference climatologies for the RCPs were computed by adding the respective ensemble mean climate signal to the RC. Finally, characteristic weather segments recorded at "Spessart" were resampled from historical record to match statistical properties of the reference climatologies simulations combine a GCM from the Coupled Model Intercomparison Project Phase 5 (CMIP5) with a RCM or an empirical statistical downscaling (ESD) method. Wheras, dynamically downscaled RCMs are forced by GCMs at their initial and lateral boundaries to produce climate simulation data on a fine regional scale (Giorgi & Gutowski, 2015), ESDs are based on transfer functions that connect observations on the GCM scale to regional records (Kreienkamp et al., 2019).
Noise-reduction was performed on each time series using Fast Fourier Transformation, through which frequencies higher than third order were eliminated. CSs for t a and rh were obtained by calculating the difference between t a and rh of the PC and the FC simulations for each ensemble member and grid point. The CSs of both RCP scenarios were averaged over all ensemble members and grid points. Thus, multimodel mean annual climatologies of the CSs with data in daily resolution were obtained for RCP2.6 and RCP8.5 (Figure 4), which were added to the RC recorded at the "Spessart" station. RC of t a , rh, and R g was defined as the reference climatology (RC) for PC.

| Resampling
A resampling method (DWD, 2017;Gelhardt et al., 2021) was applied in order to compute characteristic TRYs of t a , rh, P, R g , and O 3 in hourly resolution for PC, RCP2.6, and RCP8.5. Homogeneous weather segments recorded at "Spessart" from 1987 to 2016 in daily resolution were resampled in order to match statistical properties of the corresponding period of the reference climatologies.
For the first segment starting from 1 January, all possible 10-30 days long segments of the 30-year record (=630 segments) were tested against the corresponding segments of the reference climatologies with respect to the differences in mean values of t a , rh, R g , and in the SD of t a , as well as (for all segments but the first) the absolute t a difference to the preceding segment. The differences were listed in ascending order and scores staring from 0 (smallest difference) were assigned. Scores for the difference in mean t a were multiplied with the factor 0.3 and for the SD of t a with 0.7. Subsequently, the scores for each possible segment were summed and the weather segment achieving the lowest score was chosen. This procedure was repeated for the following segments until 31 December. Finally, the daily data of the recombined weather segments were replaced by the corresponding hourly records of t a , rh, P, R g , and O 3 . In order to smooth the transition between individual segments, t a , rh, P, and O 3 (not R g ) were linearly interpolated between 8 h to the end of one segment and 8 h after the beginning of the next segment.

| Simulation of PC, RCP2.6 and RCP8.5 in TUMmesa
The capacity of the TUMmesa CEF to operate complex climate simulations by controlling parameters in accordance with the prescribed values during long-term operation was investigated during the first experimental campaign of the three-year project "valORTree." Data were recorded on 161 operational days between 16 April and 29 September 2019. Chambers C1 and C2 executed TRY RCP8.5, C3 and C4 executed TRY RCP2.6 and C5, C6, C7, and C8 executed TRY PC with an additional O 3 gradient. Throughout the experiment, soil moisture was monitored by custom made TDR sensors and was maintained at ~80% of field capacity by adjusting the drip irrigation time. Soil moisture of C2 and C4 was reduced to 30% of field capacity for ten days in August.
Prior to the implementation of the TRYs into the TUMmesa control program, the climate series were adjusted (Table 1) to meet the climate chambers' technical requirements (Table A1 in Appendix S1) or for experimental purpose. Air temperatures exceeding 30°C were limited to 30°C. Day/night t a below 10/4°C was increased to 10/4°C and rh was limited to 75%/90%. R g was converted to PPFD by multiplying R g by a month-specific empirical conversion factor ranging from 1.90 to 2.10 (Grünhage & Haenel, 2008). PPFD between 600 µmol m −2 s −1 and the maximum intensity of 2030 µmol m −2 s −1 F I G U R E 4 Annual course of climate signals calculated for the RCP2.6 (a,b) and RCP8.5 (b,c) scenarios at the climate station "Spessart." The climate signals indicate the difference between present and predicted daily averages of air temperature (δT a ) (a,c) and relative humidity (δrh) (b,d). Mean values ± SD, median and range of nine simulated climate series per scenario and grid point are shown were scaled to the range from 600 to 800 µmol m −2 s −1 .
PPFDs < 24 µmol m −2 s −1 were increased to 24 µmol m −2 s −1 due to minimum requirements of the LEDs ( Figure A1 in Appendix S1). The LEDs responded within seconds and no gradient was programed.  (691) Note: Range and mean (in brackets) of t a , rh, CO 2 , and O 3 are shown for the simulated year, the experimental period and the derived set points for the TUMmesa climate chambers. The total sum of photosynthetic photon flux density (PPFD) over the respective period is shown. The highlighted values (in italics) in column 3 of PPFD sums represent the percentage of PPFD sums achieved in TUMmesa relative to the simulated data. For PPFD and R g , overall means and the means of daily maximum values (in brackets) are presented.

| Test reference years
Three individual TRYs were generated for the climate station "Spessart" (summarized in Table 1). The TRYs consist of annual cycles for t a , rh, P, R g , O 3 , and CO 2 in an hourly resolution, representing one possible year in each of the periods 1987-2016 (PC) and 2071-2100 (RCP2.6 and RCP8.5) ( Figure 5). By recombining measured weather segments, the natural day-to-day variation of key meteorological variables is reintroduced (black line in Figure 5, only t a is shown), which was lost in the reference climatologies (blue line) due to longterm averaging of measured t a for PC and the calculation of ensemble means for RCP 2.6 and RCP8.5. Average annual t a , rh, P, R g , and

| Comparing prescribed to measured values
The TRYs were executed in TUMmesa on 161 of 167 operational days. Averaged over the entire experimental period, the maximum differences between measured and prescribed values were 0.1°C (t a ), 0.8% (rh), 13 ppm CO 2 (at PPFD > 100 µmol m −2 s −1 ), 4 ppb O 3 , and 113 µmol m −2 s −1 PPFD (Table 3). Prescribed values were approached in one-min-steps, resulting in 1,854,720 data pairs for each parameter. Data losses were below 1%.
For t a , 99% of measured values were within a ± 0.38°C deviation of prescribed values (Figure 7a). Deviations exceeding +1°C (<0.2% of all records) were caused mainly by failure of the adiabatic precooling system or by a defective precooling of the outside air inlet.
Deviations of more than −1°C (<0.2% of all records) were related to a malfunction of the chamber heating register.

F I G U R E 5
Annual course of air temperature of the TRYs generated for PC (a), RCP2.6 (b), and RCP8.5 (c) from reference climatologies by resampling of measured weather segments. RC for PC is the mean annual cycle recorded at the climate station "Spessart"   The difference in measured PPFD to prescribed values (Table 3) is explained by that fact that LEDs are controlled in relative steps from 0% to 100%. The relative intensities are related to a PPFD in 100 cm distance from the light source. Light intensity increases with decreasing distance to the source. In this study, the PAR sensor was installed in 70 cm distance from the LEDs to avoid shading by the trees.

| Control of CO 2 concentration
In total, 49.8% of all CO 2 records deviated more than the tolerated ± 20 ppm from prescribed values (Figure 8b). Only 0.5% of all F I G U R E 6 Probability density functions (PDFs) for temperature (a) and O 3 (b) of the TRY. In addition, PDFs of the mean annual cycles of the reference period  recorded at "Spessart" ± SD are shown, as well as the historically observed range. PDFs were obtained by kernel density estimation using a bandwidth of 1.5  During the nonphotoperiod, 99.7% of records were above the tolerated range as plant respiration caused massive increase of CO 2 levels (Figure 8a), frequently exceeding 700 ppm. To ensure that CO 2 concentrations during the day were not excessively influenced by increased nocturnal values outside air with lower CO 2 concentration was blown into the chamber for 3 h before daybreak in addition to active CO 2 removal by soda lime (Figure 8a). This enabled us to maintain CO 2 within the tolerated range for 71.2% of values when there was no influence of human respiration and PPFD was above 10 µmol m −2 s −1 (Figure 8c).

| Homogeneity among chambers
To verify the homogeneity among chambers executing identical time series, correlation matrices were calculated for the parameters t a , rh, O 3 , CO 2 , and PPFD using Pearson's correlation coefficient (r, Figure 9). The correlation coefficients for t a were 1, except between C1/C2, where it was 0.99. Over 99% of t a data pairs show less than a ± 1°C deviation among chambers. Minimum r was 0.93 for rh and 0.91 for O 3 , where 98% and 99% of all data pairs were within the tolerated deviation of ± 10% and ± 20 ppb, respectively.
Minimum r was 0.94 for PPFD and more than 90% of all data pairs showed less than ± 50 µmol m −2 s −1 PPFD deviation among the chambers. However, PAR sensors were sometimes shaded by the growing canopy, such that the actual deviations among individual chambers can be assumed to be lower. For CO 2 , lowest r was 0.88.
Nevertheless, 80% of all relevant CO 2 data pairs deviated less than ± 20 ppm among chambers that performed identical time series.
With PPFD > 10 µmol m −2 s −1 and doors closed, even 89% of deviations among chambers were within ± 20 ppm.  et al., 2017), but uncertainties arise concerning their future regional and local behavior due to their long lifespan in the hemispherical background (Turnock et al., 2019). Models are further complicated by interactions between VOCs and O 3 formation (Calfapietra et al., 2013;Peñuelas & Staudt, 2010), intercontinental O 3 transport (Derwent et al., 2004;Volz-Thomas et al., 2003) and influx of stratospheric ozone (Kawase et al., 2011). Thus, regional development of tropospheric O 3 concentrations particularly depends on a number of interacting factors, the prediction of which is subject to many uncertainties.

| D ISCUSS I ON
The atmospheric CO 2 concentration was not measured at the "Spessart" station. Assuming that the CO 2 concentration at the station is influenced on a rather large scale, data from the nearest CO 2 monitoring site "Schauinsland" were used. Unlike a forest canopy, atmospheric CO 2 at the station is largely unaffected by convective boundary layer effects and canopy dynamics. Therefore, pronounced diurnal CO 2 variation-as observed in forest canopies (Murayama et al., 2003)-are not represented in the TRYs. However, due to insufficient CO 2 -removal from the chamber atmosphere, nocturnal CO 2 increase due to respiration processes was also established in the climate chambers, although unregulated and on a higher level than expected for natural forest ecosystems. By temporarily supplying outside air in combination with the soda lime columns, the targeted CO 2 concentrations could be achieved for more than 70% of the photoperiod. By identifying a single model projection, the method respects not only the covariation between climatic variables but also their projection in variability, as well as possible extreme events. However, single models are strongly influenced by uncertainty in climate predictions resulting from structural differences in the GCMs as well as uncertainty due to variations in initial conditions or model parameterization (Semenov & Stratonovitch, 2010). These uncertainties could be overcome by applying several single-model projection in parallel in an ecological experiment (as suggested by Thompson et al., 2013)

| CON CLUS ION
The TRYs generated with the methodology described here capture possible changes in the mean values of important meteorological F I G U R E 8 Typical diurnal course of CO 2 concentrations (a) in the PC scenario. The tolerated deviation from prescribed values is shaded blue. Arrows indicate (i) start of outside air inlet, (ii) activation of CO 2 removal via soda lime and (iii) the deactivation of both outside air inlet and CO 2 removal. Percentiles of difference (Δ) between actual values and set points are shown in (b). In (c), only data pairs are considered for which the door was closed for at least one full hour and PPFD was above 10 µmol m −2 s −1 . The tolerated deviation from prescribed values is shaded blue. Red numbers in (b) and (c) indicate 0 (bottom left) and 100% (top right) quantiles drivers while maintaining intra-annual variability and covariability between the multiple drivers. By calculating multimodel means, the method is, however, not capable of reproducing extreme events in a sophisticated way, and changes in climate extreme indices for our RCP TRYs represent the shift in mean values (as indicated in the PDFs) rather than the presence of extreme events. Nevertheless, the method produces dynamic multivariate climate series for the implementation in ecological CEF studies that focus on general impacts of climate change on ecological systems on a regional scale.
The TRYs are a suitable alternative to CEF climate simulations based on simple day/night cycles and incremental manipulations of single climate variables. The TRYs were adequately simulated in the TUMmesa CEFs, with particularly good reproduction of absolute values and high-resolution dynamics of temperature, relative humidity, ozone, and light.

ACK N OWLED G EM ENTS
The study was financed by the German Environment Agency (Umweltbundesamt, UBA 3717512570). We further acknowledge the valuable contributions of Fanny Kittler, Melanie Hauer-Jákli and Michael Hubensteiner. Open Access funding enabled and organized by Projekt DEAL.

CO N FLI C T O F I NTE R E S T
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
The TRY for PC, RCP2.6 and RCP8.5 are available on Dryad (DOI https://doi.org/10.5061/dryad.h1893 1zn5).