acceptodds
Withdrawn from ICLR 2027

Fast and Precise Learned Charged-Particle Trajectory Regression at the Large Hadron Collider

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We propose a training recipe that treats charged-particle trajectory parameter regression on high-energy physics detector data as a sequence-modeling task. Kalman filters and linearized least-squares fits have been the classical standard approach for this task: they are optimal estimators for sparsely sampled linear-Gaussian data and are commonly used for trajectory parameter regression (fitting). The classical fitting techniques implemented for this domain reach a final precision of one part in through detailed modeling of detector geometry, material, detection effects and precise numerical integration of the equations of motion through the detector's inhomogeneous magnetic field. With this study, we demonstrate that using a bidirectional gated linear recurrent encoder, one is able to reproduce the full precision of classical track fitting techniques. Using a custom kernel, we also achieve significantly higher throughput during GPU inference, compared to classical fitting software running on similarly priced multi-core CPU servers representing typically employed hardware. Such a speedup would lead to considerable cost savings for the pattern recognition at the Large Hadron Collider. To our knowledge, this is the first end-to-end learned track fit to reach the full precision and, at the same time, offer the opportunity to reduce the computing costs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
sequence modeling, state space models, linear recurrent neural networks, minGRU, AI for science, high-precision regression, particle physics, charged-particle tracking, track fitting, Kalman filter, GPU kernels, efficient inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_sU1kOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a method to replace the classic Kalman filter for charged-particle trajectory regression with minGRU. The experimental results show comparable regression performance to the Kalman filter-based method and much better throughput.

Critical strengths

None

Critical weaknesses

While I expect that the application of neural networks to particle trajectory regression is of considerable importance, it is unclear what insights this work offers from the perspective of representation learning.

Beyond replacing the Kalman filter with a minGRU, I find it difficult to identify the technical novelty of this work as a machine learning method or its broader contribution to the machine learning community from the current version of the paper.

Questions for authors
  • It would be helpful to clarify which aspects of this work constitute a contribution to the ICLR community.
  • What are the key technical contributions from a machine learning perspective, and what kinds of discussions within the ICLR community would be beneficial for advancing this application?
Reviewer_ycHDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes to use bidirectional gated linear recurrent encoder (referred as BGLR encoder below) to serve as a neural network surrogation of Kalman fitting for the track fitting task in silicon trackers. The BGLR encoder is then applied to a narrow subset of the COLLIDERML dataset with an assumption of perfect hit-to-track matching. By providing an analytical physical solution as the seed, and formulating this task as a residual correction, the authors demonstrate that with BGLR encoder and customized kernel optimizations on GPUs, the concurrency of track processing could be increased by 10-30 times compared to Kalman fitting on 32-core CPUs. In the meantime, BGLR encoder is able to maintain similar precision as the traditional Kalman fitting method, in terms of point estimation.

Critical strengths

The main strength of this paper is that it demonstrates the proposed BGLR encoder, together with an optimized GPU kernel, could achieve precision comparable to a classical Kalman fitter, under a specific experimental condition (single muon, no pile-ups, perfect hit-track matching)

Critical weaknesses

The idea of accelerating a traditional Bayesian state-space estimator (Kalman fitting in this case) using a neural-network surrogate is interesting and meaningful. However, I do not believe the demonstrated results are sufficient to justify decent quality as a full conference paper. With the narrow scope of the data, insufficient hardware-normalized baseline comparison, and incomplete ablation studies, the paper fails to adequately justify its main contributions and broader claims. i.e. “the first learned trajectory fit to match classical Kalman-filter precision, offering the potential for significant cost savings through higher throughput at comparable hardware cost.

1.Scope of data: In the introduction section, the author(s) mentions clearly that “A pattern recognition algorithm has to recover which hits belong to which particle and what its trajectory was. The challenge grows in compute complexity when many particles arrive at once.” However,

  • The dataset selected in this paper is a very narrow subset of the COLLIDER ML dataset, which only contains single muon tracks without any pile-ups. The muon case is the physically easiest one among various charged particles like pions and electrons, especially in the 1-110 GeV range selected in this paper.
  • The authors further assume perfect hit-to-track matching, which eliminates a large part of the challenge in realistic charged-particle track reconstruction, especially the combinatorial ambiguity of hit assignment.
  • The author(s) formulates their model as a residual correction on top of helix, and it’s totally expectable that it performs well in this severely limited and constraint muon case.
  1. Unfair baseline:
  • In the related work, the author(s) already mentions GPU-based Kalman fittings. However, in table 2, they only compare their model, deployed on modern GPUs, with the Kalman fitting on CPU. It is unclear if the 10-30 times of speed up comes from the advantage of the model or systematic hardware bias.
  • One of the biggest advantages of the Kalman fitting, or, in general, the traditional Bayesian models, is that they provide the posterior of the prediction but not just point estimations. This becomes especially important for real-world applications when the hit assignment and track fitting needs to be performed jointly. Although the proposed model predicts multiple quantiles for each track parameter, table 1 only demonstrates the comparison on point-estimation precision. The authors do not evaluate the calibration of the predicted uncertainties or demonstrate whether these marginal quantiles can replace the joint covariance estimates provided by Kalman fitting.
  • In table 2, the “approximate cost” is simply the price of the device. However, this is a super naive estimation. Considering the data size from the colliders, the purchase of CPU/GPU is only a small portion in the computation cost. Also, the proposed method in this paper requires re-training for each detector geometry, under different physics conditions. The simulation cost and training cost cannot just be neglected. On the other hand, Kalman fitting is a universal, robust algorithm which could be adapted into various setups and physics conditions.
  1. Missing ablation studies: The authors did not compare their proposed approach against independent neural-network baselines. The comparison in appendix Figure A.2 only replaces the encoder backbone within the same overall framework. In this highly simplified problem setup, I highly doubt whether a simple MLP or CNN could also achieve comparable performance using the same analytical helix seed and residual correction formulation.
Questions for authors

The following questions could materially affect my assessment:

  1. Generalization to other charged particles:
  • Why are the experiments restricted to single muons?
  • Can the authors provide results for other charged particles, particularly pions and electrons, to demonstrate that the method remains effective when the physical interactions are more challenging than the current simplified setup?
  1. Realistic tracking conditions:
  • How does the proposed method perform when the perfect hit-to-track matching assumption is relaxed?
  • For example, can the authors evaluate its robustness to missing hits, incorrectly assigned hits, or outliers?
  • If such an experiment is outside the scope of this work, can the authors provide a quantitative justification of the scientific significance and challenge of the isolated and simplified track-fitting setup without imperfect hit assignment and pile-ups?
  1. GPU Kalman baseline: Can the authors provide a direct comparison with a GPU-based Kalman fitter on comparable hardware?
  2. Uncertainty estimation: Since the proposed model predicts seven quantiles per trajectory parameter, can the authors report calibration and empirical coverage of these prediction intervals?
  3. Total computing cost and transferability:
  • Can the authors provide a more detailed cost estimate accounting for simulation, training, inference, and detector-specific adaptation?
  • For example: how many reconstructed tracks are needed for the training cost to be subdominant?
  • How does the proposed model transfer to different detector geometries, magnetic fields, and physics conditions without retraining? If they do need retraining, what would be the estimated cost? And with that cost, will the claim of “cost saving” still hold?
  1. Simpler baselines and component ablations:
  • Can the authors compare against simpler neural-network architectures such as an MLP or CNN, using exactly the same analytical seed and input features?
  • Can they also quantify the individual contributions of the 1 seed residuals, 2 Fourier encoding, 3 quantile regression, and 4 second-stage training, by doing comprehensive ablation studies?
Reviewer_ZncmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper treats charged-particle trajectory fitting, regression of track parameters given an already-assigned set of detector hits as a short-sequence modeling problem, and asks whether a learned model can match the precision of a classical Kalman filter while running much faster on GPUs. The motivations for using the RNN like architecture (minGRU-biredictional) is not very clear as the mainstream is still the Transformer or GNN in the regime. The throughput and performance metrics are clearly listed in the paper draft. The method is seed-anchored. Each hit carries 12 measured features plus three seed-derived residuals.

Training uses single muons from the public ColliderML dataset on the Open Data Detector in strict fp32 precision, deliberately at small batch 2048. Validation is against the truth-seeded KF shipped with the dataset, on four held-out muon samples (fixed pT of 2/10/50 GeV at 10⁵ tracks each, and a uniform 1–70 GeV sample of 3M tracks), using the field-standard iteratively 3σ-clipped RMS of residuals plus the un-clipped RMS.

Critical strengths

The paper focuses on the short-sequence kernel work with the transferable contribution for different tasks. Standard sequence-model kernels assume 10³–10⁵ tokens, whereas a track is at most 20-40 hits. Despite the inference speed advantages mostly occurs at long sequence for linear complexity model like RNN, the smaller offset of the minGRU architecture still yields better throughputs than the Transformer model. A careful model profiling, and compiling the Fourier encoding yields 6.1×/6.6×/3.6× for minGRU/Transformer/Mamba-2. The paper reports an architecture-speed comparison with detailed hyper-parameters (depth, width, and learning rate were not tuned per encoder so the ordering reflects their implementations).

Within its stated scope, the precision evaluation is more careful than this subfield’s norm, and the result is genuinely notable. The metric is standard in the field actually uses (iteratively 3σ-clipped RMS) and the un-clipped RMS that exposes tails, as per-parameter ratios to the reference on different samples, also draws on different physics-quantity like pT or η-range. It also reports qualitatively the performance on the 5 fitted tracking parameter against the Kalman Filter.

Critical weaknesses

The evaluated problem in the paper is substantially easier than the problem the title, abstract, and cost argument claim, and the dataset the authors chose contains the harder setting they skipped. Every result is on single muons with perfect hit-to-track assignment, no pileup, no hadrons or electrons, no missing or spurious hits, no misalignment, and no out-of-time activity. This is almost impossible to hold a realistic environment with the standard HL-LHC pileup conditions.

Track level reconstruction would always have its limitation, many of the current approaches works directly on hit level with end-to-end manner, the complexity of the task is lower than the current standard. It raises the question about how different architecture, even with simple DNN or RNN, could reach similar performance. One very important question for reader to understand the paper is what is difference between original minGRN and bidirectional version with clear evidence for using the latter one with more computational overhead.

The headline precision claim is contradicted by the paper’s own table for some parameters (i.e. q/p), the deployed path’s precision is never reported, and the one piece of evidence offered for uncertainty quality does not measure what it is said to measure.

Questions for authors

Clarifications that could materially change my assessment:

  1. Which numerical precision produced Table 1 and Figures 3–4, the fp32 training path or the bf/fp16 deployment path?
  2. Please reconcile Section 4.1 with Table 1 for q/p, and state the claim in a form the table supports. Is there a known cause for the curvature deficit, the scale-free q/p residual, the per-hit-only minGRU gate, how about the performance with other models, some plots in Appendix could be moved to the main body as the result tables/plots also need to be presented with other models (Mamba-2, Transformer).
  3. Also, more comparisons are needed with more advanced model (bidirectional Mamba or GDN, or other RNN) to support the choices using minGRU specifically.
  4. Can you report a seed-only row? The analytic seed is inside the forward pass and the model predicts a bounded correction to it, so the seed’s own resolution per parameter, and ideally the model’s resolution with the seed-derived input features removed, would show how much of the final precision the encoder contributes.
  5. Can you provide coverage or pull distributions from the predicted quantiles.
  6. Do you have any result on a second detector geometry, or on degraded inputs (dropped hits, injected noise hits, misalignment)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.