I hope you don’t mind.

Let me know if these are what you’re looking for

Several professional engineering societies

I’d fudge this by inserting that the societies will no longer accept papers for publications for which Verification and convergence

I think it’s more accurate to say: societies have published policies stating they will no longer accept will no longer accept papers for publications.

The sad fact is, scholarly journals use a peer review process that involves sending papers to mostly univerisity professors, who, let me assure you, do not check to find out whether tests for grid independence etc. were done. They all agree that it should be done. Most perform these checks when doing their own computations (though I’m sure many don’t). However, the peer review system is appallingly bad at *actually checking* *this particular* sort of thing and you’ll almost never see any proof these checks were done appearing in a scholarly article.

I think in the bit on Numerical Solutions of PDEs, you might be able to say something rather stronger in regards to GCM’s. I suspect that GCMs are known to give different results using different grid sizes. Since all computations involve coarse grids, this could lead to systematic errors or biases if we decide to try to determine what might be “real” by averaging the results of a variety of codes all of which are known to use coarse grids. (Basically, if they all share the same type of flaw, the average will retain this flaw. It won’t average out.)

On the semi-implicit time step: Do GCM codes use semi-implicit time steps? (My impression from reading the text of the GISS papers is they run into the courant condition, and overcome it by inserting false numerical diffusion at the poles. Maybe they use explicit time stepping? You could look at the actual code for that– or check the wording in the papers. I don’t happen to know.)

You might have to dig to find the references that tell use we know the solutions are not grid independent, but I’m pretty sure I’ve read that somewhere.

This situation does not exempt developers and users of such models and codes from providing detailed information about the limitations of the presented results.

Absolutely true.

The presented results are numbers, not solutions of the underlying continuous equations.

I guess I’d be a bit less severe here. I think they may not be solutions. (But then, I haven’t looked at these things all that long. )

Dan,
Is it possible that if you code an AOLGCM on different software development platforms or run it on different hardware processor platforms you will get divergent runs of the same model? Same data, same algorithms, different software development and/or operating system, different hardware processors. I would bet that the same exact AOLGCM algorithms and data would produce divergent model runs on different platforms.