Drivers of Twenty-First-Century U.S. Winter
Precipitation Trends in CMIP6 Models:
A Storyline-Based Approach
1. Introduction
Regional precipitation trends are among the more challenging processes to
project in climate change scenarios. Some broadly defined patterns, such as
an increase in zonal-mean precipitation in the deep tropics, a decrease in
zonal-mean precipitation in the subtropics, and an increase in precipitation
at high latitudes, are robust across models and are expected for well-
understood theoretical reasons (Held and Soden 2006; Seager et al.
2010; Scheff and Frierson 2012). However, impacts of future precipitation
change will be felt not through the zonal mean, but locally, and at local
scales the picture is considerably more uncertain, especially over
midlatitude land regions such as the contiguous United States (Deser et al.
2012).
The sources of uncertainty in U.S. precipitation trends can be decomposed
into three categories: scenario uncertainty, model uncertainty, and internal
variability (e.g., Deser et al. 2012). The first two of these could in principle
be reduced by better constraints on future emissions and improvements in
climate models, respectively. For temperature projections, reductions in
scenario and model uncertainty by themselves would be enough to
substantially narrow the range of possible trends (Hawkins and Sutton
2009). However, actually achieving this improvement is challenging in
practice—the range in equilibrium climate sensitivity in climate models has
not narrowed across successive generations of models (Meehl et al.
2020; Zelinka et al. 2020). On the other hand, some improvements have
been made in model representations of precipitation (Fiedler et al. 2020).
The third category of uncertainty (internal variability) is an even more
challenging problem. Internal variability can be defined as the effectively
,random variability present in the climate system that is not driven by
external forcing. Although it is unforced, internal variability can still produce
substantial decadal or centennial trends, and this long-term manifestation
of internal variability can be thought of as a sort of “irreducible uncertainty”
whose distribution cannot be tightened by model improvements. The
special challenge of precipitation trends—in contrast to temperature trends,
for example—is that precipitation is especially sensitive to internal
variability. This sensitivity allows internal variability to mask forced trends in
precipitation, both for observations (Hoerling et al. 2010, 2016) and for
future projections (Deser et al. 2012; 2014). For some U.S. regions, the
forced signal in precipitation may not be distinguishable from internal
variability until after 2100 (Giorgi and Bi 2009).
Since the uncertainty in future U.S. precipitation cannot be eliminated, we
turn instead to the task of describing the range of plausible outcomes.
Internal variability in the climate system can be represented by a large
initial-condition ensemble, in which a single model is run many times with
identical forcings but with very slight differences in initial conditions [see a
recent review by Deser et al. (2020)]. The ensemble mean trend can then be
taken to represent the forced response, and the spread among ensemble
members can be taken as an estimate of the magnitude of internal
variability. Similarly, multimodel ensembles can be used to estimate the
range of outcomes due to the combination of two uncertainties: model
uncertainty and internal variability. In both cases, the projected outcome
can be represented by an ensemble mean with error bars determined by the
ensemble spread.
This traditional approach is simple and intuitive, but there are two reasons
one may wish to go beyond it. First, the method of identifying an ensemble
mean bracketed by error bars representing the spread implicitly assumes (or
may be interpreted as assuming) that the spread among ensemble members
can be interpreted probabilistically, an assumption that is not necessarily
justified (Zappa 2019). From the perspective of communication, another
drawback of this method is that representing outcomes in terms of a mean
, bracketed by error bars naturally draws one’s attention to the mean. Thus,
despite the best efforts of researchers to communicate uncertainties, this
presentation of the results may actually distract attention from them.
Second, we would ideally like to know why the models disagree, and which
physical drivers in the climate system are responsible for the spread in
trends. This question is of scientific interest, but it also may be practically
important if some of these drivers are represented more accurately in some
models than in others.
Hence, some recent work on climate impacts has moved away from the
approach described above, and instead has explicitly focused on exploring
multiple outcomes in detail. Several methods have been proposed for doing
this (e.g., Lenderink et al. 2014; Hazeleger et al. 2015). These often involve
regression-based approaches. That is, the regression of precipitation trends
(for example) onto the trends in some dynamical driver across an ensemble
of model runs is computed, and the regression slope is used to estimate the
sensitivity of precipitation to that driver (Bladé et al. 2012; Manzini et al.
2014; van den Hurk et al. 2014a,b; Deser et al. 2017).
One particularly interesting version of a regression-based method is the so-
called storylines approach of Zappa and Shepherd (2017) (see also Shepherd
et al. 2018; Shepherd 2019; Zappa 2019). Using CMIP5 models, these
authors 1) identified key dynamic or thermodynamic drivers linked to future
precipitation and wind speed trends over Europe, 2) determined the
sensitivity of the precipitation and wind speed trends to each process, and
then 3) constructed synthetic high- and low-impact storylines for the
twenty-first century based on combinations of these drivers. The key is that
multiple drivers are considered simultaneously, in order to quantify the
impacts of various combinations—the “storylines” of the method’s name.
The same method has since been applied to the Southern Hemisphere
midlatitudes (Mindlin et al. 2020). This method addresses both of the
problems identified above, since it demonstrates the various plausible
outcomes without requiring a probabilistic interpretation, and explicitly
Precipitation Trends in CMIP6 Models:
A Storyline-Based Approach
1. Introduction
Regional precipitation trends are among the more challenging processes to
project in climate change scenarios. Some broadly defined patterns, such as
an increase in zonal-mean precipitation in the deep tropics, a decrease in
zonal-mean precipitation in the subtropics, and an increase in precipitation
at high latitudes, are robust across models and are expected for well-
understood theoretical reasons (Held and Soden 2006; Seager et al.
2010; Scheff and Frierson 2012). However, impacts of future precipitation
change will be felt not through the zonal mean, but locally, and at local
scales the picture is considerably more uncertain, especially over
midlatitude land regions such as the contiguous United States (Deser et al.
2012).
The sources of uncertainty in U.S. precipitation trends can be decomposed
into three categories: scenario uncertainty, model uncertainty, and internal
variability (e.g., Deser et al. 2012). The first two of these could in principle
be reduced by better constraints on future emissions and improvements in
climate models, respectively. For temperature projections, reductions in
scenario and model uncertainty by themselves would be enough to
substantially narrow the range of possible trends (Hawkins and Sutton
2009). However, actually achieving this improvement is challenging in
practice—the range in equilibrium climate sensitivity in climate models has
not narrowed across successive generations of models (Meehl et al.
2020; Zelinka et al. 2020). On the other hand, some improvements have
been made in model representations of precipitation (Fiedler et al. 2020).
The third category of uncertainty (internal variability) is an even more
challenging problem. Internal variability can be defined as the effectively
,random variability present in the climate system that is not driven by
external forcing. Although it is unforced, internal variability can still produce
substantial decadal or centennial trends, and this long-term manifestation
of internal variability can be thought of as a sort of “irreducible uncertainty”
whose distribution cannot be tightened by model improvements. The
special challenge of precipitation trends—in contrast to temperature trends,
for example—is that precipitation is especially sensitive to internal
variability. This sensitivity allows internal variability to mask forced trends in
precipitation, both for observations (Hoerling et al. 2010, 2016) and for
future projections (Deser et al. 2012; 2014). For some U.S. regions, the
forced signal in precipitation may not be distinguishable from internal
variability until after 2100 (Giorgi and Bi 2009).
Since the uncertainty in future U.S. precipitation cannot be eliminated, we
turn instead to the task of describing the range of plausible outcomes.
Internal variability in the climate system can be represented by a large
initial-condition ensemble, in which a single model is run many times with
identical forcings but with very slight differences in initial conditions [see a
recent review by Deser et al. (2020)]. The ensemble mean trend can then be
taken to represent the forced response, and the spread among ensemble
members can be taken as an estimate of the magnitude of internal
variability. Similarly, multimodel ensembles can be used to estimate the
range of outcomes due to the combination of two uncertainties: model
uncertainty and internal variability. In both cases, the projected outcome
can be represented by an ensemble mean with error bars determined by the
ensemble spread.
This traditional approach is simple and intuitive, but there are two reasons
one may wish to go beyond it. First, the method of identifying an ensemble
mean bracketed by error bars representing the spread implicitly assumes (or
may be interpreted as assuming) that the spread among ensemble members
can be interpreted probabilistically, an assumption that is not necessarily
justified (Zappa 2019). From the perspective of communication, another
drawback of this method is that representing outcomes in terms of a mean
, bracketed by error bars naturally draws one’s attention to the mean. Thus,
despite the best efforts of researchers to communicate uncertainties, this
presentation of the results may actually distract attention from them.
Second, we would ideally like to know why the models disagree, and which
physical drivers in the climate system are responsible for the spread in
trends. This question is of scientific interest, but it also may be practically
important if some of these drivers are represented more accurately in some
models than in others.
Hence, some recent work on climate impacts has moved away from the
approach described above, and instead has explicitly focused on exploring
multiple outcomes in detail. Several methods have been proposed for doing
this (e.g., Lenderink et al. 2014; Hazeleger et al. 2015). These often involve
regression-based approaches. That is, the regression of precipitation trends
(for example) onto the trends in some dynamical driver across an ensemble
of model runs is computed, and the regression slope is used to estimate the
sensitivity of precipitation to that driver (Bladé et al. 2012; Manzini et al.
2014; van den Hurk et al. 2014a,b; Deser et al. 2017).
One particularly interesting version of a regression-based method is the so-
called storylines approach of Zappa and Shepherd (2017) (see also Shepherd
et al. 2018; Shepherd 2019; Zappa 2019). Using CMIP5 models, these
authors 1) identified key dynamic or thermodynamic drivers linked to future
precipitation and wind speed trends over Europe, 2) determined the
sensitivity of the precipitation and wind speed trends to each process, and
then 3) constructed synthetic high- and low-impact storylines for the
twenty-first century based on combinations of these drivers. The key is that
multiple drivers are considered simultaneously, in order to quantify the
impacts of various combinations—the “storylines” of the method’s name.
The same method has since been applied to the Southern Hemisphere
midlatitudes (Mindlin et al. 2020). This method addresses both of the
problems identified above, since it demonstrates the various plausible
outcomes without requiring a probabilistic interpretation, and explicitly
