DSI without real observations?

[RyanConway91]

Apologies for any butchering of terminology...

I want to use some sort of surrogate model to predict mine inflows for a slow model. I have been going through DSI (not at PESTPP product, I know...) tutorials and I can't see how to apply this process to my model, as I do not have measurements of mine inflow. DSI fundamentally works by building a "covariance matrix that links the measured past to the modelled future" I don't have a measured past.

I feel like what I want is PCA surrogate model where X and Y are "monte carlo" draws from the prior parameter and observation ensembles (IES with -1 iterations). This relatively straight forward to code in python, but want to check if there is already existing PESTPP/pyemu tools for this.

Or am I missing something in DSI?

[rhugman]

Hi @RyanConway91 -

We have DSI implemented in pyemu (see the GMDSI tutorials for examples). Fyi - we are going to release a new Class to handle emulation soon.

Do you have measured data other than pit inflows against which to history match? If not, I dont think you'll gain advantage by just running the DSI prior; unless you want to do some DSIVC type optimization/scenario testing.

[RyanConway91]

I am calibrating to pumping tests, so I guess I could combine my calibration and future models into a single PEST run and then use DSI. Thanks!

[rhugman]

Right on. Or run each version independently but with the same parameter ensemble and stitch the obs ensemble together before passing to DSI.

[robinkeegan]

I have been doing what Rui suggested (running separate calibration and forecast models and stitching the results together before using DSI). Seems like a good approach when you have several difference forecast scenarios and don't want to run the calibration period with each scenario.

[neilmanewell]

https://www.frontiersin.org/journals/water/articles/10.3389/frwa.2023.1334022/full

I think this might be something of interest perhaps. Using a surrogate model ensemble, I simulate a pumping test in a fine scale model. Then I use dsi to predict my posterior t and s in upscaled model cells around the test. A different way of thinking nonetheless. The output is a pumping test informed posterior ensemble you can use to predict mine inflows in your slow model.

One day when I have a bit more free time i want to write some functions (utilising pyemu) to automate this process. (At the moment it's a bit manual and in fortran/external python).