| 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! |