HEPTv2: End-to-End Efficient Point Transformer for Charged Particle Reconstruction

2026-06-18 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Physical Sciences & Chemistry · Depth: Expert, quick

Summary

HEPTv2 is an end-to-end point-transformer architecture designed for charged-particle tracking, a fundamental high-energy-physics inference problem at the High-Luminosity Large Hadron Collider (HL-LHC). This system reconstructs trajectories from sparse detector measurements in one trainable pipeline, eliminating the need for auxiliary stages like graph construction, clustering, or filtering. HEPTv2 integrates a locality-aware point encoder, which uses locality-sensitive hashing for efficient local attention, with a track decoder that predicts complete trajectories via sectorized decoding and direct hit-to-track prediction. The full pipeline is optimized end-to-end under joint encoder-decoder supervision. On the TrackML dataset, HEPTv2 achieves 98.6% double-majority tracking efficiency and a 0.8% fake rate. It requires approximately 15 ms inference time and 0.4 GB peak memory per event on an NVIDIA A100 GPU, scaling linearly for events up to 5×10^5 hits. This architecture improves efficiency by 4.5% over previous transformers and 1.1-2.2% over optimized graph-based pipelines, while reducing latency by factors of 7 and 38-52.

Key takeaway

For Machine Learning Engineers developing particle reconstruction systems for high-luminosity colliders, HEPTv2 demonstrates that end-to-end point transformers can significantly outperform traditional graph-based or multi-stage transformer approaches. You should evaluate integrating similar end-to-end architectures, leveraging locality-aware encoding and direct hit-to-track prediction, to achieve superior accuracy and drastically reduced latency. This approach is crucial for meeting the real-time processing demands of future experiments like the HL-LHC.

Key insights

HEPTv2 is an end-to-end point transformer for charged particle reconstruction, achieving high accuracy and efficiency without graph processing.

Principles

• Locality-sensitive hashing preserves tracking-relevant geometry for efficient local attention.
• End-to-end optimization improves efficiency and accuracy in complex reconstruction tasks.
• Sectorized decoding and direct hit-to-track prediction resolve combinatorial ambiguities.

Method

HEPTv2 combines a locality-aware point encoder using locality-sensitive hashing with a track decoder that predicts trajectories via sectorized decoding and direct hit-to-track prediction, optimized end-to-end.

In practice

• Real-time particle reconstruction at HL-LHC.
• Efficient processing of high-density collision data.

Topics

• Charged Particle Reconstruction
• Point Transformers
• High Energy Physics
• Locality-Sensitive Hashing
• HL-LHC
• Inference Efficiency

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• GPT-Based Fast Simulation of CLAS12 Detector Hits via Conditional Autoregressive Generation — A GPT-style transformer rapidly simulates CLAS12 detector hits, maintaining physics fidelity with significant speedup over traditional methods.
• #AAAI2026 invited talk: machine learning for particle physics — Machine learning, especially graph neural networks, can overcome limiting assumptions in particle tracking to discover new physics.
• 🍙Any Resolution, Any Geometry🍙 👉Ultra Resolution Geometry Transformer (URGT) for... — URGT achieves state-of-the-art depth-normal estimation for arbitrary ultra-high resolutions.
• H2VLR: Heterogeneous Hypergraph Vision-Language Reasoning for Few-Shot Anomaly Detection — H2VLR uses a heterogeneous hypergraph for high-order visual-semantic reasoning in few-shot anomaly detection.
• LongStream: Long-Sequence Streaming Autoregressive Visual Geometry — LongStream enables stable, metric-scale 3D reconstruction for ultra-long sequences by decoupling gauge and optimizing Transformer caches.
• DeepSeek build a New Topological Transformer — DeepSeek's MHC transformer uses manifold-constrained hyperconnections for stable, learnable multi-lane information routing.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.