Skip to content

Commit a31ed86

Browse files
ereinhaereinha
committed
Add HEPSIM Project
1 parent 05281dd commit a31ed86

7 files changed

Lines changed: 346 additions & 0 deletions

File tree

‎_gsocprojects/2026/project_HEPSIM.md‎

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,10 @@
1+
---
2+
project: HEPSIM
3+
layout: default
4+
logo: HEP_SIM.png
5+
description: |
6+
High Energy Physics Simulation applies modern machine learning and agentic AI methods to Monte Carlo event generation and simulation in high-energy physics. Projects range from developing symbolic regression techniques for event reweighting and curating jet observable libraries, to deploying ML algorithms that quantify simulation biases between generators such as Pythia and Herwig. More advanced projects explore agentic workflows for automating exclusion limit extraction and Lagrangian-level model identification from collider data.
7+
---
8+
9+
10+
{% include gsoc_project.ext %}

‎_gsocproposals/2026/proposal_HEPSIM1.md‎

Lines changed: 47 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,47 @@
1+
---
2+
title: Physics-informed neural network shape optimization
3+
layout: gsoc_proposal
4+
project: SPINN
5+
year: 2026
6+
organization:
7+
- Alabama
8+
- Cerium
9+
10+
---
11+
12+
## Description
13+
In many engineering and physics domains, the shape or geometry of an object directly influences its performance with regard to some metric. Simple examples include the shape of an airplane wing affecting its lift-to-drag ratio, or the geometry/structure of a bridge influencing the maximum load weight which can safely traverse it. “Shape optimization” refers to the task of identifying the ideal shape/geometry of such an object which can maximize or minimize a metric of interest with respect to that object. This project looks to leverage Physics Informed Neural Networks (PINN) and Coordinate Projection Networks (also called shape networks in the literature) to develop machine learning architectures which can quickly and efficiently perform this task.
14+
15+
## Duration
16+
Total project length: 175 hours.
17+
18+
## Task ideas
19+
* Identify relevant use-case scenario and develop a PINN to accurately model underlying physics.
20+
* Perform model optimization of both PINN and coordinate projection network.
21+
22+
## Expected results:
23+
* PINN model with demonstrated ability to perform shape optimization.
24+
25+
## Difficulty level
26+
Advanced
27+
28+
## Requirements
29+
* Essential: Experience in Python, PyTorch (or related ML framework), and machine learning are required.
30+
* Preferred: Familiarity with basic physics concepts (classical mechanics, electromagnetics, etc.) at an undergraduate level.
31+
* Will help you stand out: Experience performing physics simulations using standard simulation software (COMSOL, ANSYS, etc.) or an equivalent open source framework, and familiarity with associated concepts (numerical methods techniques, how to handle boundary conditions, etc.).
32+
33+
<!---
34+
## Test
35+
Please use [this link](https://docs.google.com/document/d/10tITU-OJDtaZkLmXpoCpzIM55XcV4Z_7ykS7N9tyISA/edit?usp=sharing) to access the test for this project.
36+
--->
37+
38+
## Mentors
39+
* [Dale Julson](mailto:ml4-sci@cern.ch) (Cerium Labs)
40+
* [Eric Reinhardt](mailto:ml4-sci@cern.ch) (University of Alabama)
41+
* [Dinesh Ramakrishnan](mailto:ml4-sci@cern.ch) (University of Alabama)
42+
43+
Please **DO NOT** contact mentors directly by email. Instead, please email [ml4-sci@cern.ch](mailto:ml4-sci@cern.ch) with Project Title and **include your CV** and **test results**. The mentors will then get in touch with you.
44+
45+
## Links
46+
* [Paper 1](https://www.nature.com/articles/s41598-024-57137-4)
47+
* [Paper 2](https://arxiv.org/abs/2201.05624)

‎_gsocproposals/2026/proposal_HEPSIM2.md‎

Lines changed: 47 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,47 @@
1+
---
2+
title: Physics-informed neural network shape optimization
3+
layout: gsoc_proposal
4+
project: SPINN
5+
year: 2026
6+
organization:
7+
- Alabama
8+
- Cerium
9+
10+
---
11+
12+
## Description
13+
In many engineering and physics domains, the shape or geometry of an object directly influences its performance with regard to some metric. Simple examples include the shape of an airplane wing affecting its lift-to-drag ratio, or the geometry/structure of a bridge influencing the maximum load weight which can safely traverse it. “Shape optimization” refers to the task of identifying the ideal shape/geometry of such an object which can maximize or minimize a metric of interest with respect to that object. This project looks to leverage Physics Informed Neural Networks (PINN) and Coordinate Projection Networks (also called shape networks in the literature) to develop machine learning architectures which can quickly and efficiently perform this task.
14+
15+
## Duration
16+
Total project length: 175 hours.
17+
18+
## Task ideas
19+
* Identify relevant use-case scenario and develop a PINN to accurately model underlying physics.
20+
* Perform model optimization of both PINN and coordinate projection network.
21+
22+
## Expected results:
23+
* PINN model with demonstrated ability to perform shape optimization.
24+
25+
## Difficulty level
26+
Advanced
27+
28+
## Requirements
29+
* Essential: Experience in Python, PyTorch (or related ML framework), and machine learning are required.
30+
* Preferred: Familiarity with basic physics concepts (classical mechanics, electromagnetics, etc.) at an undergraduate level.
31+
* Will help you stand out: Experience performing physics simulations using standard simulation software (COMSOL, ANSYS, etc.) or an equivalent open source framework, and familiarity with associated concepts (numerical methods techniques, how to handle boundary conditions, etc.).
32+
33+
<!---
34+
## Test
35+
Please use [this link](https://docs.google.com/document/d/10tITU-OJDtaZkLmXpoCpzIM55XcV4Z_7ykS7N9tyISA/edit?usp=sharing) to access the test for this project.
36+
--->
37+
38+
## Mentors
39+
* [Dale Julson](mailto:ml4-sci@cern.ch) (Cerium Labs)
40+
* [Eric Reinhardt](mailto:ml4-sci@cern.ch) (University of Alabama)
41+
* [Dinesh Ramakrishnan](mailto:ml4-sci@cern.ch) (University of Alabama)
42+
43+
Please **DO NOT** contact mentors directly by email. Instead, please email [ml4-sci@cern.ch](mailto:ml4-sci@cern.ch) with Project Title and **include your CV** and **test results**. The mentors will then get in touch with you.
44+
45+
## Links
46+
* [Paper 1](https://www.nature.com/articles/s41598-024-57137-4)
47+
* [Paper 2](https://arxiv.org/abs/2201.05624)

‎_gsocproposals/2026/proposal_HEPSIM3.md‎

Lines changed: 75 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,75 @@
1+
---
2+
title: ML-Based Simulation Bias Analysis: Pythia vs Herwig vs Data
3+
layout: gsoc_proposal
4+
project: HEPSIM
5+
year: 2026
6+
organization:
7+
- Alabama
8+
- Fermilab
9+
- Rutgers
10+
11+
---
12+
13+
## Description
14+
Different Monte Carlo event generators, most notably [Pythia](https://pythia.org/) and [Herwig](https://herwig.hepforge.org/), employ distinct physics models for parton showering, hadronization, and the underlying event. These modeling choices lead to systematic differences in predicted distributions that directly impact physics measurements and searches for new phenomena at the [LHC](https://home.cern/science/accelerators/large-hadron-collider). Quantifying these inter-generator biases, and their discrepancies with data, is crucial for assigning robust systematic uncertainties.
15+
This project applies **machine learning classifiers** to perform a systematic, high-dimensional comparison of Pythia, Herwig, and (where available) unfolded experimental data. Rather than examining observables one at a time, ML classifiers can detect subtle multi-dimensional correlations that distinguish generators. The classifier output itself, and the features driving its decisions, reveal which regions of phase space carry the largest modeling uncertainties.
16+
The student will train classifiers (boosted decision trees, neural networks, or similar) to discriminate between generator samples and/or data, then extract interpretable information about the nature and location of biases using techniques such as SHAP values, feature importance, and learned reweighting.
17+
18+
19+
20+
21+
## Duration
22+
Total project length: 175/350 hours.
23+
24+
## Task ideas
25+
* Prepare matched Pythia and Herwig samples for standard LHC processes (e.g., Z+jets, tt̄, dijet)
26+
* Compute a comprehensive set of observables using the jet observable library (HEPSIM2) or equivalent
27+
* Train binary classifiers (BDT, neural network) to discriminate Pythia vs Herwig events
28+
* Extract feature importances and SHAP values to identify which observables and phase-space regions drive the largest inter-generator differences
29+
* Where public unfolded data is available, extend the analysis to MC-vs-data comparison
30+
* Explore the use of the classifier output as a learned reweighting function and compare with the symbolic regression approach (HEPSIM1)
31+
* Quantify the resulting uncertainty envelopes for key distributions
32+
33+
34+
35+
36+
## Expected results:
37+
* Trained ML classifiers that quantify Pythia–Herwig and MC–data discrepancies
38+
* Interpretability analysis (SHAP, feature importance) identifying the leading sources of simulation bias
39+
* Comparison of ML-derived bias estimates with traditional uncertainty prescriptions
40+
* Open-source code and documentation
41+
42+
43+
44+
45+
## Difficulty level
46+
Intermediate/Advanced
47+
48+
## Requirements
49+
* Python, PyTorch or TensorFlow, scikit-learn
50+
* Experience with ML classification and interpretability techniques
51+
* Some familiarity with HEP simulation tools or willingness to learn
52+
* Ability to work independently on a research project
53+
54+
55+
56+
57+
<!---
58+
## Test
59+
Please use [this link](https://docs.google.com/document/d/10tITU-OJDtaZkLmXpoCpzIM55XcV4Z_7ykS7N9tyISA/edit?usp=sharing) to access the test for this project.
60+
--->
61+
62+
## Mentors
63+
* [Steve Mrenna](mailto:ml4-sci@cern.ch) (Fermilab)
64+
* [Konstantin Matchev](mailto:ml4-sci@cern.ch) (University of Alabama)
65+
* [Tony Menzo](mailto:ml4-sci@cern.ch) (University of Alabama + Fermilab)
66+
* [Ian Pang](mailto:ml4-sci@cern.ch) (Rutgers University)
67+
68+
Please **DO NOT** contact mentors directly by email. Instead, please email [ml4-sci@cern.ch](mailto:ml4-sci@cern.ch) with Project Title and **include your CV** and **test results**. The mentors will then get in touch with you.
69+
70+
## Links
71+
* [Pythia 8](https://pythia.org/)
72+
* [Herwig 7](https://herwig.hepforge.org/)
73+
* [SHAP Values](https://arxiv.org/abs/1705.07874)
74+
* [Classifier-based reweighting](https://arxiv.org/abs/2007.11586)
75+
* [The Pareto frontier of resilient jet tagging](https://arxiv.org/pdf/2509.19431)

‎_gsocproposals/2026/proposal_HEPSIM4.md‎

Lines changed: 82 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,82 @@
1+
---
2+
title: Agentic Exclusion Limit Extraction
3+
layout: gsoc_proposal
4+
project: HEPSIM
5+
year: 2026
6+
organization:
7+
- Alabama
8+
- Fermilab
9+
- Rutgers
10+
11+
---
12+
13+
## Description
14+
Thousands of BSM searches at the [LHC](https://home.cern/science/accelerators/large-hadron-collider) have produced exclusion limits archived across ATLAS/CMS papers and databases like [SModelS](https://smodels.github.io/). For a theorist proposing a new model, determining what parameter space is already ruled out requires extensive manual literature searches, cross-referencing final states, extracting digitized limit curves, and mapping them onto the model of interest — a tedious, error-prone process that represents a major bottleneck in phenomenological research. This project aims to automate and generalize that approach using agentic programming.
15+
This project builds an **agentic system** within the [HEPTAPOD](https://github.com/tonymenzo/heptapod) framework that automates this process. Given a BSM model specification (particle content, decay channels, production modes), the agent searches the literature and public databases, identifies relevant analyses, extracts applicable bounds, and compiles a structured summary with references and applicability caveats.
16+
17+
18+
19+
20+
21+
## Duration
22+
Total project length: 350 hours.
23+
24+
## Task ideas
25+
* Survey existing limit databases and formats (HEPData, SModelS, ATLAS/CMS auxiliary material)
26+
* Implement HEPTAPOD tool modules for querying limit databases, parsing exclusion contours, and matching experimental final states to a user-specified model
27+
* Build agent workflows that autonomously retrieve and compile limits given a model specification
28+
* Add validation checks flagging when published analysis assumptions may not apply
29+
* Test on benchmark models (simplified SUSY, leptoquarks) where the correct relevant limits are known
30+
* Implement audit trails recording every search query, database hit, and reasoning step
31+
* Explore whether the agent can identify coverage gaps, parameter regions where no published search applies
32+
33+
34+
35+
36+
37+
## Expected results:
38+
* Working agentic system integrated into HEPTAPOD for automated limit compilation
39+
* Validation against at least two well-characterized BSM scenarios
40+
* Transparent audit trail and decision log
41+
* Open-source code with documentation and examples
42+
43+
44+
45+
46+
47+
## Difficulty level
48+
Advanced
49+
50+
## Requirements
51+
* Strong Python skills, experience with at least one ML/AI framework
52+
* Familiarity with LLM APIs and tool-use / function-calling paradigms
53+
* Interest in BSM phenomenology and limit-setting concepts
54+
* Ability to work independently with significant design freedom
55+
56+
57+
58+
59+
60+
<!---
61+
## Test
62+
Please use [this link](https://docs.google.com/document/d/10tITU-OJDtaZkLmXpoCpzIM55XcV4Z_7ykS7N9tyISA/edit?usp=sharing) to access the test for this project.
63+
--->
64+
65+
## Mentors
66+
* [Steve Mrenna](mailto:ml4-sci@cern.ch) (Fermilab)
67+
* [Konstantin Matchev](mailto:ml4-sci@cern.ch) (University of Alabama)
68+
* [Alex Roman](mailto:ml4-sci@cern.ch) (University of Alabama)
69+
* [Tony Menzo](mailto:ml4-sci@cern.ch) (University of Alabama + Fermilab)
70+
* [Ian Pang](mailto:ml4-sci@cern.ch) (Rutgers University)
71+
72+
Please **DO NOT** contact mentors directly by email. Instead, please email [ml4-sci@cern.ch](mailto:ml4-sci@cern.ch) with Project Title and **include your CV** and **test results**. The mentors will then get in touch with you.
73+
74+
## Links
75+
* [HEPTAPOD Repository](https://github.com/tonymenzo/heptapod)
76+
* [HEPTAPOD Paper](https://arxiv.org/abs/2512.15867)
77+
* [Orchestral AI](https://arxiv.org/abs/2601.02577)
78+
* [CheckMATE](https://arxiv.org/abs/1312.2591)
79+
* [GAMBIT](https://arxiv.org/abs/2512.14109)
80+
* [SModelS](https://smodels.github.io/)
81+
* [pyhf](https://pyhf.readthedocs.io/en/v0.7.6/)
82+

‎_gsocproposals/2026/proposal_HEPSIM5.md‎

Lines changed: 85 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,85 @@
1+
---
2+
title: Agentic Lagrangian Extraction from the Literature
3+
layout: gsoc_proposal
4+
project: HEPSIM
5+
year: 2026
6+
organization:
7+
- Alabama
8+
- Fermilab
9+
- Rutgers
10+
11+
---
12+
13+
## Description
14+
Hundreds of BSM Lagrangians have been proposed in the literature (extended Higgs sectors, leptoquarks, SUSY variants, dark matter portals, EFT extensions at various operator dimensions). When a researcher wants to simulate a particular scenario, they must locate the Lagrangian in the literature, translate it into a [FeynRules](https://feynrules.irmp.ucl.ac.be/) model file, and validate the implementation. This literature-to-simulation pipeline is a significant manual bottleneck.
15+
This project builds an **agentic system** within the [HEPTAPOD](https://feynrules.irmp.ucl.ac.be/) framework that automates this workflow. Given a BSM scenario description, the agent searches the literature, identifies the relevant Lagrangian and conventions, generates a syntactically correct FeynRules `.fr` model file, and validates it, checking particle content, symmetry properties, and where possible reproducing known cross sections or decay widths using HEPTAPOD's existing FeynRulesToUFOTool and downstream event generation tools.
16+
17+
18+
19+
20+
21+
22+
## Duration
23+
Total project length: 350 hours.
24+
25+
## Task ideas
26+
* Survey publicly available FeynRules model files (model database, HEPTAPOD examples) to catalog conventions and patterns
27+
* Implement HEPTAPOD tool modules for searching the HEP literature (INSPIRE, arXiv), extracting Lagrangian terms, and generating FeynRules .fr files
28+
* Build agent workflows that go from a scenario description to a validated model file, using HEPTAPOD's existing tools for UFO compilation and event generation
29+
* Implement validation: particle spectrum consistency, gauge invariance, reproduction of known decay widths or cross sections where available
30+
* Test against benchmark models already in the FeynRules database (scalar leptoquarks, 2HDM, simplified dark matter)
31+
* Implement audit trail recording literature sources, extracted terms, and model-building decisions
32+
33+
34+
35+
36+
37+
38+
## Expected results:
39+
* Working agentic system integrated into HEPTAPOD for Lagrangian extraction and model file generation
40+
* Validation against at least two models with known FeynRules implementations
41+
* Audit trail documenting the agent's reasoning from literature to model file
42+
* Open-source code with examples
43+
44+
45+
46+
47+
48+
49+
## Difficulty level
50+
Advanced
51+
52+
## Requirements
53+
* Strong Python skills, experience with AI/ML frameworks
54+
* Familiarity with LLM tool-use and agentic programming patterns
55+
* Interest in particle physics theory (Lagrangians, EFT, BSM), coursework or reading a plus
56+
* Exposure to Mathematica or FeynRules helpful
57+
* Ability to work independently on an open-ended research project
58+
59+
60+
61+
62+
63+
64+
<!---
65+
## Test
66+
Please use [this link](https://docs.google.com/document/d/10tITU-OJDtaZkLmXpoCpzIM55XcV4Z_7ykS7N9tyISA/edit?usp=sharing) to access the test for this project.
67+
--->
68+
69+
## Mentors
70+
* [Steve Mrenna](mailto:ml4-sci@cern.ch) (Fermilab)
71+
* [Konstantin Matchev](mailto:ml4-sci@cern.ch) (University of Alabama)
72+
* [Alex Roman](mailto:ml4-sci@cern.ch) (University of Alabama)
73+
* [Tony Menzo](mailto:ml4-sci@cern.ch) (University of Alabama + Fermilab)
74+
* [Ian Pang](mailto:ml4-sci@cern.ch) (Rutgers University)
75+
76+
Please **DO NOT** contact mentors directly by email. Instead, please email [ml4-sci@cern.ch](mailto:ml4-sci@cern.ch) with Project Title and **include your CV** and **test results**. The mentors will then get in touch with you.
77+
78+
## Links
79+
* [HEPTAPOD Repository](https://github.com/tonymenzo/heptapod)
80+
* [HEPTAPOD Paper](https://arxiv.org/abs/2512.15867)
81+
* [Orchestral AI](https://arxiv.org/abs/2601.02577)
82+
* [The Denario project: Deep knowledge AI agents for scientific discovery](https://arxiv.org/abs/2510.26887)
83+
* [FeynRules](https://feynrules.irmp.ucl.ac.be/)
84+
* [FeynRules Model Database](https://feynrules.irmp.ucl.ac.be/wiki/ModelDatabaseMainPage)
85+
* [INSPIRE](https://inspirehep.net/)

‎images/HEP_SIM.png‎

336 KB
Loading

0 commit comments

Comments
 (0)