Phase-space database for machine learning

[Gimzou]

Hi everyone,

I'm exploring the idea of replacing phase space sources with virtual source models, as suggested in the development project ideas discussion #790.

To my knowledge, current virtual source models are either trained with proprietary/private phase-space datasets or with the IAEA phase-space files available from the IAEA website.

Would it make sense to expand the IAEA dataset with more publicly available phase-space sources to foster innovation/development in the virtual source area ? more data, better accessibility => trusted virtual models (?)

Is there a possibility you hinted out that IAEA is modernising their phase-space database @rtownson ? (reddit comment on ParticleZoo ).

I'd be glad to have your thoughts on what should be done to move forward with those deep learning related tasks you pointed out in #790.

[rtownson]

1. Would it make sense to expand the IAEA dataset with more publicly available phase-space sources to foster innovation/development in the virtual source area ? more data, better accessibility => trusted virtual models (?)

Sure, more well-verified phase-spaces would be helpful for the community. The question is in how to produce them without violating proprietary terms. If you have access to more phase-spaces then I imagine the IAEA website could host them.

2. Is there a possibility you hinted out that IAEA is modernising their phase-space database?

We are working with them to develop a new IAEA phase-space file format. I'm not aware of any additions to the database, in terms of new data.

3. I'd be glad to have your thoughts on what should be done to move forward with those deep learning related tasks you pointed out in EGSnrc development project ideas #790.

In terms of using deep learning to produce phase-space files or optimize a virtual source, this would certainly take careful consideration to ensure its accuracy. Since we're aiming for high accuracy with MC simulations, there has to be a lot of justification to use anything less accurate than an "as accurate as possible" phase-space. Using deep learning is not guaranteed to be bias free, unlike proper use of variance reduction techniques.

[Gimzou]

Thanks a lot for the reply — really helpful.

I’m still very new to this area and don’t have any phase-space files to contribute myself (yet), so I’m mainly trying to understand where the real gaps are. From what you said, the main barrier is clearly not just data volume, but trust and accuracy, especially if ML-based approaches are involved.

That actually made me rethink my original question a bit. Maybe the bigger issue is not only hosting more phase-space data, but whether we have a community framework to properly evaluate virtual/ML-based source models properly in the first place.

So a more concrete question would be:

Is there interest in the community in defining a benchmarking and validation framework for virtual source / ML-based models, tied to MC-generated phase-spaces?

For example: agreed observables, accuracy criteria, and reference phase-spaces specifically meant for comparison.

If this is considered a useful direction, I’d actually be very interested in contributing on that side (datasets, validation ideas, benchmarking structure). I’m still early in the field, but this seems like a meaningful place where I could learn while working on something broadly useful.

(Separately, for logistics: if new non-proprietary phase-spaces were produced, is the IAEA still the natural place to host them — including for lower-energy / diagnostic ranges?)