Charged Particle Tracking Residuals

---
name: charged-particle-tracking-residuals
description: Evaluates the accuracy of neural networks in regressing charged particle hit positions (x, y) and incident angles (alpha, beta) from pixelated silicon sensor charge patterns. It measures how well on-chip inference models reconstruct particle trajectories compared to ground-truth simulation and traditional reconstruction algorithms. Use when the user wants to benchmark on Simulated silicon tracker charge clusters, or asks about evaluating this task. Reports residual_68pct_interval.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2602.15946
  bibtex_key: das2026onchip
  confidence: high
---

# charged-particle-tracking-residuals

> On-chip probabilistic inference for charged-particle tracking at the sensor edge — Arghya Ranjan Das et al. (2026) (arXiv:2602.15946, 2026)

## What this evaluates

Evaluates the accuracy of neural networks in regressing charged particle hit positions (x, y) and incident angles (alpha, beta) from pixelated silicon sensor charge patterns. It measures how well on-chip inference models reconstruct particle trajectories compared to ground-truth simulation and traditional reconstruction algorithms.

## Datasets

- **Simulated silicon tracker charge clusters** — total ?; splits: test (-1)

## Metrics

- `residual_68pct_interval` **(primary)** — range: microns or degrees
  - The minimum interval containing 68% of the residual distribution $R_v = v - v_{true}$ for each output variable $v \in \{x, y, \alpha, \beta\}$. Lower values indicate better resolution.
- `mean_residual_bias` — range: microns or degrees
  - The mean of the residual distribution $\bar{R}_v = \frac{1}{N}\sum (v - v_{true})$. Measures systematic offset.

## Input / output format

**Input**: 2D grid of digitized charge values (0-3) across a 16x16 pixel array, collected over 2 time frames separated by 3.8 ns.

**Output**: Regression targets: hit position (x, y), and incident angles (alpha, beta) or their cotangents (cot_alpha, cot_beta). Full/Max models also output per-variable uncertainty sigma_v.

## Scoring recipe

```python
def compute_metrics(predictions, true_values):
    residuals = predictions - true_values
    mean_bias = np.mean(residuals)
    sorted_res = np.sort(residuals)
    idx = int(0.68 * len(sorted_res))
    interval_68 = sorted_res[idx] - sorted_res[-idx]
    return mean_bias, interval_68
```

## Common pitfalls

- The evaluation uses simulated data with specific sensor geometry (16x16 pixels, 50x12.5 um pitch) and magnetic field (3.8 T), which limits direct generalizability to other detector configurations.
- Non-ML baseline comparisons (e.g., LocalReco, Geometric) rely on optimistic assumptions like perfect knowledge of angles or Lorentz drift corrections, making the ML model's advantage appear larger than in strict real-world conditions.

## Evidence (verbatim from paper)

> A summary of the residuals $R_{v}\=v-v_{\text{true}}$ for $v\in{x,y,\alpha,\beta}$ is shown in Figure [3] for all models. The points represent the mean of the residual distribution ($\bar{R}_{v}$), and the solid lines represent the minimum interval containing 68% of clusters in the test set, $I_{v}(68\%)$.

## Citation

```bibtex
@misc{das2026onchip,
  title={On-chip probabilistic inference for charged-particle tracking at the sensor edge},
  author={Arghya Ranjan Das et al. (2026)},
  year={2026},
  note={arXiv:2602.15946}
}
```

- arXiv: 2602.15946