Uncertainty in future climate:

1
Uncertainty in future climate What you need to
know about what we dont know
Philip B. Duffy Climate Central, Inc.
climatecentral.org
1
2
Outline

• Basics
• Climate vs. weather
• How we express uncertainty
• Why is future climate uncertain? Imperfect
  knowledge of
• Initial conditions
• Drivers of climate change (forcings)
• Climate system response to drivers

2
3
Outline

• Basics
• Climate vs. weather
• How we express uncertainty
• Why is future climate uncertain? Imperfect
  knowledge of
• Initial conditions
• Drivers of climate change (forcings)
• Climate system response to drivers

3
4
Outline

• How do we estimate climate uncertainty?
• Expert elicitation
• Ensembles of opportunity
• Perturbed physics ensemble
• Why none of these is perfect
• Guidance for decision-makers
• Avoid excessive risk

4
5
THIS TALK APPROVED FOR
climatecentral.org
5
6
Outline

• Basics
• Climate vs. weather
• Why is future climate uncertain?
• How do we estimate climate uncertainty?
• Guidance for decision-makers

6
7
Weather Conditions at specific time(s) and
location(s)
An example of weather Precipitation on 31 May
2007.
7
8
Climate a statistical description of
weather(averages and variability)
An example of climate Multi-year mean
precipitation
8
9
9
Source Roberto Buizza, European Centre for
Medium-Range Weather Forecasting
10
Climate prediction skill depends on knowledge of
external drivers
Particulate pollution
Greenhouse gases
Surface properties
Surface properties
10
11
Weather vs. climate prediction summary

• Weather prediction
• Uses models that are very similar to climate
  models.
• Predict conditions at specific times and
  locations.
• Skill of predictions depends on very accurate
  initialization of model.
• Time horizon is at most a week or so.
• Skill is constantly evaluated (and constantly
  improves).

• Climate prediction
• Climate models also predict weather! We analyze
  the statistics of the predicted weather, but not
  the weather itself.
• Skill depends on knowing future perturbing
  influences.
• Time horizon is typically decades, but can be
  even longer.
• Skill of predictions cannot be directly
  evaluated.
• We think models are improving.

11
12
Outline

• Basics
• Climate vs. weather
• How we express uncertainty
• Why is future climate uncertain?
• How do we estimate climate uncertainty?
• Guidance for decision-makers

12
13
Probability Density Function (PDF) and Confidence
Interval
PDF
Relative Probability
95 confidence interval (3º to 7º)
0 1 2 3 4 5 6
7 8 9 10
Change in annual-mean temperature (ºF)
13
14
Why is future climate uncertain?
14
15

• initial conditions in the atmosphere, etc.
• Future climate forcings, e.g. greenhouse gas
  concentrations
• how the system responds to forcings.
• These errors arise from
• numerical discretization
• Imperfect representation of unresolved phenomena
• relevant processes that are omitted (including
  unkown unknowns).

15
16
16
Source Roberto Buizza, European Centre for
Medium-Range Weather Forecasting
17
Example of initial condition uncertainty
Simulated and observed regional sea-surface
temperatures courtesy Ben Santer
1900 1920 1940 1960
1980 2000
17
18

• initial conditions in the atmosphere, ocean,
  etc.
• future behavior of climate forcings, e.g.
  greenhouse gas concentrations
• how the system responds to forcings.
• These errors arise from
• numerical discretization
• unresolved phenomena
• relevant processes that are omitted.

18
19
Future CO2 concentrations are unknowable this is
true of other influences also
19
20
About half of future uncertainty in temperature
comes from uncertainty in future CO2 emissions.
Each vertical bar shows the range of
results obtained for one greenhouse gas emissions
scenario
0.6o C is the amount of warming that occurred
during the 20th century.
Global T will increase by 1.4º - 5.8 ºC before
2100.
20
21

• initial conditions in the atmosphere, etc.
• future behavior of climate forcings, e.g.
  greenhouse gas concentrations
• how the climate system behaves.
• These errors arise from
• Imperfect representation of unresolved phenomena
  (notably clouds)
• numerical discretization
• unknown unknowns.

21
22
Computers only tell you what you already know.
Ernesto Colnago
22
23
Different models respond differently to same
inputs
Name of model
23
24
Outline

• Basics
• Why is future climate uncertain?
• How do we estimate climate uncertainty?
• Expert elicitation
• Ensembles of opportunity
• Perturbed physics ensemble
• Why none of these is perfect
• Guidance for decision-makers

24
25
Expert Elicitation

• Fancy term for asking a bunch of so-called
  experts.
• Why I dont like this approach
• Its completely subjective
• (but often made to look quantitative)
• Groupthink creates false consensus

25
26
Ensemble of opportunitya collection of
results from a number of available models

• Analyze results of a number of climate models

Results from 15 models, each simulating 3 CO2
scenarios
26
27

• Whats good about quantifying
• uncertainty in this way?
• Its a start

27
28

• Whats good about quantifying
• uncertainty in this way?
• Its a start
• 2. The mean of a large number of models
• consistently performs better than any
  single model
• This is true in climate simulation and in
  seasonal weather prediction
• So having results form multiple models seems to
  give a better estimate of the most likely outcome.

28
29

• Whats bad about quantifying
• uncertainty in this way?
• Results can be influenced by selection of models,
  which can be haphazard.
• Can be misleading because errors common to many
  models may be important. I.e., even if models
  agree with each other, they could all be wrong.
• Superiority of mean model suggests that this is
  not important
• Hence this approach measure consensus more than
  uncertainty

29
30
Whats bad 3. Some evidence that GCMs have
been unconsciously tuned
Source Kiehl, GRL (2007)
30
31

• Whats bad
• Often values all models equally, which cant be
  optimal
• But we cant agree on best way to combine
  models

Temperature change-gt
Temperature change-gt
31
Tebaldi C., Knutti R. Phil. Trans. R. Soc.
A20073652053-2075
32
Whats bad 5. Does not include outcomes that
all agree have low (but non-zero) likelihood.
Temperature response to 2 x CO2 (ºC)
A range of model results estimates the
uncertainty in the most likely outcome, not the
full range of possible values.
32
Source Roe and Baker, UW
33
Outline

• Basics
• Why is future climate uncertain? Imperfect
  knowledge of
• How do we estimate climate uncertainty?
• Expert elicitation
• Ensembles of opportunity
• Perturbed physics ensemble
• Why none of these is perfect
• Guidance for decision-makers

33
34
Perturbed Physics Ensemble

• Many simulations performed with one model,
  varying values of parameters that are uncertain.
• E.g. Climateprediction.net

34
35
Perturbed Physics Ensemble

• Good A better way to estimate the full range of
  possible outcomes.
• Bad Based on only one model (but does not have
  to be).
• Bad highly demanding computationally.

Relative likelihood
35
Temperature response to doubling atmospheric CO2
36
Outline

• Basics
• Why is future climate uncertain? Imperfect
  knowledge of
• How do we estimate climate uncertainty?
• Expert elicitation
• Ensembles of opportunity
• Perturbed physics ensemble
• Why none of these is perfect
• Guidance for decision-makers

36
37
Two quasi-fundamental barriers to good estimates
of climate uncertainty

• Future climate forcings, e.g. greenhouse gas
  concentrations, may be unknowable.
• It is very difficult to know if climate models
  share important errors.

37
38
Outline

• Basics
• Why is future climate uncertain? Imperfect
  knowledge of
• How do we estimate climate uncertainty?
• Why is this inadequate?
• Guidance for decision-makers

38
39
Things to keep in mind

• All decisions involve uncertainty
• A common approach is to avoid excessive risk
• (also known as CYA)
• Risk probability x badness of outcome
• Postponing a decision may be OK but dont
  expect climate science to improve markedly from
  year to year.

39
40
Ask yourself 5 questions

• Is there consensus among models?
• Does what the models predict seem sensible?
• Is the predicted change seen already in
  observations?

If yes to these, then no reason to doubt the
models.
4. What would happen if the models are
right and you ignore them? (How much
trouble would you be in???) If you
dont like the answer to this, then ask the 5th
question
40
41
5. Do you feel lucky?
41
42

• Parting Thoughts (1)
• We are only starting to think seriously about
  climate uncertainty.
• We are learning how to estimate uncertainty, but
  need to do better.
• There are some major barriers to good uncertainty
  quantification.
• It is important to work with decision-makers to
• find better ways to make good climate-related
  choices
• do the best we can with todays knowledge

42
43

• Parting Thoughts (2)
• Given all the limitations, what can
• we learn using todays climate projections?
• We can
• develop methodologies for making well-informed
  decisions
• assess what aspects of climate decisions are
  sensitive to
• determine if future climate is too uncertain to
  allow us to draw reliable conclusions
• improve our understanding of how natural and
  human systems respond to climate change.

43
44
44
45
 Predictions of better models are
indistinguishable from projections of worse
models. Climate model evaluation is based on
the assumption that better ability to reproduce
observations implies better predictions of the
future.  The evidence does not support this
assumption.
45
46
Weighted based on different basis variables and
metrics
46
Source Levi Brekke (USBR)
47
Parting Thoughts

• Climate models work amazingly well.
• Climate models have serious errors.
• Some important sources of error in future climate
  predictions are irreducible.
• Climate prediction is no longer an academic
  exercise!
• The need to incorporate climate change into
  real-world decisions has raised the bar for
  climate modelers.
• Quantifying and reducing uncertainties are major
  challenges.

47
48
Thanks for dinner!
48
49
Most of the observed increase in globally
averaged temperatures since 1950 is very likely
gt90 due to the observed increase in
anthropogenic greenhouse gas concentrations
The balance of evidence suggests a discernible
human influence on global climate
There is new and stronger evidence that most of
the warming observed over the last 50 years is
attributable to human activities
49
50
A better and cooler way to quantify uncertainty
climateprediction.net
48,000 participants are running a climate model
in background on their computers. 43,672,873
simulated years had been run as of April
23.  Each participant runs a slightly different
model version, with a unique combination of
parameter values. The result is a thorough
exploration of parameter space.
Relative likelihood
Temperature response to doubling atmospheric CO2
50
51
How do climate projections depend on apparent
model skill?
They dont!!!
51
52
Societal impacts of climate changeThe basis of
policy decisions
Air quality
Agriculture
Extreme events
Recreation
Human health
Water availability
52
53

• Societal-impacts studies need climate projections
  having
• Fine resolution
• to provide regional-scale fidelity
• Reliable information on extremes
• because these have disproportionate societal
  impacts
• Quantified uncertainties
• usually by analyzing a large family of
  simulations

Its difficult impossible to make
projections having all these properties!
53
54
Ed Lorenz (1917-2008) Discovered the concept of
chaos as a meteorologist at MIT
54
55
Model weighting does not affect
PDFs of future temperature
55
Source Levi Brekke (USBR)
56
Let me get this straight

• Climate model evaluation is based on the
  assumption that better ability to reproduce
  observations implies better predictions of the
  future.
• Available evidence does not support this
  assumption
• Predictions of better models are
  indistinguishable from projections of worse
  models.
• 3. Therefore, point (1) is a religious belief.

56
57
Drink Responsibly
57
58

• Parting Thoughts
  (1)
• Outstanding technical problems
• Uncertainty quantification presents inherent and
  practical difficulties
• We need quantitative methods for incorporating
  probabilitstic climate projections into decision
  processes
• Dynamical downscaling is arguably superior, but
  producing comprehensive libraries of dynamically
  downscaled results takes a lot of work!
• Statistical methods allow easy downscaling of
  ensembles of GCM simulations.
• Simulation and downscaling of extremes,
  especially of precipitation, are challenges.
• In coastal regions a fine-resolution
  ocean-atmosphere model is probably needed to
  represent coupled phenomena like fog,
  upwelling, etc.

58
59

• Whats wrong with quantifying uncertainty in this
  way?
• It combines uncertainties from all sources the
  contributions of individual sources cant be
  disentangled.
• It can be misleading because errors common to
  multiple models may be important. I.e. even if
  models agree with each other, they could all be
  wrong.
• It does not give more weight to models that
  reproduce observations well.
• It does not show the full range of possibilities,
  because each model tries to give the best answer.
  I.e. it does not show outcomes that all agree
  have low likelihood.

59
60
Typical uncertainty quantification
Results from 15 models, each simulating 3 CO2
scenarios
60
61
61
Source Roberto Buizza, European Centre for
Medium-Range Weather Forecasting
62
Weather prediction is more difficult than climate
prediction
Climate is driven by influences that are knowable
Weather is highly random
62
Source Roberto Buizza, European Centre for
Medium-Range Weather Forecasting