Description
TrueTrack is an independent research effort focused on advancing machine learning approaches for particle track reconstruction, with a particular emphasis on the TrackML dataset and related high-energy physics challenges.
The project brings together model design, detector-aware data representation, scalable processing, track reconstruction, and evaluation in a unified experimental framework for studying hit-to-track association under realistic event complexity.
Model design
At its core, TrueTrack explores geometry-aware Transformer-based methods for learning contextual relationships between detector hits. Rather than framing track reconstruction as a fixed-class classification problem, the project treats it as a relational learning task in which the model learns which observations are likely to originate from the same underlying particle trajectory.
Noise can be modelled explicitly within the same learning framework, allowing structured particle trajectories and stochastic background hits to be treated as related but distinct reconstruction problems. The aim is to learn representations that capture both detector context and the relational structure required for track formation without relying on globally fixed particle labels.
Working with TrackML data
A central objective of TrueTrack is reconstruction under the complexity of full TrackML-style events rather than relying on a problem definition in which difficult hits or noise are removed in advance. The pipeline is designed to retain event complexity, including noise hits, so that model behaviour can be studied under conditions that more closely reflect the intended reconstruction problem.
At the same time, the framework supports controlled reduced and simulated datasets for development, comparison, and diagnostic studies. This makes it possible to study how event complexity, noise, and data composition influence reconstruction performance without changing the broader training and evaluation framework.
Data representation and sharding
TrueTrack uses detector-aware hit representations that combine measured information with derived geometric context. This gives the learning system multiple complementary views of detector observations while preserving a consistent representation across events.
To make large-scale events tractable, events are decomposed into spatially structured subsets, or shards, which serve as the fundamental units for training and inference. Sharding allows the model to process manageable regions of an event while retaining the local geometric and topological relationships needed for track association.
The decomposition is designed to preserve continuity near regional boundaries, allowing local predictions to contribute to a coherent event-level reconstruction. The same representation also supports flexible experimentation with locality, overlap, and feature construction.
Scalable training pipeline
TrueTrack places strong emphasis on computational practicality. Event decomposition, variable-size batching, and memory-aware data access are used to process large detector events without requiring complete events or complete datasets to reside in accelerator memory simultaneously.
The implementation is designed with moderate research hardware and memory constraints in mind. Both indexed and sequential data-access strategies can be used within the same broader training workflow, allowing experiments to balance convenience, throughput, and memory use without changing the model or evaluation objectives.
The overarching goal is to shift the primary bottleneck away from hardware constraints and toward model quality, data representation, and learning efficiency.
Reconstruction and evaluation
Training a model is only one part of the TrueTrack workflow. Learned hit associations are subsequently converted into reconstructed particle-track candidates and evaluated in the context of complete detector events.
Evaluation is performed at several levels. In addition to measuring the quality of individual learned associations, TrueTrack examines the completeness and purity of reconstructed tracks, the consistency of the resulting event partition, and performance under TrackML-compatible scoring.
This distinction between learning metrics and reconstruction metrics is important because strong local association performance does not necessarily imply strong complete-track reconstruction. The evaluation framework is therefore designed to expose different reconstruction failure modes rather than reduce performance to a single optimisation metric.
Current status
The current TrueTrack implementation provides an end-to-end experimental framework covering detector-data preparation, model training, track reconstruction, and event-level evaluation. Development is presently focused on establishing and comparing reconstruction performance across model, data-representation, and training configurations.
Quantitative results and detailed methodological descriptions will be released when the evaluation has reached sufficient maturity for scientific reporting.
Outlook
TrueTrack sits at the intersection of machine learning systems design and detector-aware modelling. By combining expressive architectures, realistic data assumptions, scalable infrastructure, and reconstruction-focused evaluation, the project aims to provide a foundation for next-generation approaches to particle track reconstruction.