TPCpp-10M: Simulated proton–proton collisions in Time Projection Chamber for AI Foundation Models
收藏资源简介：
TPCpp-10M Dataset A large-scale particle physics dataset designed for machine learning applications in high-energy physics research. Dataset Overview The TPCpp-10M dataset contains simulated particle collision events from a Time Projection Chamber (TPC) detector, commonly used in particle physics experiments. This dataset is specifically designed for: Foundation model training for particle and nuclear physics applications Particle track reconstruction Particle identification Signal vs. noise discrimination Dataset Characteristics Total Events: over 10 million simulated proton-proton collision events at 200GeV Input Data Format: Structured spacepoint data with 4D coordinates (energy and x,y,z positions) Labels: including track IDs, particle types, and noise classification Splits: unlabeled (for pretrain): 10M labeled train: 70k validation: 13k test: 7k Data type: Numpy compressed data (.npz) Data Components Spacepoints: 4D spacepoints with energy and spatial coordinates Track IDs (only for labeled data): Unique identifiers for particle trajectories Particle ID Labels (only for labeled data): Classification labels for different particle types Noise Tags (only for labeled data): Binary labels distinguishing signal from detector noise Scripts demo.ipynb: data loading, interactive plots, statistics of the dataset plot.py: plot functions Publication FM4NPP: A Scaling Foundation Model for Nuclear and Particle Physics (https://arxiv.org/abs/2508.14087) Files TPCpp-10M.zip unlabeled spacepoints_[000-099].npz labeled: train spacepoints_[000-006].npz track_ids_[000-006].npz pid_labels_[000-006].npz noise_tags_[000-006].npz validation spacepoints.npz track_ids.npz pid_labels.npz noise_tags.npz test spacepoints.npz track_ids.npz pid_labels.npz noise_tags.npz scripts demo.ipynb plot.py TPCpp-10M_labeled_test.zip: The same as labeled/test in TPCpp-10M.zip for a quick peek TPCpp-10M_scripts.zip: The same as scripts in TPCpp-10M.zip for a quick peek
TPCpp-10M 数据集 一款面向高能物理研究中机器学习应用的大规模粒子物理数据集。 ### 数据集概览 TPCpp-10M 数据集包含来自时间投影室（Time Projection Chamber，TPC）探测器的模拟粒子碰撞事件，这类探测器广泛应用于粒子物理实验。本数据集专为以下场景设计： 1. 面向粒子与核物理应用的基础模型训练 2. 粒子径迹重建 3. 粒子识别 4. 信号与噪声甄别 ### 数据集特性 - 总事件数：超过1000万次200GeV能量下的模拟质子-质子碰撞事件 - 输入数据格式：包含4维坐标（能量以及x、y、z空间位置）的结构化空间点数据 - 标签：包含径迹ID、粒子类型以及噪声分类 - 数据集划分： - 未标注（用于预训练）：1000万条 - 已标注： - 训练集：7万条 - 验证集：1.3万条 - 测试集：7000条 - 数据类型：Numpy压缩格式数据（.npz） ### 数据组成 - 空间点：包含能量与空间坐标的4维空间点数据 - 径迹ID（仅已标注数据可用）：粒子运动轨迹的唯一标识符 - 粒子识别标签（仅已标注数据可用）：不同粒子类型的分类标签 - 噪声标签（仅已标注数据可用）：区分信号与探测器噪声的二元标签 ### 脚本文件 - demo.ipynb：用于数据加载、交互式绘图以及数据集统计分析 - plot.py：绘图功能函数 ### 相关出版物 FM4NPP：一款面向核与粒子物理的规模化基础模型（https://arxiv.org/abs/2508.14087） ### 文件结构 TPCpp-10M.zip ├─ unlabeled │ └─ spacepoints_[000-099].npz ├─ labeled │ ├─ train │ │ ├─ spacepoints_[000-006].npz │ │ ├─ track_ids_[000-006].npz │ │ ├─ pid_labels_[000-006].npz │ │ └─ noise_tags_[000-006].npz │ ├─ validation │ │ ├─ spacepoints.npz │ │ ├─ track_ids.npz │ │ ├─ pid_labels.npz │ │ └─ noise_tags.npz │ └─ test │ ├─ spacepoints.npz │ ├─ track_ids.npz │ ├─ pid_labels.npz │ └─ noise_tags.npz └─ scripts ├─ demo.ipynb └─ plot.py 此外还提供两个用于快速预览的压缩包： - TPCpp-10M_labeled_test.zip：与TPCpp-10M.zip中labeled/test目录内容一致 - TPCpp-10M_scripts.zip：与TPCpp-10M.zip中scripts目录内容一致
