⚛️ high-energy experiments

Panda Diplomacy: Foundation Model Pre-training across Particle Imaging Detectors for High Energy and Nuclear Physics

The paper introduces "Panda Diplomacy," a generalizable point cloud self-distillation framework that enables foundation models to be pre-trained across diverse particle detector modalities (LArTPC, collider TPC, and water Cherenkov) with minimal architectural changes, achieving state-of-the-art performance in particle reconstruction and identification using orders of magnitude fewer labeled events than specialized baselines.

Original authors: Samuel Young, César Jesús-Valls, Kazuhiro Terao

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Samuel Young, César Jesús-Valls, Kazuhiro Terao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep inside massive underground caverns and atop high-energy colliders, scientists are trying to solve a puzzle that has existed since the birth of the universe: what are the fundamental building blocks of matter, and how do they interact? To find the answers, they build enormous detectors that act like giant, three-dimensional cameras. When a subatomic particle zips through these machines, it leaves behind a trail of energy, much like a spark flying through a dark room. The challenge for physicists is to look at the millions of scattered points of light and energy in these images and figure out exactly what happened: which particles were there, where they started, which way they were moving, and how fast. For decades, the software used to interpret these images has been custom-built for each specific machine. A detector designed to catch particles in liquid argon requires a different set of rules than one designed to catch light in a giant tank of water. This means that every time a new experiment is built, scientists have to start from scratch, teaching a new computer program how to see, even though the basic physics of the particles remains the same.

A team of researchers has now demonstrated that this approach is no longer necessary. They have developed a new method, which they call "Panda Diplomacy," that allows a single computer model to learn how to see particles across three completely different types of detectors. The researchers took a powerful artificial intelligence system and trained it on raw data from three distinct sources: a detector filled with liquid argon that captures ionization trails, a collider detector that tracks particles moving through a magnetic field, and a water-based detector that records flashes of light. Crucially, they did not teach the computer the specific rules of any one machine. Instead, they let it learn by looking at the raw patterns of points in space, asking it to predict missing parts of the image based on what it could see. This process, known as self-distillation, allowed the model to build a universal understanding of how particles move and interact, regardless of the medium they travel through.

The results of this experiment are striking. When the researchers tested this single, pre-trained model on new data, they found that it could be adapted to perform complex tasks with a tiny fraction of the usual training data. In the past, teaching a computer to identify and group particles in a specific detector required millions of labeled examples, where humans had manually drawn the correct paths for the machine to learn from. With this new approach, the model achieved state-of-the-art performance using only one thousand labeled examples. For the liquid argon detector, the model matched the accuracy of previous systems that had been trained on a million labeled events, but it did so with one thousand times less human supervision. In the collider detector, it matched the performance of a system trained on seventy thousand examples while using seventy times fewer labeled events. This suggests that the model had learned a deep, reusable language of particle physics that applies across different technologies.

Beyond simply recognizing particles, the researchers discovered that the model had learned to understand the physical laws governing them, even without being explicitly told what those laws were. When they examined the internal "thoughts" of the model, they found that it had organized the data in ways that corresponded to real physical properties. In the liquid argon detector, the model could identify the direction a particle was traveling, effectively creating an "arrow of time" that showed the sequence of events, even though the raw data did not contain any time information. In the collider detector, the model learned to recognize the curvature of a particle's path, which is directly related to its momentum. In the water detector, it could reconstruct the position and direction of particles with high precision. These findings indicate that the model had not just memorized patterns but had internalized the underlying geometry and causality of particle interactions.

This work represents a significant shift in how scientists approach data analysis in high-energy physics. By proving that a single, general-purpose model can learn from diverse detector types and be adapted with minimal effort, the researchers have opened the door to a more unified way of studying the universe. The model, which they call Panda V2, acts as a foundational tool that can be applied to new experiments as soon as they come online, reducing the need for years of custom software development. While the model still requires some fine-tuning for specific tasks and does not yet match the precision of the most highly optimized, specialized algorithms used by major collaborations, it demonstrates that a general approach is not only possible but highly efficient. The ability to learn from three different sensing mechanisms with a single framework suggests that the future of particle physics may rely less on building unique tools for every experiment and more on developing versatile systems that can understand the fundamental language of matter, no matter how it is observed.

Technical Summary: Panda Diplomacy

Problem Statement

Foundation models in particle and nuclear physics have largely been constrained by detector-specific architectures and pre-training objectives. Existing approaches typically rely on specialized representations or task-specific goals tied to a single sensing modality (e.g., liquid argon time projection chambers, collider TPCs, or water Cherenkov detectors). This specificity limits the reuse of learned representations across fundamentally different experimental setups, which vary significantly in geometry, sensing mechanisms, and physical structures. While recent work demonstrated sensor-level foundation models for a single detector type, it remained unknown whether a unified representation-learning mechanism could operate effectively across qualitatively different particle detectors with minimal architectural changes.

Methodology

The authors introduce Panda V2, an updated backbone architecture and a unified pre-training framework designed to be applied across three distinct detector modalities:

Liquid Argon Time Projection Chambers (LArTPC) (PILArNet-M dataset).
Collider TPCs (sPHENIX TPC, TPCpp-10M dataset).
Water Cherenkov Detectors (Super-Kamiokande-like, WAND dataset).

Architecture (Panda V2 Encoder)

The model utilizes a four-stage point-native encoder based on Point Transformer V3 and LitePT. Key architectural features include:

Sparse 3D Convolutions: Used at high spatial resolutions.
Local Self-Attention: Employed at coarser scales.
Processing Flow: The network processes native voxelization, pools by a factor of two, switches from convolution to attention at the same resolution, pools by a further factor of four, and concludes with a final attention stage.
Multi-scale Features: Features are upsampled and concatenated to produce contextual features for each input point.
Parameter Efficiency: The architecture removes non-residual pre-layer normalization in attention-only blocks and utilizes 3D rotary positional embeddings (PointRoPE).

Pre-training Objective: Observable Prototype Self-Distillation

The framework employs a self-distillation objective where an Exponential Moving Average (EMA) teacher and a student encode independently augmented views of the same image. The student predicts the teacher's assignments over learned prototypes.

Masking Strategy: Following DOS, masked measurements are removed rather than explicitly represented; only observable locations are distilled.
Generalization: The objective contains no detector-specific propagation or reconstruction rules. Strong generic spatial augmentations (rotations, translations, jittering) are applied.
Hyperparameters: Only hyperparameters tied directly to the numerical and spatial scale of the input (e.g., grid size, mask size) are adjusted between modalities.

Downstream Adaptation

Downstream tasks utilize parameter-efficient fine-tuning with only 1,000 labeled events per task:

Semantic Segmentation: Linear classifiers with Low-Rank Adaptation (LoRA).
Particle Clustering: A lightweight two-stage Graph Neural Network (GNN) with LoRA.
Single-Ring Water Cherenkov: Linear probes on frozen event representations or equivariant coordinate-feature moments.
Multi-Ring Water Cherenkov: The "Panda Detector" module (DETR-style set prediction) with LoRA.

Key Results

1. Data Efficiency and Performance

Panda V2 matches or exceeds specialized foundation model baselines trained with orders of magnitude more supervision:

LArTPC (PILArNet-M): With only 1,000 labeled events, Panda V2 achieves a particle clustering Adjusted Rand Index (ARI) of 0.973, matching the performance of Panda V1 trained on 1 million labeled events. For semantic segmentation, it reaches a macro-F1 of 0.976, comparable to Panda V1 trained on 100,000 events.
Collider TPC (TPCpp-10M): Using 1,000 labeled events, Panda V2 achieves a track clustering ARI of 0.943, closely matching FM4NPP's result of 0.945 obtained with 70,000 labeled events. It also outperforms FM4NPP in particle identification (0.838 vs. 0.713) and noise tagging (0.942 vs. 0.912) macro-F1 scores.
Water Cherenkov (WAND): Lightweight probes reconstruct single-particle kinematics (vertex, direction, momentum) with 1,000 labels. While performance is lower than highly optimized algorithms like fiTQun and APFit, it demonstrates that useful detector-level observables are accessible via inexpensive adaptation. Multi-ring reconstruction recalls 80.4% (1-ring), 44.4% (2-ring), and 96.7% (≥3-ring) events.

2. Emergent Physical Structure

The authors demonstrate that the frozen representations encode physically meaningful latent structures without explicit supervision:

Arrow of Time (LArTPC): A single learned vector, when projected onto per-point features, orders points along particle trajectories (Spearman $\rho = 0.89$ ), recovering an emergent "arrow of time" despite the absence of timing information in the input.
Transverse Momentum (sPHENIX): A latent direction encodes track curvature and transverse momentum ( $p_T$ ) with a Spearman $\rho = 0.957$ .
Rotational Invariance: The self-distillation objective enforces weak rotational invariance, with cosine similarity distributions remaining $\geq 0.98$ across rotations.

Significance and Claims

The paper claims that the machinery required for foundation-model pre-training on sensor-level particle reconstruction can be general rather than detector-specific.

Unified Framework: It establishes that a single architecture and self-distillation objective can be independently pre-trained on three qualitatively different detector modalities with minimal changes.
Label Efficiency: The framework supports full reconstruction (segmentation, clustering, property extraction) from as few as 1,000 labeled events, significantly reducing the dependency on large labeled datasets.
Physical Interpretability: The learned representations naturally organize around physical concepts like particle causality ("arrow of time") and momentum, suggesting that self-supervised learning on raw sensor data can recover fundamental physical structures.

Limitations and Future Work

The authors explicitly note several unresolved questions:

Scaling: The study does not characterize scaling laws regarding model size, pre-training data volume, or compute resources.
Transferability: It is not yet tested whether representations learned on one detector type transfer directly to another (cross-domain transfer) or from simulation to real data.
Performance Gap: In Water Cherenkov reconstruction, the method currently lags behind highly optimized, domain-specific algorithms (e.g., fiTQun), indicating room for improvement in specific detector applications.

The authors plan to address scaling and transferability questions in a forthcoming extended update.