High-Energy Physics (HEP) / Machine Learning

GNN-Based Background Rejection

Graph Neural Networks are employed to improve the existing muon reconstruction algorithm by filtering out background hits. The sparse geometry of hits in the Muon Spectrometer makes GNNs a natural fit for classifying signal against background noise. To ensure computational viability, graphs are constructed dynamically from "Muon Buckets" (higher-order clusters of hits), rather than individual hits.

The deployed GNN leverages an EdgeConv architecture, propagating information through a local neighborhood of connected muon buckets. This approach ensures local spatial correlations are captured while maintaining a sparse graph structure for efficient message passing. This method significantly reduces the data load for subsequent pattern recognition stages.

Vision Transformers for End-to-End Tracking

The second approach explores an end-to-end muon tracking solution using state-of-the-art Vision Transformers (ViTs), specifically an adaptation of the Mask2Former architecture. This leverages advances in attention mechanisms and computer vision to solve the combinatorial problem of track finding and parameter estimation.

The architecture treats individual detector hits as separate tokens, incorporating a physics-informed prior by sorting hits in azimuthal angle and using windowed Flash Attention for computational efficiency (scaling as O(W × N)). A critical hit-filtering stage precedes the main tracking, performing binary classification to discriminate signal from noise at the individual hit level.

Quantified Performance & Impact

The GNN-based Bucket Filter achieved a 97% background bucket rejection rate at μ=60, leading to a 15% reduction in total reconstruction time (from 255 ms to 217 ms) for high-occupancy events (μ=200) on NVIDIA H100 GPUs, without compromising signal reconstruction efficiency.

The ViT-based tracking proof-of-concept delivered ultra-fast approximate muon reconstruction in 2.3 ms on consumer-grade GPUs, with a high 98% tracking efficiency. The integrated hit-filtering stage boasted an AUC of 0.9997, increasing hit purity from 0.6% to 66.5% and achieving 99.7% background rejection. This reduces event occupancy from 6,900 to just 55 hits per event, with 99.7% of muon tracks remaining reconstructable.

Key metrics include an average double matching efficiency of 94.59% and a charge sign classification accuracy of 96.35%. While track parameter regression precision is still developing, the pattern recognition capabilities are highly promising.

Strategic Implications & Future Work

This work demonstrates the immense potential of ML-based methods for the ATLAS Muon Event Filter pipeline. The GNN approach offers immediate speed improvements for existing systems, while the ViT proof-of-concept showcases a path towards a transformative, end-to-end tracking solution.

Future research will focus on integrating the global ViT filtering stage into the baseline reconstruction chain, optimizing for long-lived particle decays, and improving parameter regression precision. Runtime optimizations like pruning, quantization, and model compilation are crucial for further deployment. The reliance on attention mechanisms, backed by major industrial support, ensures these technologies will remain sustainable and continue to accelerate, offering a robust solution for high-throughput HEP applications.