Can climate models reliably predict meteorological extremes?

Question

This question is related to and extension of my previous question which was about the methodology used to predict future climatic extremes:

What is the methodology to analyze future climatic extremes using the results of climate models?

Although the answer to that question is quite detailed and informative, it is still not clear to me whether the models used in global climate (GC) simulations can predict rare (~1000-year return period events).

• I understand that if we have multi-thousand simulations, we can calculate the 0.1% quantile, but I do not know whether the applied models can predict rare events reliably. Under reliably, I mean that in some alignment with the observed extremes or with the extremes statistically inferred from the observations.
• All the papers I saw so far, which dealing with extremes due to climate change, are considering 20-year return period events as maximum and comparing the simulations to observed quantiles. This is one source of my suspicion regarding the reliability of GCMs.
• Are GCMs calibrated to observed mean or multiple quantiles? Daily, monthly, annual, etc. averages are used? Averages of larger areas or grid points (corresponding to the applied resolutions) are adopted in the calibration, validation processes? What are the typically used target parameters used in the calibrations?
• Here is an example to make it more clear: one wants to determine the probability distribution function of extreme wind speed in a given location from daily measurements/simulation results. The question: If we take one simulation for the 1950-2000 period and assume for a moment that model uncertainty is negligible compared to climate variability, can we treat the simulation data as a sample with more or less equivalent confidence than that of the actual observations, corresponding to the same period, to infer large return period extremes?

One way to check this is would be to compare the large return period extremes inferred from observed data and from multiple simulations, I have not seen any papers with above 20-year return level comparison yet. I see that we cannot expect very accurate models regarding e.g. 1000-year extremes, because we cannot reliably calibrate and validate GCMs:

• The available measurements are from a 100-200 year observation period (at best);
• We do not have too much data from different climatic conditions (there might be some phenomena which are not (or not adequately) taken into account in current models, we are somehow biased by the observed, prior 100 year conditions.

Any help, idea, reference, link are welcomed.

Answer 1

Here's one idea to account for meteorological extremes in our climate models, though I'm not sure if it answers your question and am less certain if it qualifies as reliable (open for feedback on that too):

Take a regional climate model which has simulated into the past and the future. We also have a few decades of observed data to compare the same time period of simulated past. The observed data is going to capture the extreme events much better in every way of course. So, in projecting into the future, if we want to maintain those extremes, we can use the change factor (based on averages, e.g. monthly means) from past-simulation==>future-simulation in order to project past-observed==>future-predicted. Those future-predictions will have the extreme events as they came in the past but with the changed magnitude based on the climate model's past versus future simulations.

This could be applied to meteorological data such as precipitation and other weather variables in order to inform us about the driving weather forces we'll face in future scenarios. In this way our climate models account for meteorological extremes we've observed recently and in the past, which the models may otherwise fall short on.

---

Otherwise, as folks have explained in comments on the OP, we simply don't have extensive enough observed data that we can rely upon to calibrate these rarer frequency events. I think your last two bullet points sum up our predicament fairly well:

• The available measurements are from a 100-200 year observation period (at best);
• We do not have too much data from different climatic conditions (there might be some phenomena which are not (or not adequately) taken into account in current models, we are somehow biased by the observed, prior 100 year conditions.

However, here are a few studies and leads you could explore which try to overcome these challenges in modeling extreme event return intervals and Probable Maximum Precipitation. These seem like they might be addressing your last question but hopefully it helps clarify the potential out there.

This one using extreme-value theory: Van Den Brink et al, 2004. Estimating return periods of extreme events from ECMWF seasonal forecast ensembles.

...applications to meteorological and hydrological situations are always hampered by the brevity of the available datasets, as the required return levels exceed the record lengths by a factor of 10 to 100. In order to overcome this problem, we u se archived data from all past seasonal forecast ensemble runs of the European Centre for Medium-Range Weather F orecasts (ECMWF) since 1987 as input for extreme-value statistics analysis. We make use of the fact that the seasonal forecast has little seasonal skill for the Netherlands, which implies that the ensembles can be regarded as i ndependent sets that cumulate to over 1500 years. We investigate the hydraulic response in the Netherlands to extreme synoptic-scale weather systems by studying the extreme-value distributions of sea storm surge levels, wa ves and river discharges. The application is detailed in four practical examples originating from coastal protection, ri ver flooding protection, and water management problems. The long record length of the ECMWF data reduces the uncertainty in the 10 3 -year and the 10 4 -year return values considerably with respect to the results based on observational time ser ies. The ECMWF dataset gives the opportunity to explore the distribution of events that depend on several kinds of extreme (Van Den Brink et al, 2004).

This one using the growth factor method, or "NERC M5 method", as explained on page 10: Skaugen and Førland, 2011. Future change in extreme precipitation estimated in Norwegian catchments.

And here is a study which sums the methods up pretty well and proposes a new one for further research, which sounds good to me and relatively like what I described in pretty broad, simple terms in the first part of my answer: Dyrrdal, 2012. Estimation of extreme precipitation in Norway and a summary of the state-of-the-art. There's many quote-worthy passages in this one so I think I'll leave it to you to explore. Sections 3, 4, and 5 are what I imagine you're most interested in because they cover, respectively, the state-of-the-art guidelines/methods from the World Meteorological Organization and various nations; a proposed new-and-improved method; and a conclusion summarize the above.