Speaker
Description
We present performance results for charged-particle tracking on H1 and ePIC simulated collision data as well as archived LEP data using the FM4NPP [1,2] foundation model. This study is part of a DOE Genesis project on multimodal and multifacility applications of FM4NPP, extending the model beyond its original sPHENIX time projection chamber setting to silicon trackers and electron colliders. FM4NPP separates representation learning from downstream adaptation: a backbone pretrained on low-level detector measurements is kept frozen while lightweight task-specific adapters learn to associate hits into track candidates. The results on e+e- data and e-h and e-A simulations are compared with standard reconstruction approaches and with the original FM4NPP results for a large sample of sPHENIX p+p simulated data [1,3]. The studies demonstrate how representations learned through self-supervision can transfer across detector geometries and collision environments to efficiently perform reconstruction tasks.
[1] D. Park et al., “FM4NPP: A Scaling Foundation Model for Nuclear and Particle Physics”, arXiv:2508.14087 (2025).
[2] Y. Ren et al., “A roadmap toward scaling, reasoning and self-evolving foundation models for nuclear and particle physics”, International Journal of Modern Physics A 41, 2650086 (2026). https://doi.org/10.1142/S0217751X26500867
[3] S. Li et al., “TPCpp-10M: Simulated proton-proton collisions in a Time Projection Chamber for AI Foundation Models”, arXiv:2509.05792 (2025).