Speaker
Description
FM4NPP is a foundation model that aims to perform track finding and particle identification from sPHENIX TPC hits with one pretrained representation, adapted to each task by a head trained on a small fraction of labeled data. Its Mamba-2 backbone is pretrained to predict nearby hits at larger radii, then frozen while lightweight supervised heads are trained for downstream tasks. We outline an extension toward a detector-wide representation that would place silicon hits and calorimeter hits in a common sequence. The central challenge is to account for differences in geometry, resolution and hit multiplicity while choosing pretraining objectives that remain meaningful across detector subsystems. The programme explores latent prediction, alternative sequence backbones and heavy-ion data for scaling studies. Evaluation will distinguish the contribution of pretraining from that of the task-specific heads and examine the computational demands of larger events. These design and validation questions are relevant to ePIC's silicon tracking, calorimetry and particle identification.