Sobol samples and batch size guidance for SWAT+ Toolbox

[Ali Akram]

Hi Celray,

I am using Temple’s HPC with 128 cores for SWAT+ Toolbox sensitivity analysis and calibration. I have HRUs in the range of 25–30k, and I am running the analysis at a monthly time scale.

Could you please advise how many samples would be ideal for the Sobol algorithm, and what batch size I should use during calibration, keeping in view the availability of 128 cores?

Thanks. 

[Celray James]

first if you do not need any output, turn it off... over 20K HRUs is an enormous number of HRUs so avoid writting unnecessary outputs to disk. second, use a short simulation period, like 3 or 4 years for sensitivity analysis. third, you can do with about 800 parameter samples (note, this is not the same as seed) - a compromise due to the large number of HRUs... however if parameters are more than 15, you may eventually need more samples... fourth for calibration, IF you are using CALSI, I would recommend not more than 64 processes because the more the processes, the more disk space you will need and the more the concurrent writes which results in issues of disk speed becoming a bottle neck (btw, make sure you are running in SSD/NVME partition, not in HDD partition). so you could try 60 processes with 20 batch size.

[celray.chawanda]

3-4 years run is a compromise. it is important to recognize that. also, you will still need a few 2-3 year warmup on top of that.

for calibration, it is a bit tricky but i would run it with 5 years + warm up of 3 years. another option is to simplify the model so you have fewer HRUs and everything runs faster, then you do not have to compromise on sensitivity and calibration.

-------- Original message --------

From: Ali Akram <ali.nia...@gmail.com>

Date: 1/2/26 20:06 (GMT-06:00)

To: Celray James <celray....@gmail.com>

Cc: SWAT+ Toolbox <spto...@googlegroups.com>

Subject: Re: [SWAT+ Toolbox] Re: Sobol samples and batch size guidance for SWAT+ Toolbox

Thank you so much for your response.

For the Colorado River Basin, I am using climate data from 1990-2024 with a 10-year SWAT+ warm-up period. As you suggested, I should use only 3-4 years for sensitivity analysis.

To clarify the workflow: if I run the model using a 3-4 year period for sensitivity analysis, and then want to use a longer simulation period for calibration, do I need to rerun the SWAT+ Toolbox using the longer period for calibration?

Thanks again for the clarification.

--
You received this message because you are subscribed to the Google Groups "SWAT+ Toolbox" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sptoolbox+...@googlegroups.com.
To view this discussion, visit https://groups.google.com/d/msgid/sptoolbox/30aca281-a2dc-431c-9cbe-fb3532005fccn%40googlegroups.com.

[Ali Akram]

Thank you so much for your response.

For the Colorado River Basin, I am using climate data from 1990-2024 with a 10-year SWAT+ warm-up period. As you suggested, I should use only 3-4 years for sensitivity analysis.

To clarify the workflow: if I run the model using a 3-4 year period for sensitivity analysis, and then want to use a longer simulation period for calibration, do I need to rerun the SWAT+ Toolbox using the longer period for calibration?

Thanks again for the clarification.

On Fri, Jan 2, 2026 at 7:21 PM Celray James <celray....@gmail.com> wrote:

--