点云量热器簇射生成中的跨几何迁移与模型崩溃
Cross-geometry transfer and model collapse in point cloud calorimeter shower generation
浏览论文内容
中文总结 AI 辅助
本研究探讨点云量热器簇射生成模型CaloClouds II的跨几何迁移能力，通过预训练与微调实现高效适配，并分析其模型崩溃现象。
中文摘要 AI 辅助
粒子簇射模拟是高能物理中的一项主要计算成本。Geant4等蒙特卡洛方法精度高但代价昂贵，而大多数机器学习替代方法则受限于特定探测器几何结构，每次设计变更都需要重新训练。我们研究了基于CaloClouds II的跨几何迁移学习，这是一种生成点云而非体素、并可投影到任意探测器读出上的生成模型。我们在国际大型探测器（ILD）中的光子簇射上进行了预训练，并适配到圆柱形CaloChallenge数据集3中的电子簇射。仅使用100个目标簇射，微调相比从头训练将到Geant4的几何平均Wasserstein距离降低了约51%。仅偏置微调（BitFit）在仅更新扩散网络17%参数的情况下，与完全微调的性能差距保持在5%以内。我们还通过在其自身生成的簇射上连续世代重新训练其归一化流和扩散模型，考察了CaloClouds II中的模型崩溃现象。
英文摘要
Particle shower simulation is a major computational cost in high-energy physics. Monte Carlo methods such as Geant4 are accurate but expensive, while most machine learning surrogates are tied to specific detector geometries and require retraining for each design change. We study cross-geometry transfer learning with CaloClouds II, a generative model that produces point clouds rather than voxels and can project onto arbitrary detector readouts. We pre-train on photon showers in the International Large Detector (ILD) and adapt to electron showers in the cylindrical CaloChallenge Dataset 3. With only 100 target showers, fine-tuning reduces the geometric mean Wasserstein distance to Geant4 by about 51% over training from scratch. Bias-only fine-tuning (BitFit) stays within 5% of full fine-tuning while updating only 17% of the diffusion network parameters. We also examine model collapse in CaloClouds II by retraining its normalising flow and diffusion model on its own generated showers across successive generations.
发表机构
- Universität Hamburg（汉堡大学）
- Deutsches Elektronen-Synchrotron DESY（德国电子同步加速器）
机构由 AI 辅助整理，请以论文原文为准。
