NHM uncertainties

Uncertainties used

Initially the same uncertainties as the current shallow crustal validation workflow were used, these consisted of:

Realisations

The generated realisation files had the following properties, the first realisation is unperturbed and used as a baseline:

Values have been truncated to 4 d.p. Full details are attached as a csv to this page.

Common properties of all realisations are:

Future runs

Fault geometry parameters

Upon reflection as the physical geometry of the fault is well defined the trace location should not be perturbated in future runs.

As the magnitude is related to the area of the fault, this should not be perturbated either.

As the top and bottom of the fault have some uncertainties provided the future runs could incorporate them. Perhaps using dbottom_sigma as the sigma value, with dtop_min and dtop_max as the bounds, relative to dtop. As both dtop and dbottom would be moved the same amount by this method, the width of the fault would be conserved, and therefore the magnitude (as well as slip and moment).

To perturbate the dip the width of the fault would need to be calculated, the dip perturbated and then dbottom determined from the (conserved) width and perturbated dip.

Care must be take when perturbating the dip_dir in order to ensure the fault does not have planes with opposing strike (ie a fault with planes (0, 0) to (1, 0) and (1, 0) to (1, 1) and dip_dir 45 degrees would have this issue)

Rake is given with no uncertainty in the NHM, so should probably not be perturbated.

Hypocentre location is already perturbated from the existing workflow, so has no issues continuing as it is.

The length and length standard deviation size of the fault is provided in the nhm. As the trace is considered more reliable than the given length, this given value is ignored.

Other perturbated parameters used in the workflow

The property qsfrac needs more investigation to determine if it should be included in future.

The properties sdrop, rvfac, kappa were all taken as they are from the validation workflow, and can probably remain.

Any other parameters

For accurate IM comparison the HF seed should be same across all runs.

Core hour requirements

As these perturbations are only modifying existing parameters there is no increase in required core hours, compared with the current unperturbed workflow.

A total of approximately 16 CH were needed for each realisation (About 100 CH needed for the entire run)

Results

IM ratio plots were generated, comparing the median IMs with the perturbated IMs. As the HF seeds and the hypocentres were all randomised the plots seem to have a lot of noise

This has shown that while in need of refinement, the uncertainty workflow does work for srfs generated from nhm sources.

This can be particularly seen in the srf plots below, the location, dip and rake all being visibly different for each realisation.

Plots

The srf plots are available below:

The following pSA 10.0 im ratio plots are all taken with respect to the unperturbed realisation, using the same scale to demonstrate the relative differences. 180 additional plots are available on request.

For reference the magnitudes are given below again:

Uncertainties used

For the minimal perturbation run:

For the minimal perturbation run the only parameter that was perturbated was the srf generation slip seed

For the base case run:

For the base case run the only things that were perturbated were the hypocentre and srf generation slip seed, as per regular nhm based srf generation.

For the full perturbation run:

All the parameters identified in the pilot test were used in the full perturbation run

Results:

Median and standard deviation spatial IM plots were created from the output of the simulations. Ratio plots between the slip only and other runs were also created

IM	Seed only	Seed and hypocentre	All perturbations	Seed only - All ratio
PGA
PGV
pSA 0.01
pSA 1.0
pSA 10.0

Combined with IM - rrup plots, it appears that the pSAs for the frequencies associated with the HF part of the workflow were approximately the same, while the frequencies in the LF were significantly lower.

To investigate the cause of this another 50 realisations with varying levels of perturbation are being run, so that each perturbation component can be checked for the changes it makes.

Step by step comparison of perturbations

In order to determine the changes from each variable that was perturbated, the 50 realisations of Moonshine that were run with full perturbation were run with differing levels of perturbation removed.

5 batches of realisations were created, the first two being independent, and the rest building on the previous batch.

These batches were used with the 3 original runs to produce a range of

Perturbations	pSA10.0	pSA1.0	pSA0.1
Slip seed - full perturbation (as above)
slip seed - slip hypocentre
slip hypocentre - Batch 5
Batch 5 - Batch 4
Batch 4 - Batch 3
Batch 3 - full perturbation

In order to understand the differences between each pair of cohorts the response spectra for each realisation and their mean were plotted.

In each of these the less perturbated cohort is in blue, while the more perturbated cohort is in red.

Due to the layout the blue in each row is the red from the row above (except the top/second row).

As the largest drop in intensity seemed to come from adding magnitude and stress drop perturbations these were investigated further.

In the below plots each realisation is given a colour based on its normalised perturbation z score, while the median has a default colour.

There would not seem to be a (strong or positive) correlation between magnitude and intensity, as would be expected

It would appear that there is a correlation between sdrop perturbation and intensity, however discussions with Sarah indicate that  the expected effect of changing sdrop is a similar change in intensity, particularly at higher frequencies, the opposite of what is observed here.