Paper: Challenges and Implementations for ML Inference in High-energy Physics

[sanjibansg]

This PR adds a paper submission in SciPy Proceedings 2025 for the paper on Challenges and Implementations for ML Inference in High-energy Physics by Sanjiban Sengupta and Lorenzo Moneta

Hi, I am Sanhita Joshi, @sanhitamj . I will serve as the editor for this submission. Reach out to me if any assistance needed.

[github-actions]

Curvenote Preview

Directory	Preview	Checks	Updated (UTC)
papers/sofie	🔍 Inspect	✅ 73 checks passed (24 optional)	Oct 14, 2025, 6:08 PM

[ameyxd]

Inviting reviewers: @Schefflera-Arboricola

[Schefflera-Arboricola]

Hi, just checking — should I start reviewing the paper, or is the PR still in draft due to the CI failure or ongoing content updates?

[ameyxd]

Hi, just checking — should I start reviewing the paper, or is the PR still in draft due to the CI failure or ongoing content updates?

You should be able to review the content. @sanjibansg - please review the failed checks.

[Schefflera-Arboricola]

Thank you very much for submitting this paper @sanjibansg !
The paper looks pretty good to me. I learned a lot while reviewing :)

Below are some general comments to consider for improving clarity, structure, and accessibility. I’ve also left in-line comments to highlight specific instances and suggestions tied to these broader points:

• consider adding a "Conclusion" or “Discussion” or "Summary" section with insights on -- trade-offs of SOFIE vs others, future improvements, where SOFIE is not suitable, etc.
• Some of the in-line comments are about explaining certain terms and phrases-- if you feel those things might be affecting the flow/direction of your narrative you can consider creating an "Appendix" and adding those details in that section.
• Adding transition sentences to enhance the narrative flow and the message you want to convey. Refer in-line comments for more.
• Using full-forms of acronyms when they are used for the first time (instances of this are highlighted below in in-line comments).
• more references to SOFIE and all the other tools, concepts, benchmarks, claims, etc. mentioned in the paper (https://root.cern/manual/tmva/#sofie ; how this paper(https://repository.cern/records/r9zqe-y7v55) led to this paper in this PR) ; Providing sources for such technical details can help reinforce the paper’s credibility. Refer in-line comments for more.
• The paper seem to be focused more on SOFIE, so it might be good to include SOFIE in the title and make it less general and more specific; just a suggestion : SOFIE: Efficient ML Inference for C++-Based HEP Workflows and a Review of Modern ML Deployment in HL-LHC Environments (it's too long, i know)
• Please ensure a consistent character limit per line throughout the paper. While 88 characters is common in many open-source projects, this paper is a .md file. So, if there are no specific guidelines, I'd recommend choose a reasonable limit and apply it consistently. Long single-line paragraphs make it harder to review and the suggestion diff in in-line comments does not appear clearly.

Thanks again for the thoughtful work -- looking forward to the next iteration! Please feel free to disagree if you don't want to include certain suggestions :)

Also, it might be nicer and easier to view the in-line comments by going over to the "Files changed" tab.

[sanhitamj]

@sanjibansg Please address the comments on the PR. The content of the paper is interesting, but the form needs work for readability. @Schefflera-Arboricola has made great suggestions as comments or as suggestions for better readability.

The deadline to work with the reviewers has been extended to 22 Aug; and the deadline to finish working on the comments is 5 Sept.

[himaghna]

Hi! Thanks for submitting this paper @sanjibansg! Wanted to quickly introduce myself. I am Himaghna Bhattacharjee and I will be helping review this paper

[himaghna]

Could be helpful to provide a more exhaustive list of missing functionalities in existing PyTorch extensions for research tasks to help strengthen the evidence. Would be even better to provide references to published literature using GNN

[sanjibansg]

AFAIK it is not a missing functionality, but more like such extensions have more python heavy APIs therefore not accessible through Libtorch which is a C++ interface.

We only mentioned PyG since it is widely used in the development of GNNs, other extensions are more integrated within the PyTorch ecosystem, and their usage in the Physics experiments varies according to the experiments' usage.

I am not sure what the references to published literature using GNN mean, do you imply to provide references to models developed using PyG that are used in High-energy physics research?

[sanhitamj]

@sanjibansg The deadline to work reviewers' comments in the paper is 5th September; next Friday. If you'd like your contribution to be included in the final Proceedings, please address all the comments by the reviewers.

[fwkoch]

Checks are failing because of missing DOIs. These articles do have DOIs, though, so I added them.