The paper presents a new framework called BUFF (Boosted Decision Tree based Ultra-Fast Flow matching) that utilizes flow matching and gradient boosted tree (GBT) models to enable efficient high-dimensional data simulation for various tasks in high energy physics.
The key highlights are:
The authors adopt the conditional flow matching approach and integrate the usage of GBT models, creating a novel model called flowBDT. This allows them to overcome the limitations of traditional normalizing flow models in handling high-dimensional tabular data.
The flowBDT model demonstrates impressive performance on end-to-end fast simulation of high-level jet variables, achieving negligible inference time and fast training across multiple CPU cores compared to traditional flow matching.
The authors scale up the dimensionality of the simulation task, showing that flowBDT can still provide decent results for simulating hundreds of calorimeter cells within irregular geometry and jet constituents.
The conditional generation capability of flowBDT is explored, showcasing significant improvements in correlation matching for unfolding tasks compared to unconditional generation. The model can also refine calorimeter showers by using approximate shower information as conditions.
The authors identify key enhancements to the original flow matching approach, such as the use of higher-order ODE solvers, distinct GBT training objectives, and batch training strategies to further improve the efficiency and scalability of the framework.
Overall, the BUFF framework demonstrates the potential of leveraging GBT models in flow matching for fast and accurate high-dimensional data simulation in high energy physics applications.
To Another Language
from source content
arxiv.org
ข้อมูลเชิงลึกที่สำคัญจาก
by Cheng Jiang,... ที่ arxiv.org 04-30-2024
https://arxiv.org/pdf/2404.18219.pdfสอบถามเพิ่มเติม