Professional Documents
Culture Documents
https://doi.org/10.1038/s41561-022-00950-1
Geological evidence of active tropical glaciers reaching sea level during the Neoproterozoic (1,000–541 Ma), suggesting a
global ocean completely covered in ice, was the key observation in the development of the hard Snowball Earth hypothesis.
These conditions are hard to reconcile with the survival of complex marine life through Snowball Earth glaciations, which led to
alternative waterbelt scenarios where a large-scale refugium was present in the form of a narrow ice-free strip in the tropical
ocean. Here we assess whether a waterbelt scenario maintained by snow-free dark sea ice at low latitudes is plausible using
simulations from two climate models run with a variety of cloud treatments in combination with an energy-balance model. Our
simulations show that waterbelt states are not a robust and naturally emerging feature of Neoproterozoic climate. Intense
shortwave reflection by mixed-phase clouds, in addition to a low albedo of bare sea ice, is needed for geologically relevant
waterbelt states. Given the large uncertainty in mixed-phase clouds and their interaction with radiation, our results strongly
question the idea that waterbelt scenarios can explain the Neoproterozoic geology. Hence, Neoproterozoic life has probably
faced the harsh conditions of a hard Snowball Earth.
P
alaeomagnetic evidence indicates that during the subsidence of the Hadley circulation suppressed precipitation and
Neoproterozoic, large fractions of Earth’s surface were cov- promoted evaporation of high-albedo snow deposited on sea ice11.
ered by ice, allowing tropical land glaciers to extend down Hence, subtropical sea ice was snow free and relatively dark, which
to sea level1–5. Geochronology indicates that rapid glaciation and weakened the ice–albedo feedback when sea ice entered the sub-
deglaciation, marked by characteristic cap carbonate layers, hap- tropics11. The weakening of the ice–albedo feedback was so drastic
pened synchronously across palaeo-regions, pointing towards a as to halt the runaway, and thus the climate did not rush into a hard
global event5. Neoproterozoic geology is traditionally explained by snowball but stabilized in a waterbelt state11. Under near-global ice
the ‘hard’ Snowball Earth hypothesis, which proposes fully glaci- cover, silicate weathering was largely shut down, and CO2 released
ated oceans on the basis of a runaway ice–albedo feedback and an by volcanic eruptions accumulated in the atmosphere11. Due to high
entire life cycle of the glaciations1,6. Corresponding long-lasting planetary albedo, Earth’s climate stayed within the waterbelt state
global ice cover implies harsh evolutionary conditions. Previous until the greenhouse effect overcame the strong ice–albedo feed-
work has suggested potential habitats for advanced marine species back11. Hence, hysteresis between waterbelt and temperate climate
that require liquid near-surface water5,7–10. Yet, especially explaining states occurred over a considerable range of atmospheric CO2 con-
the persistence of sponges through the Neoproterozoic glaciations centrations, which is consistent with the deposition of characteristic
remains a challenge5. cap carbonate layers during the termination of the glaciations11.
Alternative scenarios assume climate states with tropical ice-free To be geologically relevant, waterbelt states need to exist over a
oceans, which provide more hospitable conditions for life11–13. Ref. 13 substantial range of CO2 concentrations and be accessible from the
suggested a soft Snowball Earth regime with an ice edge around 25° temperate climate. The latter condition requires waterbelt states to
latitude, but modelling studies indicate that tropical land glaciers are exist at a lower CO2-radiative forcing than the temperate climate.
unable to form in this scenario14. By contrast, waterbelt climates with Ref. 11 found geologically relevant waterbelt states in idealized aqua-
a narrow strip of open tropical ocean and an ice edge near 10° latitude planet simulations with two general circulation models (GCMs)
are consistent with Neoproterozoic geology. Waterbelt states were pos- and tied the existence of stable waterbelt states to the presence of
tulated to exist thanks to the convergence of ocean heat transport12 and dark subtropical sea ice and well-established atmospheric physics.
the albedo contrast between high-albedo extratropical and low-albedo Furthermore, ref. 11 and subsequent studies15–17 discussed poten-
subtropical sea ice in the so-called Jormungand hypothesis11. tial factors that might promote or counteract geologically relevant
The Jormungand hypothesis is particularly attractive as it waterbelt states. However, the core of the Jormungand hypothesis
also proposes an entire life cycle of the glaciations and rests on itself, that is, the robustness of the stabilizing impact of dark sub-
well-understood atmospheric physics, that is, the Hadley circula- tropical sea ice, has not been questioned. We here investigate the
tion and corresponding surface net evaporation in the subtropics. robustness of the Jormungand hypothesis in idealized aquaplanet
Similar to the Snowball Earth hypothesis, the Jormungand hypoth- simulations to assess whether the Jormungand hypothesis indeed
esis suggests that enhanced continental weathering decreased qualifies as a viable alternative to the Snowball Earth hypothesis.
pre-glacial CO2 and drove Earth towards a colder climate with
increasing ice cover and strengthening ice–albedo feedback5,11. After Waterbelt states are not a robust feature of climate
reaching a critical extent of global ice cover, a bifurcation occurred We perform simulations with two GCMs with idealized
leading to a runaway ice–albedo feedback11. In the subtropics, Neoproterozoic boundary conditions: the Community Atmosphere
Institute of Meteorology and Climate Research, Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany. 2Department of Meteorology and
1
a b c
90 90 90
Default
pCOOKIE
60 60 60
45 45 45
Ice-edge latitude (°)
20 2 20 20
10 1
10 10
0 0 0
2,000 5,000 7,500 11,250 15,000 1,500 3,000 4,000 5,000 5,000 6,000 7,000 8,000 9,000
Atmospheric CO2 (ppmv) Atmospheric CO2 (ppmv) Atmospheric CO2 (ppmv)
Fig. 1 | Low-latitude clouds control the existence and absence of waterbelt states in the GCMs CAM and ICON. a–c, Bifurcation diagrams of global-mean ice
edge versus atmospheric CO2 concentrations for CAM in its default configuration and with clouds made transparent to radiation in a narrow tropical region
(pCOOKIE) (a), the default configuration of ICON (b) and ICON WBF (c). Filled symbols show stable states. Circles show simulations initialized from ice-free
conditions, squares show simulations initialized from stable waterbelt states and diamonds represent simulations initialized from transient waterbelt states.
Unfilled diamonds mark slowly drifting simulations that remain in a waterbelt-like state for at least 40 years, with arrows indicating the drift of the ice edge.
Lines are drawn as best guesses of equilibrium states, with solid lines indicating stable and dashed lines indicating unstable states. T , W 1 and W 2 here label
the bifurcation points (nose points) corresponding to the CO2 thresholds referred to in the text. Bifurcation points mark the unstable transitions between
temperate and waterbelt/snowball climate ( T ), between waterbelt and snowball climate ( W 1) and between waterbelt and temperate climate ( W 2).
Model v.3.1 (CAM)18 and the Icosahedral Nonhydrostatic atmo- The most pronounced difference arises from low-level clouds in the
sphere GCM (ICON)19. We configure the models as in ref. 11 using subtropics and tropics (Fig. 2d,e).
an idealized aquaplanet setup with a 50-m-deep mixed-layer ocean, The differences in cloud cover do not arise from differences in
thermodynamic sea ice and a reduced solar constant of 1,285 W m−2 the large-scale atmospheric state. The vertical and zonal structure
(Methods). We conduct simulations for a range of atmospheric CO2 of air temperature as characterized by the 273, 235 and 192 K iso-
concentrations and initial sea-ice cover (Extended Data Fig. 1) to therms is similar in both GCMs (Fig. 2d,e) as is the Hadley circula-
compute each GCM’s bifurcation diagram. tion (Extended Data Figs. 3 and 4). As the 273 and 235 K isotherms
CAM exhibits three stable climate regimes (Fig. 1a): a temper- mark the limits of pure liquid and pure ice clouds20, the low-level
ate regime with no or little ice cover, the waterbelt regime with an clouds in the subtropics and tropics are mixed-phase clouds.
ice edge between 9° and 15° latitude and the snowball regime with The treatment of mixed-phase clouds is highly uncertain
global ice cover. Our simulations reproduce the findings of ref. 11. in GCMs21 and is based on various microphysical modelling
There are three relevant bifurcation points exhibiting rapid climate approaches. The main difference between CAM and ICON arises
transitions. (1) At 3,000 ppmv CO2, the temperate climate transi- from liquid–ice partitioning (LIP): CAM applies explicit LIP based
tions into the waterbelt climate. We mark this CO2 threshold by on air temperature, while in ICON, LIP results from a comprehen-
the symbol T . (2) The transition from waterbelt to snowball cli- sive set of conversion rates between liquid and ice phases (Methods).
mate occurs at W 1 = 1, 750 ppmv CO2 (ref. 11). (3) The transition The different treatment of cloud micropyhsics has a clear impact
from waterbelt to temperate climate occurs at W 2 = 15, 000 ppmv on the simulated liquid condensate fraction (LCF). Sampling clouds
CO2 (ref. 11). Thus, the waterbelt regime exists over W 2 − W 1 ≈ by air temperature and LCF shows that ICON generates relatively
13,000 ppmv CO2 and is easily accessible from the temperate cli- more cloud ice at higher temperatures than CAM (Fig. 2f) and indi-
mate because W 1 < T . cates that the liquid-to-ice conversion is more efficient in ICON.
ICON yields a strikingly different bifurcation diagram (Fig. 1b), A lower LCF leads to optically thinner clouds because cloud ice
which questions the robustness of the Jormungand-waterbelt sce- crystals are typically larger than cloud liquid droplets and works in
nario. In ICON, waterbelt states exist only over a narrow range favour of less cloud cover since ice crystals settle more efficiently
of CO2. Even more important, in ICON, waterbelt states are not than smaller cloud droplets22. Hence, the more icy mixed-phase
accessible from the temperate state because T = 1,594 ppmv CO2 clouds in ICON explain the considerably lower planetary albedo
is smaller than W 1 ≈ 4,000 ppmv CO2. Thus, although waterbelt over ice-free ocean compared with CAM.
states exist, they are geologically irrelevant. This result is robust to The temperature at which LCF crosses a value of 0.5 in our
applying different sea-ice models (Methods). simulations is well within the large temperature range simu-
lated by 26 state-of-the-art GCMs and the observational range21
Cloud radiation and microphysics control waterbelt states (Fig. 2f). Furthermore, there is considerable uncertainty arising
CAM and ICON both exhibit bare sea ice in the subtropical region from Neoproterozoic atmospheric boundary conditions, including
and thus low reflection at the surface and in clear-sky conditions aerosol amount and composition23,24, which may strongly impact
at the top of the atmosphere, consistent with the Jormungand mixed-phase cloud physics22, as well as the applied convection
hypothesis (Fig. 2a and Extended Data Fig. 2). Yet, planetary albedo schemes that were developed for present-day climate. Although the
α is much lower in ICON (Fig. 2a) because ICON simulates fewer treatment of mixed-phase clouds is more comprehensive in ICON,
clouds than CAM and, as a consequence, a weaker shortwave these uncertainties mean that we cannot judge whether the clouds
cloud-radiative effect (SWCRE) over ice-free ocean (Fig. 2b,c). in CAM or ICON are more realistic.
90
0.6
200 200
α
0.4 75
CAM
0.2
ICON
400 400 60
Pressure (hPa)
ICON WBF
30
Total cloud cover (%)
80
800 800
15
60
40 0
60° S 20° S 20° N 60° N 60° S 20° S 20° N 60° N
Latitude Latitude
60° S 20° S 20° N 60° N
Latitude
c f
100 1.0 CMIP5 models
80 0.8 Observations
SWCRE (W m–2)
60 0.6
LCF
40 0.4
20 0.2
0 0
60° S 20° S 20° N 60° N 210 220 230 240 250 260 270 280
Latitude Air temperature (K)
Fig. 2 | Differences in clouds and their SWCRE as obtained from the GCMs CAM and ICON as well as ICON WBF. a–c, Zonal-mean annual-mean
planetary albedo α for all-sky (solid) and clear-sky (dashed) conditions (a), total cloud cover (b) and SWCRE (c) averaged over all stable waterbelt states
for each model. The grey band in a indicates the range of global-mean ice-edge latitudes for all stable equilibrium states found in CAM, ICON and ICON
WBF simulations. d,e Zonal-mean cloud cover together with 273 K, 235 K and 192 K isotherms in CAM (d) and ICON (e) for simulations with comparable
global-mean ice cover (10,000 ppmv CO2 in CAM; 4,063 ppmv CO2 in ICON). The orange dotted box in d shows region of CAM pCOOKIE modification for
the Northern Hemisphere. Purple contours in e show the cloud-cover difference between ICON WBF (at 6,000 ppmv CO2) and ICON (contour interval of
3%; positive differences in solid). f, LCF from simulations shown in d and e. The range of temperatures for which liquid and ice are equally prevalent for 26
Coupled Model Intercomparison Project phase 5 (CMIP5) GCMs is shown in grey. The red line shows the combined observational range of ground-based
LIDAR and aircraft measurements21.
We demonstrate the sensitivity of waterbelt states to clouds by version in mixed-phase clouds strongly depends on the Wegener–
running additional simulations in which we manipulate cloud radi- Bergeron–Findeisen process (WBF)26. The WBF occurs as the
ation and microphysical processes (Extended Data Fig. 5). saturation pressure over water is greater than the saturation pressure
First, we locally disable cloud-radiative effects (CRE) in CAM over ice, which allows ice crystals to grow at the expense of cloud
partial Clouds On-Off Klimate Intercomparison Experiment25 droplets in mixed-phase clouds. We weaken the WBF in ICON by
(pCOOKIE) simulations (Methods and Extended Data Fig. 6). We modifying a loosely constrained tuning parameter within plausible
disable CRE in the region that captures the strong differences in ranges27 (Methods).
low-latitude low-level clouds between CAM and ICON in each hemi- Weakening the WBF drastically increases cloud liquid, that
sphere (orange box in Fig. 2d). Yet, the impact on the global energy is, LCF, SWCRE and planetary albedo (Fig. 2 and Extended Data
balance is small. Therefore, T is only marginally affected (Fig. 1a). Fig. 7), and results in a strong cooling effect. As a consequence,
When low-level tropical clouds are made transparent in CAM, the higher atmospheric CO2 concentrations are required to maintain a
CO2 range over which the waterbelt climate exists shrinks from 13,000 temperate climate, and T increases to 6,250 ppmv (Fig. 1c). More
to approximately 5,000 ppmv (Fig. 1a). Moreover, the CO2 required to important, however, reduced liquid-to-ice conversion via the WBF
initiate a snowball from a waterbelt climate increases to W 1 = 2,375 allows ICON to exhibit waterbelt climates that are accessible and
ppmv. As a result, the waterbelt state is almost unaccessible, and CAM exist over a substantial range of CO2, namely, from W 1 = 5,250
with modified clouds approaches the behaviour of ICON in terms of ppmv to W 2 = 8,500 ppmv (Fig. 1c).
the geological relevance of the waterbelt climate regime. In summary, our GCM simulations clearly show that tropical
Second, we reduce the efficiency of liquid-to-ice conversion in clouds and their impact on planetary albedo can determine the geo-
ICON to increase LCF to a value close to CAM. Liquid-to-ice con- logical relevance of the waterbelt climate regime.
a b c
90
3.0 No hyteresis of waterbelt regime
0.75
Temperate regime
60
60
45
Ice-edge latitude (°)
le
la e
λ (K (W m–2)–1)
0.50
ib
t p gim
us
no re
45
lly elt
30 Unstable
ica rb
αo
ys ate
ph W
20 30
Waterbelt regime
0.25
Fig. 3 | Analysis of the waterbelt regime in a one-dimensional EBM. a, Bifurcation diagrams of ice-edge latitude versus CO2-radiative forcing F calculated
from the EBM with parameters estimated from the GCMs. b, Climate-sensitivity parameter λ determined from the EBM as a function of the ratio of
planetary albedo over ice-free ocean αo to bare sea-ice albedo αi,b and ice-edge latitude. c, Map in αo–αi,b space showing domains of stable and accessible
waterbelt states (white) and stable but unaccessible waterbelt states (light grey). The black dashed lines indicate the margins within which the waterbelt
regime can exist. The red dotted box indicates the range of plausible values for αo and αi,b. The black circle marks albedo values for CAM.
Clouds mask the ice–albedo feedback The combination of GCM simulations with the EBM analysis
To understand how low-latitude clouds shape the waterbelt states distils an important mechanism by which low-latitude clouds con-
via their impact on planetary albedo, we apply a one-dimensional trol the width and accessibility of the waterbelt regime. If clouds are
energy-balance model (EBM). The EBM balances absorbed solar abundant and cause a high albedo over ice-free ocean, the waterbelt
radiation with outgoing longwave radiation L and meridional heat regime is wide and easily accessible. By contrast, if clouds are scarce,
transport by considering local insolation, albedo and surface tem- the waterbelt regime is narrow and potentially inaccessible. This
perature at each latitude ϕ11,28,29. The EBM considers two opposing shows that the stark differences between CAM and ICON are to a
climate feedbacks: the negative longwave feedback, which is con- large extent related to their strong differences in low-latitude clouds,
sidered to be constant, and the positive ice–albedo feedback. The of which there are many in CAM and few in ICON.
ice–albedo feedback depends on the ice-edge latitude because ice
albedo is high for snow-covered ice in the extratropics αi,s = 0.8 and Diminished geological relevance of waterbelt states
low for bare ice at low latitudes αi,b = 0.45 (ref. 11). Planetary albedo We further use the EBM to assess whether a waterbelt scenario is a
over ice-free ocean αo is homogeneous and is dominated by short- likely explanation for Neoproterozoic geology. The waterbelt regime
wave reflection by clouds (Fig. 2a). Hence, the ice–albedo feedback can exist over a wide range of αi,b and αo (enclosed by black dashed
in the waterbelt regime is determined by the ratio αo/αi,b. lines in Fig. 3c). Yet, its accessibility requires αo to exceed a thresh-
The EBM captures the qualitative differences among GCM old, which increases with increasing αi,b (white domain in Fig. 3c). If
simulations (Fig. 3a). We estimate the EBM parameters from GCM αo is too small, the subtropical ice–albedo feedback is too strong to
simulations for the longwave feedback, αo and meridional heat escape the runaway feedback towards a hard snowball state.
transport (Methods). If the EBM is driven with ICON parameters, The albedo values for CAM are located within the domain of
the waterbelt regime is not accessible and has substantially weaker accessible waterbelt climate states (black circle in Fig. 3c). Yet,
CO2-hysteresis compared with CAM parameters. If driven by a subtle decrease of αo makes the waterbelt regime unaccessible
CAM pCOOKIE instead of CAM parameters, the hysteresis of the (Extended Data Fig. 8), comparably to our CAM pCOOKIE simula-
waterbelt regime shrinks, and the waterbelt state becomes inacces- tions with modified clouds. Furthermore, the EBM analysis indi-
sible from the temperate climate. We do not analyse the EBM with cates that a small increase of αi,b would also make the waterbelt
ICON with the less-efficient WBF process (ICON WBF) parameters regime in CAM unaccessible.
because it fails to satisfy the requirement of a constant longwave We define a domain of physically plausible albedo values on the
feedback (Extended Data Table 1). basis of the range of αo found in our GCM simulations and estimates
Using the EBM, we calculate the climate-sensitivity parameter for αi,b that range from 0.42 to 0.68 (ref. 31) (red dotted rectangle
λ for equilibrium climate states (Methods). We limit our analysis in Fig. 3c). The EBM predicts stable waterbelt states for the entire
to αo < αi,b and use αi,b = 0.45. The EBM exhibits stable states with plausible domain. However, the overlap of the plausible domain and
positive finite λ for the temperate, waterbelt and hard snowball accessible waterbelt states is small and confined to low plausible αi,b.
regimes (not shown) and unstable climate regimes with negative λ In fact, the geologically relevant waterbelt regime simulated by CAM
in between (Fig. 3b). results from the combination of rather extreme values for αo and αi,b.
Larger αo allows waterbelt states to extend to lower latitudes Hence, a geologically relevant waterbelt regime is physically possible,
because the ice–albedo feedback becomes weaker as αo is increased. but its viability is limited to a narrow range of key parameters.
At higher αo, changes in ice cover have a smaller radiative impact at
the top of the atmosphere. A similar behaviour is observed for the Conclusions
cloud impact on the radiation budget over the present-day Arctic30. Waterbelt states were proposed as alternative solutions for the
As a result, high cloud contribution to planetary albedo supports pan-glaciations of the Neoproterozoic that, in contrast to a hard
stable waterbelt states by muting the shortwave impact of changes Snowball Earth, can easily explain the survival of life by provid-
in ice cover. ing a narrow strip of ice-free tropical ocean11. Waterbelt states were
Methods forcing and therefore shift the bifurcation diagram to lower CO2 concentra-
GCM simulations. We apply the National Center for Atmospheric Research’s tions. Besides the shift of the bifurcation diagram, the additional forcing is
CAM18 as described in ref. 11 and the Max Planck Institute’s ICON with climate not expected to alter the structure of the bifurcation diagram and the stability
physics (version AES 1.3.00) as described in ref. 19 with modifications as given of the waterbelt state. The ozone concentration applied in ICON was derived
in the following. We apply a circular Kepler orbit with zero eccentricity and an from the ozone climatology applied in CAM by calculating the annual mean
obliquity of 23.5° for both GCMs, following ref. 11. We run CAM in resolution and symmetrizing across the Equator34. Therefore, the ozone concentrations
T42 (~310 km nominal grid spacing) and ICON in resolution R2B4 (~160 km in CAM and ICON exhibit similar magnitudes and are not expected to lead
nominal grid spacing). For our analysis, we use monthly mean zonal-mean profiles to qualitative differences in the bifurcation diagrams.
with a latitudinal resolution of 2.8125° for CAM and 1.875° for ICON. We regrid (3) The fundamental difference between the sea-ice schemes applied in CAM and
three-dimensional output from ICON to the same 26 pressure levels between ICON is the vertical resolution of sea ice. ICON applies a zero-layer scheme37,
993 hPa and 3 hPa used for the CAM output. whereas CAM applies a four-layer scheme39. The multi-layer sea-ice scheme
We determine the global-mean ice-edge latitude as the sine of the global applied in CAM keeps track of ice internal energy and therefore buffers sum-
ice-free surface fraction. mer and daytime melting32, which facilitates snowball initiation40. We confirm
this by conducting ICON simulations with a three-layer sea-ice scheme41.
Using the three-layer sea-ice scheme in ICON drift more readily initiates a
ICON. We modify ICON compared with the version described in ref. 19 to
snowball climate compared with using the zero-layer sea-ice scheme. There-
simulate an aquaplanet climate with no continents and with a thermodynamic
fore, we apply the zero-layer scheme in ICON because it facilitates the simula-
mixed-layer ocean of 50 m depth without ocean heat transport. We implement a
tion of stable waterbelt states.
thermodynamic sea-ice scheme that is described in more detail in the following.
We do not consider any greenhouse gases except CO2 (with a constant and uniform
concentration) and water vapour. Direct radiative effects of aerosols are set to CAM pCOOKIE simulations. Our CAM pCOOKIE simulations are inspired by
zero. The ozone concentration is constant in time and taken from the aquaplanet the Clouds On-Off Klimate Intercomparison Experiment25. We locally set cloud
simulations of ref. 34. ICON keeps track of snow deposited on sea ice, which allows cover to zero when calling the radiation scheme. Specifically, we do this in the
us to interactively simulate the strong albedo gradient arising from the transition lower troposphere in the tropical region around 12° latitude as indicated for the
of snow-covered to bare sea ice in the subtropical region. The cold albedo (at a Northern Hemisphere by the orange box in Fig. 2d. This intervention is limited
surface temperature T ≤ −1 °C) of bare sea ice and snow is set to 0.45 and 0.79, to the radiation scheme, and clouds, latent heating and precipitation still form as
respectively. The warm albedo (T = 0 °C) of bare sea ice and snow is set to 0.38 usual. Furthermore, the modification mainly affects SWCRE since it is applied to
and 0.66, respectively. For the temperature range between –1 °C and 0 °C, albedo is low-level clouds (Extended Data Fig. 6). We find a qualitatively similar behaviour if
linearly interpolated between the cold and warm values. This is an approximation we disable CRE in other regions (not shown).
of the surface albedo applied in CAM.
Because we simulate climates much colder than in ref. 19, we decrease the Cloud microphysical schemes in CAM3 and ICON. CAM3. CAM applies a
model height from 83 km to 72 km by removing the two top levels (from originally one-moment cloud microphysical scheme based on ref. 42. Liquid–ice partitioning
47). This helps to avoid model instabilities in the lower stratosphere. Along with is performed explicitly on the basis of air temperature Ta at every microphysical
reducing the levels, we adjust the stretch factor for the vertical distribution of model time step18. xi indicates the ice fraction within a cloud within the
levels from 0.900 to 0.949 to maintain a distribution similar to the the setup with temperature range Ta,min = –40 °C to Ta,max = –10 °C.
47 levels. However, some instabilities remain under high CO2 concentrations and
large ice cover. Therefore, if global ice cover exceeds about 20% in the simulations, Ta − Ta,max
xi = , Ta,min ≤ Ta ≤ Ta,max (1)
we also increase the Rayleigh damping of vertical velocity near the model top and Ta,min − Ta,max
decrease the model time step35,36. This stabilizes the model, and we have verified
that these modifications have a negligible impact on the simulated climates. xi = 1 for Ta < Ta,min and xi = 0 for Ta > Ta,max. Cloud liquid l and cloud ice I within each
Specifically, starting from ice-free states, we conduct simulations with a time grid cell are calculated on the basis of xi from the previous total cloud condensate
step of 10 min and a Rayleigh damping coefficient of 0.1. Subsequently, the model (l′ + I′ ).
time step is reduced to 8 min and the Rayleigh damping coefficient is increased
l = (l′ + I′ )(1 − xi ) (2)
to 10 after the model first becomes unstable. If further instabilities occur, the
time step is reduced to 6 min, and we have not seen further model crashes then.
The ICON simulation at 1,500 ppmv CO2 that is initiated from an ice-free state I = (l′ + I′ )xi (3)
runs into a snowball without any instabilities with an 8 min time step. Yet, for
consistency, we conduct the branched-off simulations with increased atmospheric After the explicit partitioning of cloud condensate, conversion rates from cloud
CO2 with a time step of 6 min. Thus, all simulations in the temperate climate liquid to rain and from cloud ice to snow are calculated.
regime use a time step of 10 min and a Rayleigh damping coefficient of 0.1, and all
simulations in the waterbelt climate regime use a time step of 6 min and a Rayleigh
damping coefficient of 10 (including the simulations with altered WBF). ICON. ICON applies a one-moment cloud microphysical scheme based on ref. 43.
As part of our implementation of the mixed-layer ocean, we implement a Liquid-to-ice conversion is modelled by conversion rates that depend on three
zero-layer Semtner sea-ice scheme37. Sea ice forms if T < Tf = −1.9 °C, that is, the different temperature regimes. First, homogeneous freezing of cloud liquid occurs
freezing point of water at a salinity of 35 psu (ref. 38). The transformation of ice-free instantly below air temperatures of −35 °C. Second, complete melting of cloud ice
ocean to sea ice requires the formation of a sea-ice layer of a minimum thickness occurs above air temperatures of 0 °C. Third, between −35 °C and 0 °C, stochastic
hi,min = 0.05 m. If T < Tf but is still too high to form an ice layer of thickness hi,min, and heterogeneous freezing occurs.
T is set to Tf and a residual heat flux corresponding to the difference T − Tf is Furthermore, conversion from the vapour phase directly to the ice phase by
calculated. The residual flux is added to the energy balance of the mixed-layer deposition is considered in addition to condensation if air temperatures either
ocean during the next model time step. Dynamics of sea ice are not considered. are below − 35 °C or are below 0 °C and the ratio of cloud ice to cloud liquid xi
exceeds a threshold xi,thr. This can be considered as an idealized parameterization
of the WBF process. A higher xi,thr corresponds to a weaker WBF and favours the
Differences between CAM and ICON simulations due to initialization, boundary existence of cloud liquid because more condensation instead of deposition takes
conditions and the applied sea-ice scheme. Although using the same idealized place. The value of xi,thr is a loosely constrained tuning parameter27. We weaken the
aquaplanet setup, CAM and ICON simulations exhibit differences in (1) the WBF by increasing xi,thr by a factor of ten to the upper limit of the plausible range27.
initial and (2) boundary conditions as well as (3) the applied sea-ice scheme. In all
three aspects, there is no qualitative impact on our key results, as outlined in the
Determination of bifurcation points. The thermodynamic mixed-layer ocean
following.
limits the timescales to the order of decades to a century. Therefore, we here
(1) The initial conditions differ in terms of sea surface temperature (SST) pat- consider climate states to be stable if they stay within a climate regime for at least
terns between CAM and ICON. In CAM, homogeneous SSTs of 300 K are 100 years after initialization without considerable drift of the ice edge.
used while in ICON a zonally symmetric SST profile reaching from 273.5 K We define the atmospheric CO2 concentration of bifurcation points to be in
at the poles to 302.5 K at the Equator is applied. We conducted ICON simula- the middle between the last simulated stable climate state and the next simulated
tions (at 1,500 ppmv CO2 and 3,000 ppmv CO2) with initially homogeneous climate state that transitions to a colder (warmer) climate regime for decreasing
SSTs of 300 K and found no difference in the equilibrium state that ICON (increasing) CO2. All our results and conclusions are independent of this choice.
drifts to. In fact, the large number of simulations that we conduct near the bifurcation
(2) Boundary conditions between ICON and CAM differ in the amount points for all model setups allows us to pinpoint the critical CO2 values within
of well-mixed greenhouse gases and the ozone climatology. ICON uses about 100 ppmv.
no well-mixed greenhouse gases except CO2. CAM uses the well-mixed
greenhouse gases CH4, N2O, CFC11 and CFC12 with the volume mixing ratios Calculation of LCF. We calculate LCF for 40 yr periods of single simulations for
defined in ref. 4. Additional greenhouse gases exert an additional radiative CAM (10,000 ppmv CO2), ICON (4,063 ppmv CO2) and ICON WBF (6,000 ppmv
dT ∂T dT ∑ Data availability
λ= = + fi (8) The data and corresponding run scripts that support the findings of this study45 are
dF ∂F dF i
available at https://doi.org/10.5445/IR/1000144276.
In the EBM (equation (4)) the ice–albedo feedback is the only feedback besides
the Planck feedback. The sensitivity parameter corresponding to the ice–albedo Code availability
feedback is The custom computer code used to generate the results45 is available at https://doi.
org/10.5445/IR/1000144276.
∂T dα
fα = . (9)
∂α dT
References
Using the global mean of equation (4), fα is given by 34. Liang, X.-Z., Wang, W.-C. & Boyle, J. S. Atmospheric Ozone Climatology for
Use in General Circulation Models (Lawrence Livermore National Laboratory,
I dα 1997).
fα = − . (10)
B dT 35. Klemp, J., Dudhia, J. & Hassiotis, A. An upper gravity-wave absorbing layer
for NWP applications. Mon. Weather Rev. 136, 3987–4004 (2008).
After rearrangement, we find 36. Zängl, G., Reinert, D., Rípodas, P. & Baldauf, M. The ICON (ICOsahedral
Non-hydrostatic) modelling framework of DWD and MPI-M: description
1 1 of the non-hydrostatic dynamical core. Q. J. R. Meteorol. Soc. 141,
λ= . (11)
B 1 − fα 563–579 (2015).
37. Semtner, A. J. A model for the thermodynamic growth of sea ice in
numerical investigations of climate. J. Phys. Oceanogr. 6, 379–389
Estimation of parameters for EBM from the GCM simulations. We here (1976).
estimate the EBM parameters B, αo and C corresponding to our GCM simulations 38. Fujino, K., Lewis, E. L. & Perkin, R. G. The freezing point of seawater at
with CAM, ICON, CAM pCOOKIE and ICON WBF (only B). All other EBM pressures up to 100 bars. J. Geophys. Res. 79, 1792–1797 (1974).
parameters are chosen as in ref. 11. 39. Bitz, C. M. & Lipscomb, W. H. An energy-conserving thermodynamic model
of sea ice. J. Geophys. Res. Oceans 104, 15669–15677 (1999).
Longwave feedback parameters. We derive estimates for the longwave feedback 40. Yang, J., Peltier, W. R. & Hu, Y. The initiation of modern ‘soft snowball’ and
B for each GCM configuration by a linear regression of monthly mean L on ‘hard snowball’ climates in CCSM3. Part I: The influences of solar luminosity,
monthly mean T using a least square fit. We estimate B for the temperature CO2 concentration, and the sea ice/snow albedo parameterization. J. Clim. 25,
range spanned by all simulations indicated in Extended Data Figs. 1 and 5. We 2711–2736 (2012).
Extended Data Fig. 1 | Temporal evolution of the global-mean ice edge latitude for the default GCM setups. (a) CAM and (b) ICON. Global-mean
ice edge latitude is calculated as the sine of global-mean ice-free surface fraction. Labels indicate the constant atmospheric CO2 concentration and the
global-mean ice-edge latitude of the initialisation. Thin lines represent the 10-year rolling mean. Solid lines indicate simulations considered to exhibit
(semi-)stable waterbelt states. Thick solid lines represent monthly-mean for 40-year periods that are analyzed as (semi-)stable waterbelt states.
Extended Data Fig. 2 | Low clear-sky planetary albedo due to exposure of bare sea ice in the subtropical region in CAM and ICON. Climatological
zonal-mean clear-sky planetary albedo determined from (a) CAM simulation at 10000 ppmv CO2 and (b) ICON simulation at 4063 ppmv CO2.
Dashed contours indicate the water-equivalent snow thickness of 0.01 m, which marks the minimum thickness of a snow-layer to be considered in the
determination of the surface albedo in ICON. For each simulation 40-year periods are analyzed. Further details are given in text S1.
Extended Data Fig. 3 | CAM and ICON exhibit similar large-scale thermal structure across the seasonal cycle and the simulated CO2- range. Monthly
climatological zonal-mean cloud cover and air temperature indicated by isotherms determined from CAM at 2000 ppmv CO2 and 10000 ppmv CO2, and
ICON at 4063 ppmv CO2. Each period of analysis spans 40 years after stabilizing in a stable waterbelt state.
Extended Data Fig. 4 | CAM and ICON exhibit similar large-scale circulation in CAM and ICON across the seasonal cycle and the simulated CO2- range.
Monthly climatological zonal-mean cloud cover and mass stream function indicated by contours. Contour intervals are 5 ⋅ 1010 kgs−1. Positive contours
∫ ps
are solid and negative contours are dashed. The zero contour is not shown. The mass stream function is calculated by ψ(p, ϕ) = 2ΠrE cos(ϕ)/g vdp
p
with pressure p, surface pressure ps, latitude ϕ, radius of Earth rE = 6371 km, gravitational constant g = 9.81 ms−2, and zonal-mean meridional wind v. Same
simulations and time periods as in Extended Data Figure 3 are used.
Extended Data Fig. 5 | Temporal evolution of the global-mean ice edge latitude for the modified GCM setups. (a) CAM pCOOKIE and (b) ICON WBF. See
caption of Extended Data Figure 1 for further details.
Extended Data Fig. 6 | Impact of pCOOKIE modification in CAM on clouds and cloud-radiative effects. (a) Planetary albedo α, (b) total cloud cover,
(c) shortwave CRE, and (e) longwave CRE for CAM standard and pCOOKIE. Shown are 40-year-means over all simulations after stabilizing in a stable
waterbelt state. (d) 40-year-mean of zonal-mean cloud cover from a single simulation with CAM pCOOKIE after stabilizing in a waterbelt state at
5000 ppmv CO2 along with 273 K, 235 K, and 192 K isotherms.
Extended Data Fig. 7 | Impact of WBF modification in ICON on cloud liquid and longwave cloud-radiative effect. (a) and (b): Annual-mean zonal-mean
specific cloud liquid from single simulations after stabilizing in a waterbelt state at 4063 ppmv CO2 (ICON) and 6000 ppmv CO2 (ICON WBF) along
with 273 K, 235 K, and 192 K isotherms. (c) Annual-mean zonal-mean liquid water path LWP, and (d) longwave CRE averaged over all simulations with a
(semi-)stable waterbelt state. The period of analysis spans 40 years after stabilizing in a (semi-)stable waterbelt state.
Extended Data Fig. 8 | The geologically relevant domain of waterbelt states is impacted by the intensity of meridional heat transport. Domains spanned
by planetary albedo over icefree ocean αo and bare sea ice albedo αi,b for stable and accessible waterbelt impacted by the intensity of meridional heat
transport. states (white) and stable but unaccessible waterbelt states (light gray). The orange circle marks albedo values for CAM pCOOKIE. The orange
line marks the lower boundary of the stable and accessible domain and is calculated with the meridional heat transport parameter C = 1.6B corresponding
to CAM pCOOKIE. The black dotted line marks the lower boundary of the stable and accessible domain if calculated with C = 1.5B corresponding to CAM
(similar as in Fig. 3c). Weaker meridional heat transport stabilizes the waterbelt climate because it is characterized by strong meridional temperature
gradients. The red dotted box indicates the range of plausible values for αo and αi,b.
Estimated longwave feedback parameter B in units of Wm-2K-1 using the temperature range spanned by all conducted simulations (column 1) and longwave feedback parameter Bw using the temperature
range spanned by simulated stable waterbelt states (column 2). Ratios Bw/B are used as an indicator of the (non-)linearity of the longwave feedback (column 3). See methods for further details.