Regional Climate Models |
Raymond W. Arritt | |
Iowa State University, Ames, IA 50011 | |
rwarritt@bruce.agron.iastate.edu |
Part 1: Motivation for regional climate modeling |
WhatÕs a model? |
Most generally, a model is a representation of something else | ||
Physical models: model ships, model airplanes, etc. Sometimes used to help with design of the real thing (as in wind tunnels). | ||
Conceptual models: ÒI think it works like this.Ó | ||
Mathematical models: An equation or system of equations that represents a physical system. |
The Climate System |
Climate models |
Yesterday you talked about global climate models. | ||
Regional climate models work in the same way, except that they do not cover the entire globe. | ||
many regional climate models include code in common with global climate models |
Starting point for numerical models: Laws of physics |
We have a problemÉ |
The coupled, nonlinear partial differential equations that express these physical laws are impossible to solve exactly. | ||
Not difficult, or time-consuming, or expensive, but impossible. |
Numerical solutions |
Definition of a derivative: | ||
This suggests a method for solving the equations: | ||
define space and time at discrete points x, t | ||
distance between points or times is Dx or Dt | ||
approximate derivatives in the governing equations as finite differences Dy/Dx or Dy/Dt | ||
(another method fits harmonic functions and/or polynomials in space) | ||
Discretizing a continuous world |
Still have problems |
Computers are not (never will be?) large and fast enough for Dx to be infinitesimally small: | ||
Solution degrades as Dx becomes larger. | ||
There may be important processes that occur on scales smaller than our grid. |
Increased computing power has allowed finer resolution |
North America in the Hadley Centre global climate model |
HadCM3 grid spacing relative to Tropical Storm Edouard |
Advantages of higher resolution |
Regional climate models allow use of finer resolution |
HadCM3 grid spacing is about 280 km. | |
To reduce the spacing to 50 km, we would need (280/50)3 = 175 times the computing power. | |
Proposal: Use a finer-scale model over only a limited region of interest. | |
Dynamical Downscaling |
How do regional climate models work? |
Regional models cover a limited area and so need information from global models |
Global and regional model grid points |
Even more problems |
Some examples of parameterization |
Turbulence: | ||
If the low layers are warmer than the upper layers, thermal turbulence will occur (warm air rises through cold air). Turbulent mixing acts on scales of millimeters to a few hundred meters. | ||
Parameterization: Gradually mix the layers if temperature decreases strongly with height between layers. | ||
Deep convection (thunderstorms): | ||
Thunderstorms develop when the atmosphere is warm and moist near the surface and cool aloft, and if condensation occurs. Motions are on scales of tens of meters to a few km. | ||
Parameterization: Vertically rearrange heat and moisture if the lower levels are sufficiently warm and moist and grid scale motion is upward (promotes condensation). Deposit leftover moisture as rain. | ||
Problems continue |
Mostly because of the nonlinearity of the equations, small differences in the initial state eventually grow to completely change the solution. | ||
"Does the Flap of a Butterfly's Wings in Brazil Set Off a Tornado in Texas? " (E. Lorenz, 1972, via P. Merilees or D. Lilly) | ||
Do multiple simulations starting from slightly different initial states: ensemble prediction | ||
Also create ensembles by using multiple models. |
Advantages of ensembles |
May allow some estimate of confidence or uncertainty: | ||
If two solutions disagree, at least one of them is wrong. | ||
If solutions agree, can we have greater confidence? Currently much work on this topic. Spread-skill relationships. | ||
In practice, the simple average of all the solutions – i.e., the ensemble mean prediction – often performs as well as or better than the best individual solution. |
An ensemble of IPCC model runs |
A simple ensemble |
How wide is this screen? |
Mother Of Ensembles (aka Shukla Staircase, other names) |
Summary of Part 1 |
Climate modeling is hard. | ||
We can never obtain exact solutions to the governing equations. | ||
The spatial resolution (grid point spacing) possible with present-generation computers leaves out many things we are interested in, or requires that we represent them using educated guesswork (i.e., parameterization). | ||
There are advantages to doing lots of runs using lots of models (i.e., ensemble simulations). | ||
Regional climate models give a way to use improved resolution over a particular area of interest. | ||
Part 2: Regional climate model methods and projects |
Mother Of Ensembles (aka Shukla Staircase, other names) |
How much of this is necessary? |
Minimum 4 main IPCC scenarios (A1, A2, B1, B2), about 20 global models, 6-member GCM ensemble, 10 regional models, 8 ensemble members per regional model. | ||
4x20x6x10x8 = 38,400 regional climate model runs (or 3,840 runs per regional model). Not practical! | ||
What are the greatest sensitivities in nested global-regional climate models? | ||
How can we most efficiently employ our computer time and (most important) people? | ||
Results from ensemble studies using GCMs and short-range forecast models may not apply: | ||
Regional climate models are (should be) constrained by lateral boundary conditions. | ||
A study of sources of uncertainty in regional climate models |
Compare four different types of ensembles: | ||
lagged average ensemble (sensitivity to initial conditions) | ||
perturbed physics ensemble (sensitivity to closure parameters) | ||
mixed physics ensemble (sensitivity to closure schemes or assumptions) | ||
multi-model ensemble (inter-model variability) | ||
Test case is the summer 1993 flood over the central U.S. |
Sensitivity to initial conditions: Lagged-average ensemble |
Start a forecast using an analysis at some time T for initial conditions. | ||
Perform additional forecasts starting from times T+1, T+2, T+3,... all ending at the same time as the first forecast. | ||
The overlap period gives an ensemble of forecasts starting from different but physically plausible initial states. | ||
here, forecasts begin at 00 UTC 15 May 1993 and at preceding 12-hourly times |
Lagged ensemble |
Perturbed physics ensemble |
How much variability in RCM simulations is due to settings of closure parameters? | ||
Parameters that control the behavior of parameterizations. | ||
Construct an ensemble in which each member uses a different value for a closure parameter or parameters: | ||
Must truly be an adjustable parameter (e.g., donÕt vary gravitational acceleration or specific heat). | ||
Parameter value should be reasonable. | ||
Here: in the Grell scheme of RegCM2, vary | ||
Dp = lifting depth threshold for trigger | ||
t = time scale for release of convective instability |
Perturbed physics ensemble |
Mixed-physics ensemble |
How much variability is produced by using different physical parameterizations in the same model? | ||
here, different techniques, as opposed to different parameters within the same technique | ||
Construct an ensemble whose members use the same initial conditions, but different parameterizations: | ||
convective parameterization: Kain-Fritsch, Betts-Miller, Grell | ||
explicit moist physics: simple ice, mixed phase, Reisner-2 | ||
shallow convection on or off |
Multi-model ensemble |
How much spread is created by using completely different regional climate models? | ||
Perform simulations using different models executed by different modeling groups but with specified initial and boundary conditions: | ||
Analyze 12 models from the PIRCS 1B experiment. |
Test case |
Flood over north-central U.S. | ||
1 June - 31 July 1993 | ||
Why? | ||
Extreme event of practical interest. | ||
Corresponds to PIRCS 1B period: the archive of PIRCS simulations provides a "target of opportunity." | ||
Initial and boundary conditions from NCEP/NCAR Reanalysis: | ||
this is an analysis of past observations that blends observations and a physical model (the model fills in the holes between observations) | ||
boundary data updated every 6 hours |
Analysis measures |
Mean for each group of simulations | |||
Spread (standard deviation) of results in each group | |||
For ensemble forecasts: | |||
Equitable threat score | |||
sensitive to phase error | |||
Bias | |||
Probability of detection | |||
False alarm rate |
Equitable threat score |
ETS = (H - C) / (H + F + M - C) | ||
where C = hits by ÒchanceÓ = (forecasted occurrences) x (event frequency) = (H + F) x (H + M) / N | ||
notice phase error reduces H, and increases both F and M |
Other criteria |
Bias = (forecasted occurrences) / (actual occurrences) = (H + F) / (H + M) | ||
ranges 0 to infinity; ideally = 1 | ||
Probability of detection = H / (H + M) | ||
ranges 0 to 1; ideally = 1 | ||
False alarm rate (aka probability of false detection) = F / (F + O) | ||
ranges 0 to 1; ideally = 0 |
Area-averaged precipitation in the north-central U.S. |
Verification results |
Results from longer regional climate simulations using the lagged method |
Sensitivity to source of boundary data |
In summary |
The main sources of uncertainty in regional climate modeling are: | ||
model formulation | ||
source of initial / boundary data | ||
Sensitivity to initial conditions is constrained by the continual flow of information into and out of the regional domain. |
Model intercomparison programs (MIPs) |
Run different regional climate models for the same region and time period, and evaluate performance of models. | |
Trend is away from evaluating relative skill of the models (Òbake offsÓ or Òbeauty contestsÓ) toward combining the models in an ensemble. |
A few past and ongoing regional climate MIPS |
PIRCS – Project to Intercompare Regional Climate Simulations (continental U.S., summers of 1988, 1993) | |
ARCMIP – Arctic Regional Climate Model Intercomparison Project | |
NAMAP and NAMAP-2 – North American Monsoon Model Assessment Project (southwest U.S. – Mexico) | |
PRUDENCE – Prediction of Regional scenarios and Uncertainties for Defining EuropeaN Climate change risks and Effects | |
ENSEMBLES – EU sponsored successor to PRUDENCE | |
ICTS – Inter-CSE Transferability Study (multiple regions) | |
NARCCAP – North American Regional Climate Change Assessment Program | |
See supplement to Takle et al. (2007), Bulletin of the American Meteorological Society for more. |
Current MIPS |
Regional modeling projects have begun using multiple global models to provide input for multiple regional models. Two current programs doing this are: | |
ENSEMBLES – Europe | |
NARCCAP – North America | |
North American Regional Climate Change Assessment Program (NARCCAP) |
Assess regional climate change for North America by downscaling 4 AOGCMs with 6 regional climate models (RegCM3, MM5, Scripps RSM, Canadian RCM, WRF, Hadley Centre regional model). | ||
About 12-15 combinations will be simulated. | ||
Project phases and status: | ||
Phase I: RCMs driven by reanalysis (1979-2004) to examine uncertainty in RCMs (completed) | ||
Phase IIa: RCMs driven by AOGCM output for 20th century climate (1971-2000) to examine combined GCM-RCM uncertainty (in progress) | ||
Phase IIb: RCMs driven by AOGCM output from SRES A2 scenario (2041-2070) (in progress) | ||
Slide 51 |
Slide 52 |
NARCCAP Participants |
NARCCAP Domain |
Comparison with observations |
Phase I: RCMs driven by NCEP/DOE Reanalysis 2 for 1979-2004 | ||
Evaluate errors due to RCM downscaling by using ÒobservedÓ boundary conditions. | ||
Is there value in using the models as an ensemble? How should we construct such an ensemble? |
Regions Analyzed |
Coastal California |
Mediterranean climate: wet winters and very dry summers (Koeppen types Csa, Csb). | ||
More Mediterranean than the Mediterranean Sea region. | ||
ENSO can have strong effects on interannual variability of precipitation. |
Monthly time series of precipitation in coastal California |
Correlation with Observed Precipitation - Coastal California |
Pacific Coast |
Very wet winters and moderately dry summers (Koeppen types Cfb, Csb). | |
Highly complex topography. |
Slide 61 |
Slide 62 |
Upper Mississippi River Basin |
Continental climate with hot summers and cold winters (Koeppen types Dfa, Dfb). | |
Maximum precipitation usually is in April-June. | |
Most NARCCAP models simulated this region in the PIRCS project. |
Slide 64 |
Deep South |
Humid mid-latitude climate with substantial precipitation year around (Koeppen type Cfa). | |
Past studies have found problems with RCM simulations of cool-season precipitation in this region. |
Monthly Time Series - Deep South |
Monthly Time Series - Deep South |
Maritime provinces |
Moist, cool mid-latitude climate with little seasonal variation in precipitation (Koeppen types Dfb, Dfc). | |
This region is near the outflow boundary of the regional model domain. |
Slide 69 |
Transferability |
How general are regional climate models? | ||
Does ÒtuningÓ a model for one region limit the model's skill for other regions? | ||
How well do models perform outside their Òcomfort zoneÓ of regions where they have previously been applied? | ||
Run multiple models on multiple domains | ||
A "model" is a specific configuration of a specific code. | ||
Many codes include several options for a given physical parameterization. | ||
Hold model choices constant for all domains. No adjustments for different domains. | ||
Hypothesis: Testing regional models in this way may tell us something about their applicability in a changing climate. |
Slide 71 |
Slide 72 |
Summary |
Regional climate models provide a basis for dynamical downscaling for climate impacts or process studies. | ||
The sources of variability in regional climate models are not the same as in global climate models because the solutions are constrained by the large-scale input data. | ||
As with global climate models, there appears to be benefit in producing ensembles of regional climate simulations: | ||
Given multiple simulations, how do we construct an ensemble? |
Slide 74 |
Thank You! |