Energy-weighted message passing: an infra-red and collinear safe graph neural network algorithm

Abstract

Hadronic signals of new-physics origin at the Large Hadron Collider can remain hidden within the copiously produced hadronic jets. Unveiling such signatures require highly performant deep-learning algorithms. We construct a class of Graph Neural Networks (GNN) in the message-passing formalism that makes the network output infra-red and collinear (IRC) safe, an important criterion satisfied within perturbative QCD calculations. Including IRC safety of the network output as a requirement in the construction of the GNN improves its explainability and robustness against theoretical uncertainties in the data. We generalise Energy Flow Networks (EFN), an IRC safe deep-learning algorithm on a point cloud, defining energy weighted local and global readouts on GNNs. Applying the simplest of such networks to identify top quarks, W bosons and quark/gluon jets, we find that it outperforms state-of-the-art EFNs. Additionally, we obtain a general class of graph construction algorithms that give structurally invariant graphs in the IRC limit, a necessary criterion for the IRC safety of the GNN output.

IRC-safe Graph Autoencoder for an unsupervised anomaly detection

“…where h a are the latent node features. Similar to the graph readout in a classification scenario [25], this is an IRC safe representation of the jet. The distribution of the individual components of the two-dimensional graph representation are shown in Fig.…”

“…This section presents a brief overview of the IRC safe Energy-weighted Message passing algorithm [25]. Similar to message-passing networks like the Dynamic Graph Convolutional Neural Network (DGCNN) [27] that extract local features beyond the global feature extraction via point cloud-based architectures such as deep-sets [28] and PointNet [29,30], it generalises Energy Flow Networks [31,32], an IRC safe feature extraction on point clouds to ensure local feature extraction.…”

“…It seems prudent to avoid mistakes of the past: in this paper, we devise an IRC-safe Graph Neural Network (GNN) autoencoder algorithm, employing an Energy-Weighted Message-Passing Network (EMPN) [25] for unsupervised anomaly detection. While the IRCsafe loss function is the primary observable of our autoencoder, we also study the latent space (graph) representation structure as a motivating tool for new physics discrimination [5,26] and highlight the relations to known and more "traditional" IRC-safe observables.…”

An Efficient Lorentz Equivariant Graph Neural Network for Jet Tagging

“…Based on the theory and phenomenology of jet physics, many expert-designed high-level jet substructure observables are constructed for jet tagging [2][3][4][5][6][7]. Using different jet representations, various deep learning approaches have been investigated in recent years: Multilayer perceptrons (MLPs) can be trained on a collection of jet-level observables [8][9][10][11][12][13]; 2D convolutional neural networks (CNNs) are applied to jet images [14][15][16][17][18][19][20][21][22][23][24][25][26][27]; MLPs, 1D CNNs, and recurrent neural networks are used to process a jet as a sequence of its constituent particles [28][29][30][31][32][33][34][35][36]; Graph neural networks (GNNs) are developed for the "particle cloud", i.e., an unordered set of particles [37][38][39][40][41][42][43][44][45][46]…”

Improved Constraints on Effective Top Quark Interactions using Edge Convolution Networks

“…Secondly, we can resolve to a more comprehensive extraction of information from experimental data. Such strategies are highlighted in the recent resurgence of machine learning (ML) applications to particle physics [26][27][28][29][30][31][32][33]. 'Traditional' collider observables such as transverse momenta, angles and (pseudo)rapidities, alongside rectangular cuts on these, might not fully capture the exclusion potential when all ad hoc modifications of correlations are considered, which is the key motivation of the EFT approach (in particular this extends to the inclusion of systematic uncertainties [34]).…”