PersPective
https://doi.org/10.1038/s41558-020-0731-2
1
National Center for Atmospheric Research, Boulder, CO, USA.
2
US CLIVAR Working Group on Large Ensembles, Washington, DC, USA.
3
Center for
Climate Physics, Institute for Basic Science, Busan, South Korea.
4
Pusan National University, Pusan, South Korea.
5
Cornell University, Ithaca, NY, USA.
6
Geophysical Fluid Dynamics Laboratory/NOAA, Princeton, NJ, USA.
7
Institute for Geophysics, University of Texas, Austin, Austin, TX, USA.
8
Lamont
Doherty Earth Observatory of Columbia University, Palisades, NY, USA.
9
Sorbonne University, Paris, France.
10
Canadian Centre for Climate Modelling
and Analysis, Environment and Climate Change Canada, University of Victoria, Victoria, British Columbia, Canada.
11
Department of Earth and Planetary
Sciences, Northwestern University, Evanston, IL, USA.
12
Cooperative Institute for Research in Environmental Sciences, University of Colorado Boulder,
Boulder, CO, USA.
13
Department of Atmospheric and Oceanic Sciences, University of Colorado, Boulder, CO, USA.
14
Institute for Atmospheric and Climate
Science, ETH Zurich, Zurich, Switzerland.
15
Institute of Arctic and Alpine Research, University of Colorado, Boulder, CO, USA.
16
Max Planck Institute for
Meteorology, Hamburg, Germany.
17
Institute of the Environment and Sustainability and Department of Statistics, University of California, Los Angeles, Los
Angeles, CA, USA.
18
Hokkaido University, Sapporo, Hokkaido, Japan.
19
Department of Earth System Science, University of California Irvine, Irvine, CA, USA.
20
University of Exeter, Exeter, UK.
✉
e-mail: cdeser@ucar.edu
I
dentifying anthropogenic influences on weather and climate
amidst the background of internal variability, and providing
robust projections, are central scientific challenges with prac
-
tical implications
1–6
. Since the inception of the Coupled Model
Intercomparison Project (CMIP), substantial progress has been
made on quantifying sources of uncertainty in climate projections
(for examples, see refs.
7–9
). However, such multi-model archives
confound uncertainties arising from differences in model for
-
mulation (that is, structural uncertainty) with those generated
by internal variability (variability from natural processes in the
coupled ocean–atmosphere–land–biosphere–cryosphere system).
This distinction is important because the former is potentially
reducible as models improve, whereas the latter is an intrinsic
property of each model and is largely irreducible after the memory
of initial conditions is lost, typically after less than a few years
over land
10
. This key distinction is often not widely appreciated
and communicated to stakeholder groups
11
. Indeed, internal vari-
ability accounts for approximately half of the inter-model spread
within CMIP for projected changes in near-surface air tempera
-
ture, precipitation and runoff across North America and Europe
over the next 50 years
5,8,9,12–14
.
One way to isolate the uncertainty from internal variability is to
create an ensemble of simulations with a single climate model under
a particular radiative forcing scenario, applying perturbations to the
initial conditions of each member to create diverging weather and
climate trajectories, causing ensemble spread (for examples, see
refs.
12,15–17
). Since the resulting sequences of unpredictable inter-
nal variability are randomly phased between the individual ensem-
ble members, the forced response can be estimated by averaging
over a sufficient number of members. The definition of ‘sufficient’
depends on the quantity of interest, location, spatial scale, temporal
scale and time horizon, often on the order of 10–100 members (for
example, see ref.
12
). Such ‘initial-condition large ensembles (LEs)’
conducted with fully-coupled global models are a relatively new
development in climate science, with the first efforts employing
CMIP3-era models
12,18
.
The past few years have witnessed an explosion of LEs with
newer-generation CMIP5-class Earth system models (ESMs;
Table 1). Each LE required substantial high-performance comput-
ing resources and generated hundreds of terabytes of output. For
example, the CESM1-LE used 21 million CPU hours and produced
over 600 terabytes of model output (for comparison, the entire
CESM1 contribution to CMIP5 was 170 terabytes). Making these
‘big data’ projects accessible to a wide range of users is challenging,
yet their ease-of-use for different types of analysis workflows has a
substantial impact on the scientific value gained from their produc
-
tion. A case in point is the NCAR CESM1-LE Project
19
, which was
created to serve a broad research community by responding to user
needs to provide easy access to the output and stable on-disk access.
This project has resulted in more than 860 peer-reviewed studies to
date, with approximately 400,000 data files downloaded. Remaining
nimble to new workflows and users is important, as is following the
Insights from Earth system model initial-condition
large ensembles and future prospects
C. Deser
1,2
✉
, F. Lehner
1,2
, K. B. Rodgers
2,3,4
, T. Ault
2,5
, T. L. Delworth
2,6
, P. N. DiNezio
2,7
,
A. Fiore
2,8
, C. Frankignoul
2,9
, J. C. Fyfe
2,10
, D. E. Horton
2,11
, J. E. Kay
2,12,13
, R. Knutti
2,14
,
N. S. Lovenduski
2,12,15
, J. Marotzke
2,16
, K. A. McKinnon
2,17
, S. Minobe
2,18
, J. Randerson
2,19
,
J. A. Screen
2,20
, I. R. Simpson
1,2
and M. Ting
2,8
Internal variability in the climate system confounds assessment of human-induced climate change and imposes irreducible lim-
its on the accuracy of climate change projections, especially at regional and decadal scales. A new collection of initial-condition
large ensembles (LEs) generated with seven Earth system models under historical and future radiative forcing scenarios pro-
vides new insights into uncertainties due to internal variability versus model differences. These data enhance the assessment
of climate change risks, including extreme events, and offer a powerful testbed for new methodologies aimed at separating
forced signals from internal variability in the observational record. Opportunities and challenges confronting the design and
dissemination of future LEs, including increased spatial resolution and model complexity alongside emerging Earth system
applications, are discussed.
NATURE CLIMATE CHANGE | VOL 10 | APRIL 2020 | 277–286 | www.nature.com/natureclimatechange
277
PersPective
NATurE CLImATE ChANgE
recommended big data practice of ‘bringing your analysis to your
data’. Following these principles, the CESM1-LE was made freely
available as a public dataset on the Amazon Web Services cloud
in October 2019. Access on the commercial cloud demonstrates
strong interest in LEs from industry and scientific communities well
beyond typical climate researchers. Such scrutiny and widespread
use attests to the value of LEs for a range of applications—truly a sea
change for climate and related sciences.
Strength in numbers with a Multi-Model Large Ensemble
Archive
While a single-model LE has enormous utility, a multi-model col-
lection of LEs can be leveraged for robust comparison of both the
forced response on regional or decadal scales across models and
internal variability across models. It can also advance model evalu
-
ation by providing more complete information on biases in inter-
nal variability versus those in the forced response. Unlike CMIP, a
multi-model archive of LEs allows for direct separation of projection
uncertainty into a structural component due to model differences
and an internal variability component. Despite these advantages,
most analyses to date have been limited to one or two LEs (with a
few exceptions; for examples, see refs.
20,21
), in part because of the
burdensome task of accessing large volumes of data from disparate
sources. To fill this gap, we have produced a centralized data reposi
-
tory of LEs conducted with seven different CMIP5-class ESMs
under historical and future emissions scenarios (hereafter referred
to as the ‘Multi-Model Large Ensemble Archive’ (MMLEA); Table 1).
This repository includes gridded fields of key variables at daily
and monthly resolution, and it is easily accessible via the National
Center for Atmospheric Research (NCAR) Climate Data Gateway
22
.
This Perspective seeks to illustrate what MMLEA can offer, with
the aim of widening its usage and stimulating new research direc
-
tions and Earth system applications. We also look to the future of
initial-condition LEs, in particular the opportunities and challenges
that confront their design and facilitate their accessibility to the
user community. In this regard, we offer a path forward that bal
-
ances demands for increased spatial resolution and model complex-
ity against ensemble size. We encourage future CMIPs to take on a
greater role in the design and coordination of LE simulations, data
storage and access.
New insights on separating sources of uncertainties
Individual LEs have been crucial to understand the need to con-
sider internal variability alongside forced trends in past and future
climate change at continental and smaller spatial scales
10,12,14,19,23–31
.
The MMLEA expands on this view by providing new insights on
the relative roles of internal variability and model structural differ
-
ences—two sources of projection uncertainty in addition to the radi-
ative forcing scenario. The MMLEA shows both factors can play a
first-order role in the magnitude and pattern of warming at continen
-
tal scales; for example, the distributions of trends in North American
air temperature over the last 60 years from each of the seven LEs
(Fig. 1 and Methods). While they all encompass the observed trend
value, they clearly differ in the strength of the forced trend (given by
the ensemble mean) and in the shape and width of the distribution
of trends, which emerge due to internal variability. This information
on model dependence of both the forced trend and its range due to
internal variability is unique to the MMLEA and could not have been
deduced directly from CMIP archives. It is important to note that a
LE centred on the single observed trend value does not constitute
evidence that this particular model is more realistic than any other
model (see further discussion in sub-section titled ‘Multi-model LEs
as methodological testbeds for observations’ below).
The distribution of North American temperature trends based
on CMIP5 (see Methods) is only slightly wider than that based on
an individual LE and is due to both model differences and inter
-
nal variability (see grey shaded probability distribution function
(PDF) (Fig. 1)). Moreover, the MMLEA spans a wider range than
CMIP5, suggesting that CMIP5 under-samples internal variability
at regional scales. This highlights the importance of evaluating the
realism of models’ internal variability of trends, since a model with
unrealistically large trend variability (that is, a broad distribution)
can encompass the observed trend for the wrong reason and would
also inflate uncertainty in projections. Approaches to address this
challenge are discussed in the sub-section titled ‘Multi-model LEs
as methodological testbeds for observations’ below.
Just as North American temperature trends vary across the
members of a LE, the geographical pattern of trends can also dif
-
fer strikingly (Fig. 1). This can confound comparisons of individual
simulations from different models and lead to erroneous interpre
-
tations, since internal variability might be mistaken for structural
differences. With enough members, the spatial pattern of the forced
response emerges for each model, allowing for a direct comparison
between models. Models may show similar forced patterns of pole
-
ward-amplified warming but different overall amplitudes (Fig. 1),
a conclusion that is difficult to discern without a MMLEA. Similar
issues confront the study of observed real-world trends (Fig. 1),
since these are also just one realization of many that could have hap
-
pened (see sub-section titled ‘Multi-model LEs as methodological
testbeds for observations’ below).
Table 1 | The multi-model LE archive and data repository
Modelling
centre
Model version Resolution
(atmosphere/ocean)
Years Initialization No. of
members
Forcing Reference
CCCma CanESM2 ~2.8°×2.8°/~1.4°×0.9° 1950-2100 Macro and micro 50 Historical, RCP 8.5 94
CSIRO MK3.6 ~1.9°×1.9°/~1.9°×1.0° 1850–2100 Macro 30 Historical, RCP 8.5 95
GFDL ESM2M 2.0°×2.5°/1.0°×0.9° 1950–2100 Macro 30 Historical, RCP 8.5 78
GFDL CM3 2.0°×2.5°/1.0°×0.9° 1920–2100 Micro 20 Historical, RCP 8.5 96
MPI MPI-ESM-LR ~1.9°×1.9°/nominal 1.5° 1850–2100 Macro 100 Historical, RCP 2.6, RCP 4.5, RCP 8.5 61
NCAR CESM1 ~1.3°×0.9°/nominal 1.0° 1920–2100 Micro 40 Historical, RCP 8.5 97
SMHI or KNMI EC-Earth ~1.1°×1.1°/nominal 1.0° 1860–2100 Micro 16 Historical, RCP 8.5 98
Salient characteristics of each LE, including the method of initialization. Here, the term ‘micro’, referring to micro perturbation
15
, indicates that all LE members begin from a single coupled model state, with
slight perturbations introduced only in the atmospheric component to create ensemble spread. The term ‘macro’, referring to macro perturbation
15
, indicates that the LE members begin from a variety of
coupled model states (for example, from different years in a long control simulation). Canadian Earth System Model (CanESM2) consists of a hybrid approach, with ten micro ensemble members for each
five macro ensemble members. Additionally, ‘forcing’ refers to the greenhouse gas concentrations used to drive the model simulations; ‘historical’ corresponds to observed forcing during the historical
period; and representative concentration pathway (RCP) refers to the estimated trajectory of greenhouse gas emissions corresponding to a range of radiative forcing values of 2.6 W m
–2
, 4.5 W m
–2
and 8.5
W m
–2
by 2100. Data from the multi-model LE archive are accessbible from ref.
22
.
NATURE CLIMATE CHANGE | VOL 10 | APRIL 2020 | 277–286 | www.nature.com/natureclimatechange
278
PersPective
NATurE CLImATE ChANgE
Quantifying model uncertainty requires knowledge of the forced
response in each model, but most models in past and current CMIPs
do not have enough available ensemble members to allow for a robust
estimate of the forced response. Instead, low-frequency statistical fits
to a single ensemble member are often used to estimate the forced
response (for examples, see refs.
8,9
). Consequently, internal variabil-
ity has to be estimated either from the residual of this fit or from long
pre-industrial control simulations. From these approaches, it is often
not easy or possible to robustly estimate systematic changes to inter
-
nal variability under increasing radiative forcing. The availability of
a MMLEA circumvents these limitations and assumptions. More
importantly, it allows one to separate the sources of uncertainty at
smaller spatial and temporal scales, and for quantities that are noto
-
riously variable, such as precipitation and extremes.
Decision-making and risk assessment in a variable climate
LEs are increasingly proving their utility in the context of real-world
decision-making
32
where full assessment of changing climate risks is
needed, including variability and extremes. In particular, discerning
changes in variability and extremes requires large sample sizes
33–37
:
the hallmark of LEs. Moreover, the MMLEA is critical for evaluating
the extent to which projected changes in variability and extremes
are model-dependent.
The Upper Colorado River basin—which feeds the largest reser
-
voirs in the US—is an example of where changes in mean and variabil-
ity can produce a wide range of climate risks for water managers. This
basin is located at a latitude where projected changes in precipitation
are notoriously uncertain: the transition zone between the expected
drying in the subtropics and the wetting at high latitudes
2,38–40
.
The MMLEA shows divergent outcomes regarding how decadal
mean precipitation will change in this region under a high-emissions
scenario (Fig. 2a). However, decadal variability of precipitation is
projected to increase on average by ~10% of the magnitude of the
forced change (Fig. 2b). This result by itself suggests a heightened
hazard of prolonged droughts and pluvials and could, in the absence
of consistent projections of mean change, provide useful informa
-
tion for refining water management strategies.
To illustrate the challenge of projecting extreme events, we
use an example of daily summer heat extremes for a location in
the south-central US centred on Dallas, Texas (see Methods). As
expected under global warming, daily July heat extremes at Dallas
are projected to increase over the twenty-first century; however,
their evolution is far from monotonic in any single ensemble mem
-
ber, and their rate and degree of increase varies considerably across
different realizations of future internal variability in the same model
(Fig. 3a). For instance, historical daily heat records could be broken
almost continuously starting in the late 2060s, or their occurrence
could be more punctuated, with some decades even as late as the
2090s spared from any days of record heat depending on how inter
-
nal variability unfolds (Fig. 3a). This variety of temporal expressions
of historical heat extreme exceedances should be a cautionary note
on the enormous impact of internal variability on rare events (for
examples, see refs.
31,32
). Results also differ among models, as differ-
ences in the amount of warming and in the magnitude of variability
combine into an uncertain future risk of exceeding a given thresh
-
old (Fig. 3b). Validating not only a model’s climatology or mean
trend but also its variability thus emerges again as an important step
when investigating, and ultimately constraining, future projections
in this case of extreme events
41
.
Attribution-focused LEs differ from those in the MMLEA in that
they often rely on regional or high-resolution global atmosphere–
land models to capture the small spatial scales of specific extreme
Trend (°C per 60 yr)
CESM1-CAM5 (40)
MPI-ESM (100)
Other LEs
CMIP5 (123)
Observations
North America
annual temperature
trend 1951–2010
MPI forcedObservedCESM1 forced
MPI coldestCESM1 coldest
MPI warmestCESM1 warmest
Relative density
–3 –2.5 –2 –1.5 –1 –0.5 0 0.511.522.5 3
Temperature (°C per 60 yr)
0 0.5 1 1.5 2.532
Fig. 1 | Internal variability and model differences in continental temperature trends. The distribution of 60-yr annual temperature trends (1951–2010) over
North America (24–72° N, 180–62° W) from seven ESM LEs (thin curves), 40 different CMIP5 models (grey shading), and observations (Berkeley Earth
Surface Temperature; vertical black line). The maps show the associated patterns of temperature trends: top row, observed and the forced component
(estimated by the ensemble mean) from two LEs (CESM1 in green and MPI in purple); bottom row, individual ensemble members from CESM1 (green) and
MPI (purple) with the weakest (‘coldest’) and strongest (‘warmest’) trends. Note that the individual member maps show the total (forced-plus-internal)
trends in the model LEs. Observed trends are analogous to an individual ensemble member in that they reflect forced and internal contributions.
NATURE CLIMATE CHANGE | VOL 10 | APRIL 2020 | 277–286 | www.nature.com/natureclimatechange
279
PersPective
NATurE CLImATE ChANgE
events
35–37,42,43
, and may prescribe additional boundary conditions,
such as the large-scale atmospheric circulation
44,45
. Nevertheless,
these types of ensembles highlight the large number of simulations
required to identify significant shifts in the probability of certain
events. We note that LEs can also serve these alternate types of
ensembles by providing lateral boundary conditions to more spe
-
cialized regional climate models
46
and oceanic boundary conditions
to higher-resolution global atmosphere–land models.
Multi-model LEs as methodological testbeds for
observations
Another key usage of LEs is to test methods suitable for application
to the observational record, for example those aimed at separating
the signals of internal variability and forced climate change from a
single realization (for examples, see refs.
29,30,47–51
). Using observations
alone, it is difficult to assess the skill of such separation methods due
to lack of true knowledge of the observed forced response or the full
range of variability, including extremes. However, separation meth
-
ods can be evaluated by applying the methodology to each LE mem-
ber individually and comparing the results to the model’s forced
response, estimated from the ensemble mean of the LE (Fig. 4).
Application to the MMLEA will identify if the validation has a
strong dependence on model structure.
An additional testbed application of model LEs is the development
of surrogate realizations of internal variability based on observations
(Fig. 4). Although one cannot replay the ‘tape of history’
52
with an
initial-condition perturbation in the real world, the single observed
trajectory is only one of many that could have plausibly occurred
(under the same boundary conditions and forcing). The underlying
premise of LEs is that internal variability can unfold with a different
(and largely unpredictable) chronology, creating uncertainty in the
estimate of trends that are calculated over a finite time interval. Can
the sample of internal variability in the observational record be used
to generate surrogate realizations whose statistical characteristics are
largely unchanged, but whose temporal sequences are altered? If so,
an observationally based LE can be developed, wherein these surro
-
gates are added to an estimate of the forced response (derived from
models or empirical methods applied to observations) to produce an
observationally constrained range of outcomes (Fig. 4).
Several methods for generating surrogate realizations that aim
to preserve the temporal
26
and spatio-temporal characteristics of
observed internal variability have been proposed
47,53–59
. To date,
these techniques have been applied to terrestrial temperature and
precipitation
26,47,59
, sea level pressure
47
and sea-surface tempera-
ture
53,55
. These methods interact in two important ways with model
LEs. First, model LEs can be used as methodological testbeds to
ensure that the statistical ensembles have the desired properties
(Fig. 4). Second, after the statistical ensembles are validated, they
can then be used to validate the model LEs. We demonstrate this
interplay with an example from the Observational Large Ensemble
(Obs-LE) developed by ref.
47
(see Methods).
Analogous to the approach mentioned above for estimating the
forced trend, the Obs-LE methodology can be tested in the con
-
text of a model LE by creating a statistical ensemble based on a
single member of the model LE, and assessing whether the spread
of the statistical ensemble is consistent with that of the remain
-
ing ensemble members. This procedure can then be repeated for
each ensemble member, and the resulting information pooled for a
robust estimate of the accuracy of the methodology (Fig. 4). In the
case of annual temperature trend variability over the past 50 years
on land, the fractional error of the Obs-LE methodology is generally
less than 20% over most of the globe, with slightly larger errors in
certain regions of the tropics (Fig. 5a). Assuming the properties of
the real world are not drastically different from those of the model,
this indicates that applying the same approach to generate a statisti
-
cal ensemble from the single realization of the real world is valid.
Having validated the Obs-LE approach, one can then assess the
realism of internal variability simulated by each model LE by com
-
parison with the Obs-LE. Regarding the CESM1-LE case, the model
overestimates variability of 50-year temperature trends by up to
50% in parts of western North America and northern Eurasia, and
up to 100% in areas of high terrain in the tropics (Fig. 5b). These
model biases are larger than the error of the Obs-LE methodology,
indicating they are true model biases. Similar results are found for
precipitation trend variability, which exhibits regions of both sig
-
nificant underestimation and overestimation in the CESM1-LE
47
.
One can also apply the Obs-LE to evaluate the simulated distri
-
butions of temperature trends at specific locations. For example, the
simulated temperature trend distribution for Dallas, Texas, in the
CESM1 and LE narrows considerably when the Obs-LE is used to
estimate the internal variability (Fig. 5b), which is consistent with
the model’s significant overestimation of variability at this location.
1950 2000 2050
2100
Time (yr)
-0.4
-0.2
0
0.2
0.4
0.6
Precipitation anomaly (mm d
–1
)
10-yr running mean relative to 1971–2000
CanESM2
CSIRO-Mk3-6-0
GFDL-CM3
CESM1-CAM5
GFDL-ESM2M
EC-Earth
MPI-ESM
1950 2000 2050
2100
Time (yr)
– 0.04
– 0.02
0
0.02
0.04
0.06
Change in s.d. (mm d
–1
)
Change in s.d. of 10-yr running means
relative to 1971–2000
Range across LEs
Mean across LEs
a
b
Fig. 2 | Decision-making under uncertainty: changes in mean and
variability. a, 10-yr running mean annual precipitation anomalies (mm d
–1
)
over the Upper Colorado River Basin (approximated as a spatial average
over 38.75–41.25° N and 111.25–106.25° W) relative to the reference
period 1971–2000 from each of the seven model LEs. Solid lines show the
ensemble means, and colour shading the 5–95% range across ensemble
members. b, Moving average of the change in standard deviation of 10-yr
mean precipitation (relative to 1971–2000), calculated across the individual
ensemble members of each model LE. The thick black curve shows the
mean, and grey shading shows the 5–95% range across the seven models.
Note the order-of-magnitude smaller range in the y axis in b compared to
a. s.d., standard deviation.
NATURE CLIMATE CHANGE | VOL 10 | APRIL 2020 | 277–286 | www.nature.com/natureclimatechange
280
PersPective
NATurE CLImATE ChANgE
This brings the observed trend closer to the lower tail of the dis-
tribution. It is worth emphasizing that without an observationally
based LE, it would not have been possible to assess the width of the
models’ temperature trend distributions, which indicate important
implications for constraining future projections.
An important future challenge for the LE community is to develop
effective means to evaluate and benchmark the internal variability
generated by model LEs. Meeting this challenge requires taking
advantage of historical and paleoclimate records, and developing
suitable statistical emulation methods to construct observationally
based LEs for other components of the climate system. Statistical
emulation of internal variability may also be advantageous in the
context of ESMs when the cost of conducting a sufficiently large
LE is prohibitive; for example, in the case of models with increased
spatial resolution and/or complexity (discussed further below in the
sub-section titled ‘Designing future initial-condition LEs’). These
statistical emulation methods will need to take into account any
projected changes in internal variability
60
.
Designing future initial-condition LEs
The existing LEs have been designed and created independently,
with different choices of time period, radiative forcing scenario,
number of members and method of initialization (Table 1). In addi
-
tion, they employ different protocols for data output, storage and
access. These differences must be considered when comparing LEs
across models, as each has ramifications.
Initialization. In some LEs, the initial conditions are created by
introducing miniscule perturbations (at the level of round-off error
or 10
–14
K, also known as ‘micro perturbation’
15
) into the atmo-
sphere. The rapid growth of atmospheric perturbations makes this
technique well suited for studies involving atmospheric variability
and trends. However, for persistent phenomena involving oceanic
or terrestrial processes, it may be more desirable to start each mem
-
ber from completely different initial conditions in the ocean and
other components (also known as ‘macro perturbations’) to bet
-
ter sample possible climate trajectories. Macro perturbations can
increase ensemble utility but also introduce subsurface ocean drift
in the control simulation that can influence ocean initial condi
-
tions; thus, they require long and quasi-equilibrated control simula-
tions to choose initial conditions from
61
. A combination of micro
and macro perturbations could have the most scientific benefit,
but ensemble initialization procedures need close examination and
potential coordination among LE projects.
Length of simulation and ensemble size. For a given amount of
computer time, there is a trade-off between the number of ensem
-
ble members and their length. For example, is it better (for some
purposes) to have a 100-member ensemble covering the period
1981–2040, or a 50-member ensemble extending over 1981–2100?
Furthermore, if higher spatial resolution is critical, such as for the
simulation of some climate extremes, this usually comes at the
expense of the total number of ensemble members that can be run.
The optimal balance between ensemble size and spatial resolution
will depend on the LE application (for further details, see ref.
62
).
Radiative forcing scenario. The forcing scenario may impact the
characteristics of internal variability. Is it better to run more mem
-
bers using a single choice of a forcing scenario, or multiple smaller
ensembles with differing scenarios? Even single scenarios are nor
-
mally comprised of individual forcing components (for example,
greenhouse gases and aerosols), and for the important but other
-
wise elusive goal of attribution, the use of ensembles with a single
radiative forcing (for example, only changing aerosols) can provide
critical insight into the mechanistic drivers
63,64
.
Data output, storage and access. As the LE applications expand to
broader timescales (diurnal to centuries), practical limitations arise
from the computational burden and storage requirements of main
-
taining hundreds of terabytes of data for analysis. At present, some
LEs only provide monthly averaged output, while others provide
daily averages but only for select fields. In general, practical storage
limitations require a compromise between ensemble size and choice
of output fields. Model fields can also be in non-intuitive formats
for users, limiting accessibility. Careful consideration should be
given not only to data storage, enabling workflows that bring analy
-
sis to the data, but also to format. We recommend single-variable
time series. We also encourage modelling centres to provide some
LE output interpolated onto conventional grid structures and/or
tools to accomplish this re-gridding—for example, for non-uniform
ocean model output.
Accommodating increases in model complexity and
resolution
High-resolution regional climate projections can also benefit from
MMLEs. As mentioned above, statistical and dynamical downscal
-
ing techniques can help resolve processes at smaller spatial scales.
Such efforts are currently limited by the trade-off between ensemble
Projected daily heat extreme occurrence in July at Dallas, Texas
CESM
no. 20
CESM
no. 35
CESM
no. 4
CESM
no. 8
CESM
no. 17
Projected probability of occurrence
1970 1990 2010 2030 2050 2070 2090
Time (yr)
0
10
20
30
40
50
60
Probability (%)
CanESM2
CSIRO-Mk3-6-0
GFDL-CM3
GFDL-ESM2M
CESM1-CAM5
EC-Earth
168 events
519 events
357 events
250 events
386 events
a
b
Fig. 3 | Decision-making under uncertainty: changes in extremes.
a, Vertical bars mark the occurrence of July days that meet or exceed the
historical (1950–1999) 99.9th temperature percentile for the grid box
containing Dallas, Texas, in five members of the CESM1-LE under historical
and future (RCP 8.5) radiative forcing. The 99.9th percentile is defined as
the average of the 99.9th percentile values calculated for each ensemble
member. b, Probability of exceeding the historical (1950–1999) 99.9th
percentile of daily temperature in July at Dallas, Texas, for six model LEs.
Thick coloured lines show the probability in each LE calculated over all
ensemble members, and colour shading shows the 5–95th percentile range
based on calculating the probability in each ensemble member separately.
Open circles and vertical bars show those same values for every other
decade from 1970 onwards, with models plotted in a staggered fashion
centred on year five of a given decade.
NATURE CLIMATE CHANGE | VOL 10 | APRIL 2020 | 277–286 | www.nature.com/natureclimatechange
281
