The impact of oxygen content variability is investigated with two coupled atmosphere–ocean general circulation models (AOGCMs): Hadley Centre Coupled Model version 3 (HadCM3-BL) and Hadley Centre Global Environmental Model version 3 (HadGEM3-AO).

HadGEM3-AO is an AOGCM . The atmosphere component is the UK Met Office Unified Model version 7.3 in the HadGEM3-A r2.0 climate configuration . It employs a regular Cartesian grid of 3.75∘ longitude by 2.5∘ latitude (N48). In the vertical, 60 hybrid height vertical levels are employed – “hybrid” indicating that the model levels are sigma levels near the surface, changing smoothly to pressure levels near the top of the atmosphere . The model top is 84 km, which permits a detailed treatment of stratospheric dynamics. A 20 min time step is used. The model employs a non-hydrostatic and fully compressible dynamical core, using a semi-implicit semi-Lagrangian advection scheme on a staggered Arakawa C grid . Radiation is represented using the scheme with six shortwave and nine longwave bands, accounting for the radiative effects of water vapour, carbon dioxide, methane, nitrous oxide and ozone. The Met Office Surface Exchange Scheme 2 (MOSES2) land surface scheme is used , which simulates atmosphere–land exchanges and hydrology. A fixed present-day vegetation distribution of plant functional types is employed. The ocean component of the model is NEMO-OPA (Nucleus for European Modelling of the Ocean - Océan Parallélisé; ) model version 3.0 , run at a 96 min time step. In the vertical, 31 model levels are used, which increase in thickness steadily between 10 m in the shallowest to 500 m in the deepest layer at 5 km in depth. NEMO employs a tripolar, locally anisotropic grid (ORCA2; ) which permits a more detailed treatment of the north polar region and higher resolution in the tropics. This yields an approximate horizontal resolution of 2∘ in both longitude and latitude, with an increased resolution of up to 0.5∘ in the tropics. The sea-ice component of the model is CICE (Los Alamos Community Ice CodE) at version 4.0 , run at a 96 min time step. This treats sea ice in a five-layer model, allowing the simulation of different ice types. The atmosphere and ocean/sea-ice components exchange fields every 24 h, while NEMO and CICE exchange fields every time step. HadGEM3-AO can be thought of as a close relation to the newest-generation HadGEM3 coupled model that will be used to support the next Intergovernmental Panel on Climate Change (IPCC) assessment and so represents the state of science in numerical climate models.

HadCM3-BL is an AOGCM coupled to an interactive vegetation model. The model was originally developed by the United Kingdom Met Office Hadley Centre but has since been substantially developed further by the University of Bristol. The atmosphere component of the model employs a regular Cartesian grid of 3.75∘ longitude by 2.5∘ latitude. In the vertical, 19 hybrid height vertical levels are employed. A 30 min time step is used. The primitive equation set of is solved to conserve energy and angular momentum, solved on a staggered Arakawa B grid in the horizontal. Radiation is represented using the scheme with six shortwave and eight longwave bands, accounting for the radiative effects of water vapour, carbon dioxide and ozone, amongst other radiative active species. The ocean component of the model employs the same horizontal grid as the atmosphere component of the model, 3.75∘ longitude by 2.5∘ latitude. In the vertical, 20 model levels are used which increase in depth from 10 m in the shallowest layer to 616 m in the deepest layer. A time step of 60 min is employed and the ocean and atmosphere components exchange required fields once per day. The ocean component is based on the model, solving the full primitive equation set in three dimensions. A staggered Arakawa B grid is employed in both atmosphere and ocean models. Sea ice is treated as a zero thickness layer on the surface of the ocean grid. Ice is assumed to form at the base at a freezing point of -1.8 ∘C but can also form from freezing in ice leads and by falling snow. A simple parameterisation of sea-ice dynamics is also employed and sea-ice formation due to convergence from drift being limited to 4 m thick. Sea-ice albedo is fixed at 0.8 for temperatures below -10 ∘C, decreasing linearly to 0.5 at 0 ∘C. The MOSES2.1 land surface model is employed to simulate the fluxes of energy and water between the land surface and the atmosphere . TRIFFID (Top-down Representation of Interactive Foliage and Flora Including Dynamics; ) predicts the distribution of vegetation using a plant functional type (PFT) approach. TRIFFID is run in equilibrium mode with averaged fluxes calculated over a 5-year period. TRIFFID calculates vegetation properties for five PFTs: broadleaf trees, needleleaf tree, C3 grass, C4 grass and shrubs. Grid boxes can contain a mixture of PFTs based on a “fractional coverage co-existence approach” . Net primary productivity (NPP) is also calculated, using a photosynthesis-stomatal conductance model accounting for a number of factors including atmospheric oxygen content, which affects the photorespiration compensation point . The predicted vegetation distribution impacts the atmosphere component by altering surface albedo, evapotranspiration and surface roughness.

Both models simulated the climate response to oxygen variability in a pre-industrial Holocene (PIH) climate. HadCM3-BL was additionally run for three time periods across the Phanerozoic: the Maastrichtian (late Cretaceous, 66.0–72.1 Ma), Wuchiapingian (late Permian, 254.14–259.1 Ma) and the Asselian (early Permian, 295.0–298.9 Ma). The continental reconstructions employed were developed by and are from ©Getech. These reconstructions have been widely employed in a number of previous studies using the HadCM3-BL climate model (e.g. ). All three are periods of time in which models have suggested that atmospheric oxygen may have deviated significantly from the present level of 21 % (see Fig. ). Modifications were made to alter the oxygen content of the atmosphere by adjusting the mass mixing ratios of major and minor gases, the surface pressure and other physical characteristics of the atmosphere such as the specific gas constant in an analogous way to . Figure  shows the annual average surface temperatures and Fig.  shows the annual average precipitation for the (a) PI-CM21, (b) Ma-CM21, (c) Wu-CM21 and (d) As-CM21 simulations.

A summary of the experiments performed can be found in Table . When an experiment with a particular oxygen content is referred to, it will be indicated in superscript; e.g. EXP21 indicates a 21 % oxygen simulation. The 21 % simulations (PI-CM21, As-CM21, Ma-CM21 and Wu-CM21) were integrated for 50 model years, as these simulations had already been spun up at that CO2 content. For 10 % and 35 % pO2, the model was spun off the 21 % pO2 simulation and iterated for at least 1000 model years. The 2 × PI-CM*, 2 × Ma-CM* and 2 × As-CM* experiments were spun off from the end of the PI-CM, Ma-CM and As-CM experiments and iterated for 100 years in order to perform a analysis. For HadGEM3-AO, model integrations (PI-GEM35, PI-GEM10, 4 × PI-GEM35 and 4 × PI-GEM10) were performed for 300 model years with a 10× acceleration of the deep ocean to reduce the time for equilibrium, then integrated for a further 500 years to spin up the shallow ocean without acceleration. The last 50 years were used for model analysis.

Pre-Quaternary pCO2 is poorly constrained due to the absence of glacial ice; however, there is growing evidence that CO2 is unlikely to have been significantly higher than the order of hundreds of Pa since the radiation of land plants . For the Maastrichtian, 56 Pa is used in agreement with stomatal proxy-based reconstructions . For the Asselian, 28 Pa is used in agreement with carbonate and fossil plant reconstructions . For the Wuchiapingian, 112 Pa is used .

1-D atmospheric chemistry simulations have simulated higher O3 column with increasing pO2 . More detailed 2-D model simulations, which capture critical latitudinal gradients in photolysis and zonal-mean transport , support a monotonically increasing ozone column with increasing pO2 . However, simulated ozone column was more sensitive to N2O levels than pO2 . In addition, while column ozone reduces at low pO2 in , there are increases in ozone concentration in the tropical tropopause region where the radiative effect of O3 is stronger . Changes in lightning are important for understanding future changes in tropospheric ozone ; however, they are subject to considerable uncertainty . There may be more lightning at high pO2 due to a higher pO2/pN2 ratio or less due to reduced convection . Low pO2 may also enhance isoprene emissions , which could enhance tropospheric ozone and alter cloud properties . Ozone is also sensitive to changes in CH4 and N2O; the changes to inventories of these chemically active species on the Phanerozoic timescale are highly uncertain. Ozone is also sensitive to dynamical changes. Given these large uncertainties in possible changes to chemical sources, reactivity and transport, we neglected including changes in atmospheric ozone concentration in these simulations. However, we recommend that follow-up work should focus on this specific question in detail. In HadGEM3-AO, the mass of tropospheric and stratospheric ozone is fixed at PIH values simulated by using a tropopause height-matching scheme. This prevents a rising tropopause leading to stratospheric levels of ozone existing in the troposphere, particularly in the 4 × PI-GEM experiments. Not accounting for a rising tropopause has been found to artificially increase climate sensitivity and initial tests not accounting for this led to instability for 4 × PI-GEM10. In HadCM3-BL, tropospheric ozone is set to 6 ppbv and stratospheric ozone is set to 1.66 ppmv for the 21 % simulations. These values are adjusted to conserve total ozone mass in the alternative pO2 scenarios.

Data for the Cenomanian 21 % O2 and 10 % O2 simulations were obtained from https://www.ncdc.noaa.gov/paleo/study/18776 (last access: 21 February 2018). At the time of writing, the 35 % simulation contained missing data, so it was not used for analysis.

A 1-D energy balance model (EBM) has been used to deconvolve the contributions from changes in different parts of the climate system. This 1-D EBM approach has been applied to zonal-mean quantities for climate simulations of the Eocene by following and : 1SWt↓(ϕ)1-α(ϕ)-12πR2cos(ϕ)∂F(ϕ)∂ϕ=ϵ(ϕ)σTs,ebm4(ϕ), where SWt↓ is the incident shortwave radiation at the top of the atmosphere, ϕ is the latitude, α is the surface albedo, R is the radius of Earth, ϵ is the effective surface emissivity, and Ts,ebm is the EBM surface temperature . ∂F(ϕ)∂ϕ is the divergence of total meridional heat transport and is given by 2∂F(ϕ)∂ϕ=-2πR2cos(ϕ)SWtnet(ϕ)+LWtnet(ϕ), where SWtnet and LWtnet are the net top-of-atmosphere shortwave and longwave radiative fluxes, respectively (positive downward; ). Solving for Ts,ebm, the EBM surface temperature for each latitude can be calculated using zonal- and annual-mean radiative fluxes from the GCM. Where clear-sky radiative fluxes are also available, cloud radiative effects can be deconvolved from clear-sky radiative effects. The clear-sky albedo αc and clear-sky effective surface emissivity ϵc can be calculated by 3αc=SWt,c↑SWt↓,ϵc=LWt,c↑LWs↑, where SWt,c↑ is the upward top-of-atmosphere clear-sky shortwave radiative flux and LWt,c↑ is the upward top-of-atmosphere clear-sky longwave radiative flux. When considering the temperature change between two experiments, the contributions from different components can be quantified by calculating Ts,ebm with different combinations of components from each experiment .

To estimate the climate sensitivity to CO2 changes, the linear regression methodology of is employed. This assumes a linear relationship between the changes in global, annual-mean radiative imbalance at the top of the atmosphere (N, Wm-2) and surface temperature anomalies (Ts, ∘C): 4N=F+ξΔTs, where ξ is the effective climate feedback parameter (Wm-2∘C-1) and F is the effective forcing (Wm-2) accounting for fast climate adjustments and effective radiative forcing. The effective ECS is then ΔTs when N=0. While there are weaknesses of this approach, particularly due to non-linearities in ξ as ΔTs changes , the climate response when simulations are continued to equilibrium shows an accuracy to within 10 % . Furthermore, the contributions to ξ and F from longwave (LW) and shortwave (SW), clear-sky (CS) and cloudy-sky (CRE) components can be decomposed by a linear decomposition as 5F=FCS,SW+FCS,LW+FCRE,SW+FCRE,LW for the effective forcing and 6ξ=ξCS,SW+ξCS,LW+ξCRE,SW+ξCRE,LW for the effective climate feedback parameter.

Figure  (left) shows the annual-mean surface air temperature differences between the 35 % and 10 % runs. For the pre-industrial Holocene, PI-GEM1035 (Fig. a) shows a global-mean surface temperature response of +1.50 ∘C, while PI-CM1035 (Fig. b) shows a similar global-mean surface temperature response of +1.22 ∘C. It is worth noting that HadCM3-BL and HadGEM3-AO are not completely distinct climate models, for instance, sharing the radiation scheme, so this is unlikely to capture the full variability in possible climate model responses. That the results are in reasonable agreement with the 1-D results of , who simulated a temperature response between +1.05 and +2.21 ∘C depending on assumptions about atmospheric ozone, gives some confidence in the HadCM3-BL and HadGEM3-AO results. Similarly, the As-CM1035 (Fig. c) case exhibits a global-mean surface temperature response of +1.29 ∘C. For the warmest climates, a response of -0.82∘C is simulated for Wu-CM1035 (Fig. e) and +0.17 ∘C for 4 × PI-GEM1035 (Fig. f). In the Ma-CM1035 case (Fig. d), a global-mean surface temperature response of +0.70 ∘C is simulated. This suggests that the climate response to pO2 variability depends on the background climate state.

There is a strong seasonal dependence in the surface air temperature response. Considering the changes in coolest average monthly temperature in each grid box (Fig. , middle column), the change in cool months dominates the warming response, particularly at high latitudes. By contrast, the warm-month mean is smaller/less negative in all cases (Fig. , right). A cooling of continental land masses is evident in the tropics and particularly in the Wu-CM (Fig. e) and 4 × PI-GEM (Fig. f) cases. These could be in part due to free-air lapse rate changes which should be stronger at high pO2; as for a given topographic height, the change in pressure is higher for high pO2, which should lead to a larger temperature reduction with height. The changes to the seasonal cycle are consistent with the radiative changes associated with changing oxygen content. The reduction in incident surface shortwave radiation should have its strongest effect on extratropical temperatures in the summer; therefore, the Rayleigh scattering component will most strongly affect the warm-month temperature. Warming from pressure broadening of greenhouse gas absorption lines as atmospheric mass increases will be most evident in extratropical winter, as with anthropogenic climate change, due to sea-ice and surface heat flux changes . The reduction in the amplitude of the seasonal cycle in temperature simulated by both HadGEM3-AO and HadCM3-BL is therefore supported by a consideration of the changes to atmospheric radiation.

The zonal- and annual-mean surface air temperature changes are shown in Fig. . The Northern Hemisphere Equator-to-pole temperature gradient is reduced by 6.6 ∘C in the PI-GEM1035 case (blue line) and 4.0 ∘C in the PI-CM1035 case (not shown). The zonal structure of the surface temperature change is similar in the palaeoclimate case studies. In the Maastrichtian, the Equator-to-pole temperature gradient is reduced by 2.0 ∘C (Ma-CM1035) and in the Asselian the Equator-to-pole temperature gradient is reduced by 2.3 ∘C (As-CM1035). The Equator-to-pole temperature gradient reduces even in the Wu1035 case despite the reduction in global-mean surface temperatures.

The hydrological cycle is also affected by changing oxygen content. Increases in Rayleigh scattering at high pO2 ought to reduce incident shortwave at the Earth's surface and inhibit convection , which should lead to reductions in precipitation. This is analogous to stratospheric sulfate or solar radiation management geoengineering where precipitation is reduced in geoengineering experiments with respect to an unperturbed climate with the same global-mean surface temperature . simulated large reductions in precipitation as pO2 increased in the GENESIS climate model; however, much of this could be explained by the surface temperature changes. Annually averaged precipitation change between the 10 % and 35 % oxygen content runs is shown in Fig. . In all cases, increasing oxygen content leads to a decline in global-mean total precipitation, despite the increase in surface temperatures, however, with strong regional differences. For PI-GEM (Fig. a), PI-CM (Fig. b) and 4 × PI-GEM (Fig. f), there is a clear northward shift in the tropical rainbelts. A northward shift in the Intertropical Convergence Zone (ITCZ) would be consistent with stronger warming in the Northern Hemisphere due to Bjerknes compensation . Heat transport is more hemispherically symmetric in the Maastrichtian, Asselian and Wuchiapingian cases so latitudinal ITCZ shifts are not evident. While global precipitation is reduced in Wu-CM35, the increase in ocean–land temperature contrast leads to a significant increase in tropical land precipitation. Despite the increases in global-mean surface temperatures simulated for most cases, precipitation is still reduced in all simulations.

Comparing the surface temperature and precipitation response between HadCM3-BL and HadGEM3-AO suggests that the model responses are broadly consistent. A grid-box-by-grid-box comparison of annual-mean surface air temperature and precipitation anomalies for PI-GEM1035 vs. PI-CM1035 is presented in Fig. S1 in the Supplement. The largest discrepancy in surface air temperature response between the two models occurs for the largest temperature changes simulated by HadGEM, which are strongest in Northern Hemisphere polar regions. This could be linked to differences in the representation of polar climate processes and amplification by polar ice feedbacks between the two models. There is broad consistency in cold- and warm-month means (Fig. a and b) with stronger warming in the cold-month mean and terrestrial cooling in the warm-month mean.

The drivers of the changes in surface temperature can be understood by decomposing the terms which contribute to surface temperature change in a 1-D energy balance model following . For PI-CM1035, these results are shown in Fig. . These show that the 1-D EBM can reasonably capture the temperature response in the HadCM3-BL simulations, with slight errors (where the black and grey lines are not overlapping) evident in the polar regions. This could be due to averaging over the polar rows in the HadCM3-BL model. There are positive contributions to the surface temperature change in the clear-sky emissivity and albedo at the poles. This is consistent with the increase in pressure broadening of absorption lines and the simulated reduction in sea-ice extent. By contrast, extrapolar contributions to clear-sky albedo provide a negative contribution to the temperature change which is consistent with an increase in Rayleigh scattering which would be expected to be strongest in the tropics where the maximum in incoming solar radiation is located. Combined, the clear-sky component of the temperature change is +1.45∘C and the cloudy-sky component is -0.35∘C. This suggests that HadCM3-BL supports a cloud feedback which acts to cool the climate at high pO2 and partially offset the clear-sky temperature changes.

The same analysis was performed for the HadGEM3-AO PIH simulations. Figure  shows that a somewhat weaker cloud feedback is simulated by HadGEM3-AO (-0.21 ∘C). Clear-sky contributions are slightly stronger (+1.51 ∘C). The largest differences between the simulations appear in the all-sky albedo and emissivity changes, where there appear to be competing factors which lead to a similar climate response possibly related to partitioning between the longwave and shortwave contributions to the cloud response. This is perhaps unsurprising, as cloud feedbacks to CO2 changes represent a large uncertainty in future climate change projections, and given the relatively small global-mean temperature changes, a relatively small change in cloud radiative effects has the power to considerably mediate the climate response. However, the qualitative agreement in latitudinal structure of the clear-sky albedo and emissivity changes between these structurally different models gives some confidence that the relevant climate feedbacks are well captured in these simulations.

Analysis of the palaeo case studies (As-CM, Fig. S2; Ma-CM, Fig. S3; Wu-CM, Fig. S4) shows a similar pattern. In all simulations, irrespective of surface temperature response, the clear-sky emissivity is a positive contribution to global-mean surface temperature change, while clear-sky albedo is a more negative contribution. The emissivity contribution becomes less positive as pCO2 increases from As-CM to Ma-CM to Wu-CM. By contrast, the albedo contribution becomes more negative as pCO2 increases. This is consistent with the reduction in planetary albedo as sea-ice extent is reduced on the ocean and dark vegetated surfaces increase on the land.

Here, we investigate the impact of oxygen variability on climate sensitivity. The HadGEM3-AO and HadCM3-BL results suggest that increasing CO2 content leads to a reduction in the surface temperature change on increasing pO2 (compare 4 × PI-GEM and PI-GEM in Fig. ). For reference, HadGEM3 has a climate sensitivity of +3.6 ∘C , while HadCM3 has a climate sensitivity of +3.1 ∘C . From the 4 × PI-GEM and PI-GEM experiments, a reduction in climate sensitivity of 0.65 ∘C can be inferred based on the changes in surface temperatures. For HadCM3, CO2-doubling experiments were performed, and a regression of the change in top-of-atmosphere radiative imbalance against change in surface temperature following (see also Sect. ) is shown in Fig. . The PI-CM35 climate state has a smaller ECS than PI-CM10 by 0.7 ∘C. While the changes in total radiative forcing F are very similar, ξ is less negative (-1.08 vs. -1.37 Wm-2∘C-1) at low pO2. The decomposition of these changes into their longwave and shortwave components, clear-sky and cloudy-sky components is also shown in Fig. . The clear-sky longwave radiative flux changes are slightly higher in 2 × PI-CM35 (4.0 Wm-2) than 2 × PI-CM10 (3.8 Wm-2) as would be expected due to the pressure broadening of CO2. The clear driver for the less negative ξ value are from the longwave cloud radiative effect changes, which is much steeper for 2 × PI-CM10 (+0.62 Wm-2∘C-1) than 2 × PI-CM35 (+0.17 Wm-2∘C-1). This is somewhat offset by stronger clear-sky shortwave radiative feedbacks in 2 × PI-CM35 (+1.00 Wm-2∘C-1) than 2 × PI-CM10 (+0.57 Wm-2∘C-1). This highlights the important role that cloud radiative feedbacks play in determining the climate sensitivity.

An increase in ECS appears to be robust across the HadCM3-BL experiments. For As-CM, ECS is 0.8 ∘C lower at 35 % O2 than 10 % O2 (Fig. ). For Ma-CM, this value is much larger. A 3.3 ∘C reduction in ECS is simulated, which is also driven by the longwave cloud radiative effects in conjunction with a weaker clear-sky shortwave radiative effect which tended to cool the low pO2 (Fig. ). Unlike the clear-sky shortwave effects, the longwave cloud radiative effects seem consistent across the three experiments. It should be noted that attempts were made to simulate 2× experiments for the Wuchiapingian; however, what would have been the 2 × Wu-CM10, in the nomenclature used here, was numerically unstable.

The increase in climate sensitivity appears to be linked to the reduction in temperature anomaly in a warmer climate state. We propose that this is due to more vigorous convection at low pO2 , leading to an atmospheric moistening (Fig. ) which causes warming analogously to . This is consistent with the increases in climate sensitivity observed – in a warmer climate the atmosphere can hold more water vapour, so any changes to water vapour will be amplified in their impacts on the radiative budget of the atmosphere. This water vapour feedback is also consistent with the weaker clear-sky shortwave radiative effect observed in 2 × Ma-CM and the temperature response observed in the Wuchiapingian simulations.

Changes in pO2 and surface temperatures have the potential to impact the terrestrial carbon cycle by altering the competition between the oxidative and photosynthetic metabolic pathways for RuBisCO. simulated significant changes to vegetation productivity in the Permian due to changes in oxygen content. The modelled changes to vegetation in the final 50 years of the Asselian and Wuchiapingian experiments are investigated. Focusing on changes to vegetation across the Permian, the dominant vegetation fractions for As-CM35, As-CM10, Wu-CM35 and Wu-CM10 are shown in Fig. . For low pCO2 in the Asselian, increasing pO2 leads to a reduction in the extent of broadleaf trees and greater proliferation of grasses and shrubs. This would be consistent with increases in photorespiration at high pO2. The reverse is true in the Wuchiapingian simulations with increases in the extent of tropical broadleaf forests. It should be noted that the simulation of plant functional types is carefully tuned to present-day vegetation which was likely considerably different in the past. Therefore, caution should be exercised when extrapolating to past vegetation changes.

Figure a shows the change in net primary productivity (NPP) for As-CM1035 and Wu-CM1035. The Asselian simulations shows a large reduction in net primary productivity (NPP) as pO2 is increased (Fig. a, -59 PgCyr-1), while the reverse is true in the Wuchiapingian simulations (+33 PgCyr-1). At low pCO2, it is expected that competition for RuBisCO will be won out by O2 and therefore that rates of photorespiration should lead to a decline in photosynthesis. This is reflected in the gross primary productivity (GPP, -34 %) and NPP (-52 %) response for the Asselian. During the Wuchiapingian, there may be sufficient CO2 that competition is much less sensitive to the pO2 so changes to NPP are much less significant. In fact, NPP is increased by 14 % (GPP +18 %). Tropical water use efficiency is also higher in Wu-CM35 (Fig. c), which suggests that water economy of plants could alter to adapt to a higher pO2 .

These net primary productivity changes are reflected in the total carbon storage (Fig. b) which is lower as pO2 is increased in As-CM (-338 Pg C) and higher as pO2 is increased in Wu-CM (+379 Pg C). This is dominated by changes in the tropics (in agreement with ), where broadleaf trees cover more area in Wu-CM35. Cooler terrestrial tropical temperatures, particularly in the warm months (Fig. e), reduce the pO2 inhibition of RuBisCO and reduce the rate of respiration by vegetation and soils .

As these simulations are fully coupled and changes to oxygen content affect temperatures, radiation and precipitation, it is challenging to explore all the possible contributions to differences between these results and the more idealised simulations. However, there is general agreement that changes occur in the signs of the response of NPP and total carbon storage. This supports the conclusions of that high pO2 in the early Permian may have played an important role in the evolution of plants. Note that while the atmosphere and vegetation are coupled in the physical sense, the carbon cycle is not interactive (atmospheric CO2 is fixed), so determining the impacts of atmospheric pO2 on the carbon cycle remains an outstanding problem.