Question-and-Answer Resource for the Building Energy Modeling Community
Get started with the Help page

# Radiance Convergence - Simulation parameters

Dear all, I have been following your topics about the results convergence due to Radiance Monte Carlo method. At the same time, I have been working in a complex geometry (too many details) and from previous results, I have achieved variations of 30 % for 5 simulations using the following parameters in rcontrib (based on OpenStudio code):

type maps\Estante1.map | rcontrib -ab 10 -ad 65536 -as 512 -dj 1 -dp 1 -dt 0 -dc 1 -lw 1.52e-05 -I+ -fo -e MF:4 -f reinhart.cal -b rbin -bn Nrbins -faa -o matrices\Estante1.vmx -m skyglow octree\model_ill.oct

However, I would like to increase convergence. The model is taking more or less half-hour to finish and time is not a major issue. I would kindly ask your contribution since due to the great number of setting parameters, I am struggling in the definition of those. I am running now a simulation based on the topics that found here in unmethours (increase -ab, -ad and decrease -lw):

type maps\Estante_1.map | rcontrib -ab 20 -ad 262144 -as 1024 -dj 1 -dp 1 -dt 0 -dc 1 -lw 0.38e-05 -I+ -fo -e MF:4 -f reinhart.cal -b rbin -bn Nrbins -faa -o matrices\Estante1.vmx -m skyglow octree\model_ill.oct

Thank you in advance, Best regards, Nuno Garcia Saraiva

edit retag close merge delete

Sort by » oldest newest most voted

-as, -dj, -dp, -dt, and -dc have no effect in a daylight coefficient simulation with rcontrib, so you can ignore those.

The parameters you're using look similar the parameters I recommend for the interior view matrix of a three-phase calculation, which was determined with convergence testing on the BRE-IDMP dataset.For convergence testing I recommend taking roughly 1% of the test points and running simulations with varied parameters. The points chosen should be some of the more challenging, further from the window and in corners.

The critical parameters for matrix generation (with glow materials as light sources) are usually -ab, -ad, and -lw. I'd recommend something like this:

for ab in 6 8 10 12 ; do
for ad in 1024 4096 16384 65536 262142 ; do
for lw_power in 1 1.5 2 2.5 3 ; do

rcontrib ... -ab $ab -ad$ad -lw $lw ... < fewpoints.txt > result_${ab}_${ad}_${lw}.mtx

done; done; done


Then you can either compare the difference in matrix values, or you can generate a test set of sky vectors (sunny morning, sunny noon, sunny evening, overcast, etc.) and compare illuminance results between the runs. Then if thee's a big jump between ab 6 and 8, a small jump between 8 and 10, but only a minor difference between 10, and 12, you're okay with -ab 10. You could go back and test -ab 9 in a second run, if you want to be superbly optimal. If there's still a sizable jump between 10 and 12, then maybe run some more with -ab 14 and 16. I find it helpful to plot results, that way you can see if things flatten out, and you can also recognize noise (which will be apparent with low -ad and -lw runs)

You might also want to time the simulations, so you can estimate the full run length (which is semi-linear with number of simulation points).

And finally, rfluxmtx (which calls rcontrib in the background) has simplified the process of daylight coefficient simulations greatly. You might want to look at Sarith's tutorial and consider using rfluxmtx instead of rcontrib. https://www.radiance-online.org/learn...

more

Wouldn't some ambient exclusion or geometry simplification help smooth things out as well?

( 2019-09-06 10:32:46 -0500 )edit
1

Those tactics have less potency in the case of rcontrib, which doesn't use the ambient cache. Personally I'd just run with higher parameters and let the computer work hard so I can be lazy.

( 2019-09-06 10:50:45 -0500 )edit

( 2019-09-16 03:32:00 -0500 )edit

What is relationship of the points of interest to the transfer matrix surfaces? Is there even a T matrix? In other words, is this a single phase

I ran the Daylight Coefficient method (2PM) so there is no T matrix and no 3PM. I opted for that for not having complex glazing systems (only single layer windows). However, the geometry of the windows is rather complex.

how complicated is the window interface geometry? If it's complicated, you may want to consider the three phase method.

I did not know that if the geometry was complex 3 phase method would be preferable. There is here a picture windows' geometry which represents the southern façade. The northern is similar to this one. .

My gut reaction is that you have some complex geometry at the perimeter and so even with the aggressive sampling settings you have there for the view matrix, you're getting a lot of variation.

Yes, the geometry is very complex especially near the illuminance maps (view matrices) that are in front of the bookshelves.

So is it worth to keep increase the setting parameters trying to achieve more convergence, or try to follow a three phase method? Best regards

more

Hi!

I still struggle to find the best set of parameters for lighting simulation! But, yours seem quite "high", so I would expect your results to have converged already, thus my first next guess would be to simplify the geometry.

I am not sure I am reading your facade properly, but if you want to use the 2-Phase Method, I would keep the windows the same, without removing details. The 3-phase method is better when your facade is optically complex (i.e. when there will be a lot of bounces and scattering on when the light goes through your window), not when the shape of the window is complex. Your facade seems to match better the latter case...?

Now, I do not think you need all those details in your geometry. You are calculating the illuminance on a grid and thus you can avoid the trouble of dealing with corners and other complexities. I might be wrong, but I think all those details in your railings may be "absorbing" part of your high parameters. That is, lots of rays get stack bouncing around them without really improving the results on your grid. Maybe try simplifying those and see if you get more stable results.

more

( 2019-09-16 03:32:02 -0500 )edit

It would be helpful if you could share a bit about the model. What is relationship of the points of interest to the transfer matrix surfaces? Is there even a T matrix? In other words, is this a single phase (some folks call it two-phase, mmm hmm), where you are tracing rays from the POI directly thru fenestration and to the sky, or is it three-phase? If the former, how complicated is the window interface geometry? If it's complicated, you may want to consider the three phase method. My gut reaction is that you have some complex geometry at the perimeter and so even with the aggressive sampling settings you have there for the view matrix, you're getting a lot of variation.

more