Fusion use case

[fcasson]

If you are interested in using this library please get in touch

We are using FTorch to couple plasma physics transport surrogates to an integrated plasma simulation.

The work is an offshoot of this paper: https://iopscience.iop.org/article/10.1088/1741-4326/ad240d/meta

We are very grateful for Ftorch, thanks making it! It fills a need for us, in the past some surrogates we had were manually coded evaluations in Fortran, which does not scale well when we have many new models coming in! And looser couplings probably don't have the performance we would like.

Some related projects to our use case

• https://gitlab.com/qualikiz-group/QLKNN-fortran
• https://github.com/google-deepmind/fusion_transport_surrogates
• https://github.com/ProjectTorreyPines/TGLFNN.jl

[jatkinson1000]

Thanks for this @fcasson ❤️
Great to hear the library is useful!

We like to feature work/research that uses FTorch in presentations and in our main README, so do let us know if we can publiscise anything. We are always keen to hear more about projects building on FTorch.

It may also be of interest to you that we are planning to run a 2 day workshop later in the year on the subject of ML-numerics coupling and hybrid modelling: https://cambridge-iccs.github.io/ml-coupling-workshop/

[fcasson]

Maybe in a few months we will have a publication and can let you know. Just got things up and running recently. Thanks for flagging the workshop, will think about it.

[jatkinson1000]

Brilliant, good luck.
We recently had FTorch published in JOSS which you can cite: https://joss.theoj.org/papers/10.21105/joss.07602

🙏

[fcasson]

William Hornsby will present our work and use cases at your workshop next week.

We recently open sourced the first NNs which we use with Ftorch, conference publications in preparation:

• https://github.com/ukaea/tglfnn-ukaea,
• https://github.com/ProjectTorreyPines/TurbulentTransport.jl/blob/master/utilities/README_onnx_to_pytorch_fortran.md

We're interested in any future developments that might allow us to easily run both training and inference in the same framework from Fortran. At the moment we do training and inference very separately, with clunky conversions and complex workflows in between, but ideally we would like to move towards wrappers which can build surrogates for an arbitrary expensive function, using continual learning when new parameter ranges are encountered.