Re: [fluka-discuss]: uncertainty level in MC simulations

From: 委v鋱k Aleksandras <aleksandras.sevcik_at_ktu.edu>
Date: Fri, 3 Mar 2017 16:17:55 +0000

hello again,


If I may back to my question about the uncertainty with this concrete practical example, I would appreciate the comments,


Hereby attached picture with statistical error in usrbin format ( 5 mln histories, usrbin dose scoring in the voxelised phantom, analogue rad-decay of radioactive isotope, data extraction with SimpleGeo tool) - do I understand correctly, that every bin has its own error evaluation, and my goal would be to achieve that every bin would have error less than 5E-02, i.e. 5% ?

________________________________
From: me_at_marychin.org <me_at_marychin.org>
Sent: Monday, February 20, 2017 11:00:22 PM
To: 委v鋱k Aleksandras
Cc: Alfredo Ferrari
Subject: Re: [fluka-discuss]: uncertainty level in MC simulations

Alex,

Putting our conversation back on fluka-discuss for future reference by
other users in the community; some posts were missed due to a technical
glitch.

Flair > Run > Data. You should see a list of files you are about to
combine. In case you don't, play with Add/Remove/Rules/Refresh until you
get the list. Then click Process. After a brief wait you should see a
green box announcing success, and you will find some new files generated
by Flair in your directory. At this point, you can go to Plot to plot
the results (without ever need to open or look into the new files
generated by Flair). If you want to open and look inside the files, you
may. In case the files are in binary format, you may ask Flair to
convert to ascii: Run > Files > ->Ascii.

:) mary

On 2017-02-21 03:58, 委v鋱k Aleksandras wrote:
> Mary, thank you very much for your structured answer, I had an idea
> that is something like that, just wasn't sure. Please, could you tell
> me exactly how to combine the results from all cycles and get sigma
> value with flair? There is something mentioned in the manual, but in
> such way that I could not really understand it.
>
> regards, alex
> -------------------------
>
> FROM: me_at_marychin.org <me_at_marychin.org>
> SENT: Monday, February 20, 2017 2:27:31 PM
> TO: 委v鋱k Aleksandras
> CC: owner-fluka-discuss_at_mi.infn.it
> SUBJECT: Re: [fluka-discuss]: uncertainty level in MC simulations
>
> Dear Alex,
>
> Generally:
> #1 we decide on a sigma value to aim for;
> #2 we do a series of short preliminary runs (as opposed to long
> production runs) e.g. 10 cycles of 5 minutes each;
> #3 Flair combines for us the results from all cycles, telling us the
> sigma in % -- one value for each bin. This sigma is the standard error
>
> (not the standard deviation).
> #4 From those sigma values, we can estimate the number of histories we
>
> need to achieve #1:
> Sigma is inversely proportional to the square root of the number
> of
> histories. If we get sigma=10% with N histories, we would need 4*N
> histories to bring the sigma down to 5%, and 100*N to bring the sigma
> down to 1%. This is generally true, unless there are under-sampling
> issues, which I shouldn't start elaborating at this point.
>
> From #3 we are free to plot the error bar with one, two or three ...
> sigmas. Just multiply the value by two if we want to plot the errorbar
>
> with two sigmas, multiply by three if we want to plot the errorbar
> with
> three sigmas ....
>
> We won't be able to estimate the sigma from a single run. That's why
> we
> need multiple cycles. Each cycle will give us a mean for each bin.
> Individual scores contributing to the bin are not recoverable at the
> end
> of a Monte Carlo run, that means there is no way to calculate the
> standard deviation directly (unless one intercepts the runs, which is
> not recommended).
>
> The multiple cycles are to help us estimate the standard deviation in
> the form of the standard error. >=10 cycles are recommended. With less
>
> cycles, our standard error would give a poorer estimate of the
> standard
> deviation. With more cycles, our standard error would provide a better
>
> guess of the standard deviation.
>
> For anthropomorphic phantoms in particular, the first thing to
> consider
> is the voxel size. Don't use dimensions smaller than necessary. 1 mm x
> 1
> mm x 1 mm is something I have never done and will probably never do.
> Unless one is a micro-dosimetrist, he/she needs very clear motives to
> want to do <=1 mm. I would down-sample/re-bin to, say, 5 mm or 1 cm.
> Down-sampling from 1 mm to 5 mm would give us a volume increase of 5^3
> =
> 125 times! Down-sampling can be done either at the moment of export
> (e.g. from CT workstation), or using Matlab / Python /... (just a line
>
> or two would do the job, it's straightforward).
>
> :) mary
>
> On 2017-02-20 19:23, 委v鋱k Aleksandras wrote:
>> Dear all,
>>
>> As I am learning mc on my own, I often lack some guidance even in
> the
>> general subjects, so excuse me if the question is too basic.
>>
>> Is there any simple way /rule of thumb to determine how much
> histories
>> I need to have/how much particles to run in order to get the
> desirable
>> uncertainty level? Let's say I have a voxelised human phantom and
>> measuring a dose distribution from the internal source. How many
>> histories I need to run in order to have three sigma uncertainty
>> level? Or does it vary and should be recalculated as errors
> averaging
>> the different run files?
>>
>> Any advice/links to basic tutorials/other information regarding the
>> topic will be much appreciated,
>>
>> Regards
>>
>> Alex




__________________________________________________________________________
You can manage unsubscription from this mailing list at https://www.fluka.org/fluka.php?id=acc_info

statistics.jpg
(image/jpeg attachment: statistics.jpg)

Received on Fri Mar 03 2017 - 18:26:47 CET

This archive was generated by hypermail 2.3.0 : Fri Mar 03 2017 - 18:26:49 CET