To reduce the CH usage of the HF part of our simulations it was determined that the variable dt could be doubled to reduce the number of timesteps calculated, and therefore the computational time required.
Currently:
For all simulations an HF dt of 0.005 is used, the data from the LF component, which has dt of 0.02, 0.01 or 0.005, is then interpolated and merged to form BB at a dt of 0.005.
Proposed:
If HF is run at dt=0.01 it is possible to interpolate it to merge with LF for form BB data with dt=0.005.
The main issue is whether this dt=0.01 data is scientifically valid.
Due to the number of timesteps changing to every second time step it is not possible to directly compare the output of two runs with the same seed as is normally done when changes are made to the HF binary.
A statistical approach must be used to determine if the difference between a run with dt=0.01 and a run with dt=0.005 is statistically insignificant.
Method:
30 realisations of the fault Hossack were run once with the previous workflow (HF dt=0.005, BB dt=0.005) and 30 were run with the proposed workflow (HF dt=0.01, BB dt=0.005).
The same Data directory was used.
The IM values of these faults (Without the LF component) were calculated .
For this investigation the IM PGA was inspected. For each station the ratio between the new value and the old value were calculated and summary statistics were created.
One realisation of each were also used to create IM ratio plots.
Results:
The ratios had a mean of 1.002268 and standard deviation of 0.039279. A similar comparison with an additional set of HF dt=0.005 data showed a mean of 1.001674 and standard deviation of 0.035706, so the mean and variance only increased slightly.
Example ratio plot:
Conclusion:
It was determined that the differences between dt=0.01 and dt=0.005 results are not significant and we can use the dt=0.01 workflow in future.
A time trial was conducted with a larger fault (Moonshine) and found that the WCT for dt=0.005 was 971s, while the WCT for dt=0.01 was 514s, a 47% reduction.
We still need to determine how this will work with the gmsim versions. The 400m and 200m grid will be fine as both LF and HF used a dt of at least 0.01, while the 100m grid uses LF with a dt of 0.005.