This is a broad question. You can narrow it down a bit by defining the variable(s) of interest for your study. You also need to define your evaluation goals, eg- do you want to match hourly values or daily mean or yearly mean at a given grid point or average value of a given variable in both space and time. Once that is defined, your first step should be just visual inspection in case of a map or simple regression/ scatter plot/ time-series plot etc. You will most likely not get all the variables right. You just need to focus on your target variable. If it doesn't match well with observations, you can start looking for systematic and non-systematic errors. Also, be conscious of errors/biases in observed data. For example all your measurement stations may be predominantly urban which is not reflected well in a coarse-resolution global climate model. Errors in the model may be a bit more challenging to track-down as in atmospheric sciences, everything is related to everything else. So, it is hard to distinguish what's cause and what's effect. This problem will depend on your target variable and all the feedback effects that influence it. Good Luck!

Is ensemble approach really the best option? Because most models share the same modeling approaches, the physics is the same, parameterizations are almost the same, discretization are the same.

Kenneth M Towe

Others areas worth considering is how can we improve the science behind the models, the discretizations, numerical choices, etc.

Dear Gemechu Fanta Garuma

You can see the following link. In this study, we have chosen the global models as of their ability to capture the basic climate feature of the regional climate.

Article Interannual variability of rainfall over the Arabian Peninsu...

Performance of Regional Climate Model is very sensitive on initial and boundary condition, parameterization, time etc for which you are working with. Most vital thing is the region on what you want to address your study. So, it is always better to choose your regional model after one sensitivity test...After that you will have option for tuning the model for better result!

Regards,

There is a whole hierarchy of climate models, from almost trivial zero-dimensional energy balance models, to 1- and 2-dimensional energy balance models, then to dynamic models using various approximations in various domains (e.g., quasigeostrophic or primitive equations, on beta-plane or sphere...), all the way up to "earth system models", with land, ocean, ice components all coupled together. Remember that models are just the numerical expressions of some theory that you want to test. You, the modeler, are in control, and can include or exclude particular components according to your immediate purpose.

See this paper by Isaac Held, on "The Gap between Simulation and Understanding in Climate Models": Article The Gap Between Simulation and Understanding in Climate Modeling

Myself and Lee Branscome addressed a similar question about "Minimal modeling of the extratropical general circulation" in Tellus in 1989 Article Minimal modeling of the extratropical general circulation

In doing any climate related numerical experiment, you should probably always choose the "minimum" dynamics and physics that capture the process you're interested in. E.g., if you want to investigate the MJO (or 30-60-day wave, or tropical intraseasonal oscillation), there's no real need to include separate ice or ocean models. Sure, you can look for MJO signals in ESMs that have all the bells and whistles, but you will probably learn more about the MJO by finding it in simpler models.

So rather than ask what are the minimum criteria for climate models, try to formulate more fundamental questions. E.g., what aspects of the climate do you want to simulate? What minimal features are needed to simulate them? E.g., for meridional heat flux, you probably need to allow for baroclinic instability, and enough wave-wave and wave-mean flow interactions for the heat-carrying waves to organize themselves realistically. You can add in topography to find what impact that has, and to get stationary waves. You could then add in moisture to account for latent heat transport. Then you're into shallow and deep convection, and pretty soon you've included all kinds of physics - but you find your model doesn't represent hurricanes very well, so you need to go to ultra-high resolution to get them right. By this stage you have fallen very far down the slippery slope of complexity and you will worry more about the scalability of your MPI communications than the physics of the climate....

So my advice is, try to design your models to be as simple as possible in order to reveal something useful about the real world.

Quite agree with Enda William O'Brien that “models are just the numerical expressions of some theory that you want to test." It seems we often forget this, and it's not unusual to hear people blaming models to be "unrealistic" with problems they are not designed to solve. Even models are growing more and more complex and sophisticated, they are still just simple representation of the ultra-complicated real world, based on our theoretical understanding and formulation of it. So, the minimum requirement for climate models depends on what problems you want to explain with the model. You would not expect a model with a horizontal resolution of over 200 kms to capture the number of hurricanes that hit Florida in the past decades.

Hello

It all depends on the purpose of your study. There are bunch of models out there for different purposes. A model can be good on one aspect and very bad on another aspect. There is no minimum requirement that works for all cases.

Thanks

Let me tell one important information regarding atmospheric models. There is no universal conceptual framework. It solely depends upon the atmospheric system or process you want to investigate. As a result, there are differences in performance amongst various models in simulating any particular system. The uncertainty may come from different angles such as model dynamics and physics, discretization, initial and boundary conditions, model domain, horizontal and vertical resolution etc. Importantly, the treatment of subgrid scale physical processes is very significant while performing simulation. Usually, every model offers various schemes/models to represent those process. Now it is customary to find the set of those schemes which perfectly works for your requirement and for that you need to compare results from few example simulation named as "customization". Finally, after that you can use the model for long term or forecast purpose. Therefore, there is as such no definition of "minimum requirement" and rather vague. It solely depends on what you want to study.