量子完全圖神經網路在高能物理噴流辨識的應用及量子優勢之探討

Abstract

機器學習技術，尤其是深度神經網路，近年來逐漸成為高能物理領域當中分析的重要工具。這些方法已在多項應用中展現出驚人的效果且強大的潛力。隨著機器學習與量子運算的整合，一個特別的新研究領域「量子機器學習」逐漸成形，預期能在處理更複雜的資料時提供一定的優勢。

在本論文中，我們提出一種名為「量子完全圖神經網路（QCGNN）」的方法，它是一種建立在變分量子電路所設計的演算法，目標是處理完全圖上的學習任務。該模型透過參數化的量子邏輯閘達成資料編碼，並利用量子平行性的特性，期許在一些特別場景中能比古典方法帶來更有效率的運算。

為了比較，我們將此模型應用於大型強子對撞機（LHC）中的質子－質子碰撞事件，處理噴流分類的任務。在這個情境下，噴流是由高能夸克或膠子所產生的粒子集合，並可建構為完全圖的形式。由於噴流分類對理解粒子交互作用中有著關鍵性的影響，所以更快速且精準的模型來辨識不同來源的噴流可以有更好的分析結果。QCGNN 展現出能有效處理基於圖形結構的噴流資料的能力，有望在未來高亮度 LHC 實驗中發揮其應用潛力。

此外，我們也與古典深度學習模型進行比較分析，以評估 QCGNN 的效能表現。我們也在 IBM 的量子硬體上實作並測試此模型，以探討其在真實量子設備上的可行性。整體而言，QCGNN 不僅對圖形方法中常見的運算挑戰提出了解法，也展現出在量子運算與高能物理交叉領域中具有的研究潛力。

Machine learning techniques, especially deep neural networks, have gradually become important tools in the field of high-energy physics. These methods have shown encouraging results in various applications. In recent years, the combination of machine learning and quantum computing has led to the development of a new field of research field called the "Quantum Machine Learning", which is expected to offer advantages in handling increasingly complex data.

This dissertation proposes the Quantum Complete Graph Neural Network (QCGNN) designed for learning tasks on fully connected graphs, which is a VQC-based algorithm, where VQC stands for "Variational Quantum Circuit". The QCGNN makes use of parameterized quantum circuits and aims to benefit from quantum parallelism, potentially providing computational advantages compared to classical approaches.

We apply the QCGNN to the task of jet discrimination in proton-proton collisions at the Large Hadron Collider (LHC). In this context, jets, which are collections of particles originating from energetic partons, are represented as complete graphs. Since jet classification plays a crucial role in understanding particle interactions, an effective model is necessary to distinguish between different jet types. The QCGNN shows the ability to process graph-based jet data efficiently, which may be helpful for future analyses at the high-luminosity LHC.

In addition, we compare the QCGNN with classical deep learning models to evaluate its performance. We also test the implementation of QCGNN on IBM quantum hardware to examine its feasibility on real devices. The results suggest that the QCGNN has potential for further exploration in both quantum computing and high-energy physics research.