CEPC上基于DeepSets模型的喷注标记算法研究

The Study of Jet Tagging Algorithm Based on DeepSets at CEPC

【目的】在希格斯玻色子被证实存在之后,精确测量希格斯玻色子性质以及检验电弱模型就成了高能物理领域今后一段时间内的重点工作,对喷注进行更加细致地研究,可以有效地提升测量和检验的精确度。【方法】本研究采用了新提出的、基于DeepSets的粒子流网络深度学习算法以及Z玻色子强衰变末态产生的喷注全模拟数据集,尝试对重味夸克喷注的标记算法进行研究。【结果】经过初步研究,与XGBoost等传统算法和DNN全连接网络相比,新算法性能提高约6%,使平均精度从80%提高到85%左右。【结论】这表明深度学习在高能物理实验中存在巨大的潜力,在粒子重建、事例分析等多方面有较好的应用场景,但还需更加深入地研究深度学习方法带来的影响,以提升深度学习的可信度和效果。

[Purpose]After the existence of the Higgs boson was confirmed,precision measurements of its properties and searches for new physics beyond the Standard Model become the focus of the field of high energy physics in the future,and better detection of the jet can effectively improve the precision and sensitivity.[Method]In this study,a newly proposed DeepSets-based deep learning algorithm,the Particle Flow network,and a collection of full simulation data sample of jets generated by the hadronic decay of the Z boson are used to study the tagging algorithm of heavy flavor jets.[Results]Compared with the CEPC baseline algorithm,the performance of the flavor tagging is improved by about 6%,i.e.,the average accuracy is increased from 80%to 85%.[Conclusion]This study shows that deep learning has great potential in high-energy physics experiments,and has wide application scenarios in particle reconstruction,physics analysis,and other aspects.However,more in-depth research on deep learning methods is needed to enhance the credibility and robustness.