Surrogate models?

[orbitfold]

Should we try to implement those workflows simply via appropriate analysis classes? I'm thinking krigging as proof of concept. I don't know much about it though so would be nice if someone with an application would chip in.

[wedeling]

There already is a simple surrogate model functionality in the SC analysis class (surrogate subroutine). Once all the samples are computed, the SC expansion is basically a polynomial approximation of the code output. It interpolates the code samples via Lagrange polynomials. We could implement the same thing for PC I think @jlakhlili?

[wedeling]

I'm also working on different surrogate models for time-dependent problems, but I'm not sure if EasyVVUQ is the place to incorporate these, since it does not follow the same workflow of sampling the code a number of times, and computing some stats from the output.

[orbitfold]

I'm also working on different surrogate models for time-dependent problems, but I'm not sure if EasyVVUQ is the place to incorporate these, since it does not follow the same workflow of sampling the code a number of times, and computing some stats from the output.

What workflow does it follow?

[orbitfold]

There already is a simple surrogate model functionality in the SC analysis class (surrogate subroutine). Once all the samples are computed, the SC expansion is basically a polynomial approximation of the code output. It interpolates the code samples via Lagrange polynomials. We could implement the same thing for PC I think @jlakhlili?

That class is very monolithic. Wonder if there is a way to modularize that functionality a bit.

[jlakhlili]

As I said in the VecmaTk telco. it's a good Idea. The first draft that I can do with PC, it could be like @wedeling example. After that, I can also implement/interface something using Gaussian Processes.

[orbitfold]

Also you probably would want to store the model for later use I guess? So maybe it's a good idea to think what is the best place for it. The database maybe?

[jlakhlili]

Is it not the case with CampaignDB?

[wedeling]

I'm also working on different surrogate models for time-dependent problems, but I'm not sure if EasyVVUQ is the place to incorporate these, since it does not follow the same workflow of sampling the code a number of times, and computing some stats from the output.

What workflow does it follow?

I have a subgrid-scale source term in the large-scale equation, for which I have reference snapshots in time. The surrogate is trained on those snapshots, and it then replaces the subgrid scale source term in the large-scale equation. At every time step, the surrogate takes large-scale variables as input (features) and produces a value for the source term. I then integrate the system in time for a long time, and compute statistics for the output.

The main difference with for instance SC or PC methods is that it is coupled to a (large-scale) PDE , and the inputs are not uncertain inputs, but large-scale features.

[wedeling]

As I said in the VecmaTk telco. it's a good Idea. The first draft that I can do with PC, it could be like @wedeling example. After that, I can also implement/interface something using Gaussian Processes.

That would be interesting, would you make the GPs part of EasyVVUQ?

[wedeling]

There already is a simple surrogate model functionality in the SC analysis class (surrogate subroutine). Once all the samples are computed, the SC expansion is basically a polynomial approximation of the code output. It interpolates the code samples via Lagrange polynomials. We could implement the same thing for PC I think @jlakhlili?

That class is very monolithic. Wonder if there is a way to modularize that functionality a bit.

Perhaps we could take out the Sobol indices part (which is quite large), and put it in a separate class?

[wedeling]

Also you probably would want to store the model for later use I guess? So maybe it's a good idea to think what is the best place for it. The database maybe?

The SC surrogate only needs the code samples to be stored.

[jlakhlili]

As I said in the VecmaTk telco. it's a good Idea. The first draft that I can do with PC, it could be like @wedeling example. After that, I can also implement/interface something using Gaussian Processes. That would be interesting, would you make the GPs part of EasyVVUQ?

I am thinking to start with a small example, and we can discuss about it in Poznan Hackathon.

[jlakhlili]

Also you probably would want to store the model for later use I guess? So maybe it's a good idea to think what is the best place for it. The database maybe?

The SC surrogate only needs the code samples to be stored.

I think that COLLATION_APP1 contains already the samples data.

[orbitfold]

how about I add a simple GP analysis class now and then we can further discuss it during the hackathon?

[jlakhlili]

how about I add a simple GP analysis class now and then we can further discuss it during the hackathon?

Please do it.