TPCpp-10M: Simulated proton-proton collisions in a time projection chamber for AI foundation models

Area of Science:

• Nuclear and particle physics
• Machine learning
• Foundation models

Background:

• Scientific foundation models offer potential for advancing nuclear and particle physics.
• Progress is hindered by limited open datasets, standardized tasks, and metrics.
• Specialized knowledge and software create barriers for interdisciplinary collaboration.

Purpose of the Study:

• Introduce a large, openly accessible dataset for foundation model training in particle physics.
• Facilitate interdisciplinary research between machine learning scientists and physicists.
• Enable systematic evaluation of foundation model adaptability on downstream tasks.

Main Methods:

• Generated 10 million simulated proton-proton collisions using Pythia.
• Processed simulations with Geant4 for realistic detector conditions (sPHENIX Time Projection Chamber).
• Provided dataset in NumPy format with 70,000 labeled examples for track finding, particle identification, and noise tagging.

Main Results:

• Established a large-scale, openly accessible dataset for foundation model research.
• Included downstream tasks for systematic evaluation of model adaptability.
• Ensured reproducibility of the simulation and reconstruction chain using the sPHENIX software stack.

Conclusions:

• The dataset serves as a common ground for interdisciplinary research in nuclear and high-energy physics.
• Accelerates progress in developing and evaluating foundation models for physics applications.
• Promotes collaboration by lowering barriers to entry for machine learning experts.