Reproduce “Panda Diplomacy: Foundation Model Pre-training across Particle Imaging Detectors for High Energy and Nuclear Physics” on OpenResearch

About this paper

Samuel Young, César Jesús-Valls, Kazuhiro Terao

The Panda V2 framework introduces a generalized point-native architecture and self-distillation pre-training for foundation models, enabling its application across LArTPC, collider TPC, and water Cherenkov particle imaging detectors. This method achieved reconstruction performance comparable to or exceeding specialized baselines with a 70-fold to 1000-fold reduction in required labeled data, while also showing emergent physically meaningful latent representations.

Key insights

• A single foundation model architecture and pre-training objective can learn generalizable and semantically organized representations across fundamentally different particle detector modalities.
• The self-supervised pre-training enables substantial data efficiency, allowing state-of-the-art reconstruction performance with orders of magnitude fewer labeled examples for fine-tuning.
• Physically meaningful quantities, such as particle trajectory ordering ("arrow of time") and kinematic properties like transverse momentum, spontaneously emerge and are linearly encoded within the frozen latent representations.

Summary by alphaXiv.