Source-linked AI summary
Graph Neural Networks in Particle Physics
Jonathan Shlomi, Peter Battaglia, Jean-Roch Vlimant
TL;DR
Particle-physics data often have set- and graph-like structure, motivating methods that preserve relations and avoid imposed vector, grid, or sequence representations. This review surveys GNN applications, graph constructions, architectures, and learning objectives, finding strong performance across diverse HEP tasks while identifying real-data validation as an unresolved boundary.
Problem
Particle-physics measurements often involve unordered sets with rich relations that are inconvenient for vector-, grid-, or sequence-based networks.
Method
The paper reviews GNN applications to HEP, linking graph construction and architecture choices to task requirements and organizing models through a normalized formalism.
Results
GNN applications show superior performance over other model architectures across diverse HEP tasks, supported by relational inductive bias, fewer parameters, richer losses, and natural data representations.
Takeaways & Limitations
GNNs provide a natural framework for graph-level, node-level, and edge-level predictions across triggering, reconstruction, simulation, and physics analysis.
Takeaways & Limitations
The reviewed methods were developed on simulated events, and no performance on real detector data is reported.
Abstract
from arXiv · showhide
Particle physics is a branch of science aiming at discovering the fundamental laws of matter and forces. Graph neural networks are trainable functions which operate on graphs---sets of elements and their pairwise relations---and are a central method within the broader field of geometric deep learning. They are very expressive and have demonstrated superior performance to other classical deep learning approaches in a variety of domains. The data in particle physics are often represented by sets and graphs and as such, graph neural networks offer key advantages. Here we review various applications of graph neural networks in particle physics, including different graph constructions, model architectures and learning objectives, as well as key open problems in particle physics for which graph neural networks are promising.
1. Introduction
Particle physics studies fundamental laws using sparse, heterogeneous measurements that are naturally represented as sets and graphs. This review examines how graph neural networks exploit that structure across particle-physics applications.
- Particle physics uses experimental measurements to test models of fundamental particles and forces, with data sampled sparsely across time and space.
- Data Representation: Detector measurements are heterogeneous and spatially sparse, so they do not naturally fit homogeneous grid-like data structures.Collider detectors combine tracking, calorimeter, and muon sub-detectors using different measurement technologies.
- Deep learning applied to high-level particle-physics features can improve classical analyses but does not exploit the full potential of lower-level information.
- Image representations can lose information through irregular detector geometry and sparse projections, potentially limiting what models extract from the data.
- Sequence models impose an ordering on measurements or reconstructed objects, and evidence shows permutation-invariant networks can outperform sequence-based algorithms on the same classification task.
- This review explains why graphs suit particle-physics data and surveys GNN graph constructions, architectures, applications, and open questions.It also provides a normalized model description intended to simplify GNN adoption and further development in high-energy physics.
2. Geometric Deep Learning
Graph neural networks provide a graph-native framework for learning relational functions, with the GN formalism expressing graph-to-graph computation through update and permutation-invariant aggregation functions. GN blocks can be composed sequentially or recurrently, and their locality supports generalization to unseen graphs while enabling long-range computations through depth or recurrence.
- Graph neural networks: GNNs use parameterized message passing to propagate information across graphs and compute edge-, node-, and graph-level outputs.They implement relational inductive biases for graph-valued functions.
- Graph neural networks: The GN formalism generalizes several GNN architectures, including message-passing models and Transformer-style self-attention.Specific architectures arise by removing or rearranging GN components and choosing particular forms for their functions.
- GN block formalism: A GN block contains three update functions and three aggregation functions, with fixed-size updates and variable-set-to-fixed-size aggregations.The aggregation functions operate on sets such as edges incident on a node.
- GN block formalism: The GN edge, vertex, and global blocks transform input attributes while preserving graph structure, producing G′ = (u′, V′, E′).Edge outputs are aggregated by receiving node, node outputs are computed from incident edges, and all edge- and node-level outputs contribute to the global block.
- GN block formalism: Permutation-invariant reductions such as sums, means, or maximums allow GN aggregation to maintain permutation equivariance.The update functions are often trainable neural networks, while aggregation functions use invariant reduction operators.
- GN architectures: GN blocks can be arranged in depth-wise sequences or shared recurrent configurations, allowing local computations to support complex long-range operations.The formalism also tends to generalize to graphs not seen during training because edge- and node-level computations use local inputs.
3. Survey of Applications to Particle Physics
The review surveys graph neural network applications across particle-physics event analysis, organizing methods by graph-level, node-level, and edge-level predictions. It highlights graph construction and architecture choices for classification, pileup estimation, efficiency parametrization, and particle association.
- Scope of applications: GNN applications span graph-, node-, and edge-level predictions across triggering, reconstruction, simulation, and physics analysis.The reviewed methods are developed on simulated events, with no reported performance on real data.
- Graph classification: Jet-classification studies compare learned, dynamic, fully connected, interaction-network, Deep Sets, and attention-based graph constructions.These approaches address W-versus-QCD, quark/gluon, top-tagging, and multiclass jet categorization tasks while incorporating permutation-invariant or physics-motivated designs.
- Graph classification: Threefold signal-to-background improvement was reported for IceCube muon classification using a distance-constrained graph convolution model.The model builds a graph over detector sensors, updates node features through the adjacency matrix, and sums latent node features for graph-level classification.
- Node classification and regression: GNNs estimate per-particle pileup probabilities and improve jet and dijet-mass resolution while remaining stable across a large pileup-density range.One approach uses gated recurrent updates after stacked graph layers, while another maps a global graph representation back to particle nodes.
- Node classification and regression: Message-passing GNNs improve per-jet tagging-efficiency determination by modeling dependencies among all jets in an event.The graph-based parametrization is compared with a two-dimensional histogram over jet transverse momentum and pseudo-rapidity.
- Edge classification: A universal set-to-graph formalism outperforms standard methods on jet partitioning by about 10% across multiple metrics.The approach uses an equivariant embedding, fixed broadcasting, and graph-to-graph components for particle association within a jet.
4. Formulating HEP tasks with GNN
Formulating HEP tasks with GNNs requires choosing the task objective, graph construction, and architecture to match the physical relationships in the data. These choices balance relational inductive bias, information flow, and computational cost.
- 4.1. Task Definition: Task formulation determines whether GNN outputs support classification, regression, segmentation, clustering, or partitions of the input set.Classification uses node or graph representations with a standard classifier, while segmentation and clustering can use edge classification or object condensation.
- 4.1. Task Definition: Edge-based task formulations explicitly classify relationships, whereas object condensation uses node representations and computes attractive or repulsive relationships only during loss evaluation.This avoids computing those relationships as part of the model’s inference graph.
- 4.2. Graph Construction: Graph construction is nontrivial because particle-physics relationships are often unclear, so edges may be fully connected, predefined by feature-space neighbors, or learned dynamically.If pairwise conditional dependence is absent, a DeepSet with node and global properties may be preferable.
- 4.2. Graph Construction: Edges act as communication channels, carry physics-motivated input relations such as ∆R, and store latent relational information learned during message passing.These roles determine whether edges encode observed relationships, facilitate information exchange, or represent task-relevant variables discovered by the network.
- 4.2. Graph Construction: Fully connected graphs suit small sets, but their edge count grows as Ne ∝(Nv)2, making learned edge representations or attention costly for larger inputs.Fixed, easily precomputed edge features such as detector-module distances are one workaround.
- 4.2. Graph Construction: Dynamic k-nearest-neighbor graphs adapt connectivity in a learned feature space, but nondifferentiable neighborhood indexing prevents its parameters from being optimized by gradient descent.Only connected nodes affect parameter gradients, while neighborhood-selection parameters can instead be tuned through hyperparameter search; random edges can encourage exploration.
- 4.3. Model Architecture: GN architectures differ in whether node updates aggregate pairwise edge representations or globally aggregated node representations, producing different inductive biases.The former can encode pairwise ∆R-dependent interactions, whereas DeepSet-style updates independently transform nodes before forming an energy-weighted graph representation.
- 4.3. Model Architecture: Attention, EdgeConv, and GravNet connect learned representations with adaptive information aggregation, while node embeddings and attention weights can define latent neighborhoods or weighted sums.These designs allow inputs without explicit edges to exchange information through learned connectivity and relative weights.
5. Summary and Discussion
The review presents GNNs as a common framework for diverse HEP applications, with graph connectivity and message passing adapted to task and computational constraints. It also identifies graph generation and real-data validation as important open directions.
- 5. Summary and Discussion: Reviewed GNN applications show strong performance across diverse HEP tasks, supported by relational inductive biases, fewer parameters, tailored losses, and natural graph representations.Graphs are often sparsely connected to reduce computational requirements.
- 5. Summary and Discussion: Different GNN architectures share the core operation of exchanging information across graph structures, despite differing names and formalisms.The review organizes these models through common considerations about graph construction and model building.
- 5. Summary and Discussion: Transformers can be interpreted as message passing on complete graphs with normalized learned edge weights, while Reformer and Linformer reduce traditional Transformer costs through sparse or low-rank structure.Reformer uses learned-space nearest neighbors to construct a sparse graph; Linformer uses a low-rank approximation to the soft adjacency matrix.
- 5. Summary and Discussion: Graph generative models modify or infer graph structure through autoregressive prediction, latent-edge inference, reinforcement learning, or hierarchical pooling such as DiffPool.DiffPool performs attention-based soft edge prediction across graph hierarchies, pooling lower-level graphs into higher-level ones.
- 5. Summary and Discussion: Graph-based generative modeling remains relatively unexplored in particle physics, despite naturally graph-structured targets such as calorimeter energy deposits and particle-flow candidates.The review connects this direction to the substantial computational demands of particle-physics simulation.
- 5. Summary and Discussion: Dense implementations offer fast parallel operations but quadratic memory use, whereas sparse implementations scale with edge count but can incur costly indexing operations.The dense–sparse trade-off remains an active software and hardware development issue.
- 5. Summary and Discussion: The reviewed models perform well in proof-of-concept studies, but they have not yet been tested with real detector data.The reported studies use simulated events, so real-data performance remains unreported.