DeepLearnPhysics/oracle-dataset-v1

LArTPC readout dataset

Simulated readout of a liquid-argon time projection chamber: 1,199,200 events, 12,913,985 particles, 16.15 G voxel rows, 129.8 GB.

Each event is a list of energised voxels with an integer charge on each, plus a table saying which particle produced which voxels.

An event is not a single interaction. It overlays several independent ones (up to seven), each starting from its own vertex, and none of them is a neutrino interaction.

The charge is uncalibrated. It is raw ADC, and no detector constants are included — there is no conversion to electrons or to energy anywhere in these files. That is deliberate, not an omission.

Per-file counts and sizes are in each file's attributes and in MANIFEST.txt.

---

Layout

The data is a directory of HDF5 files:

observable/   obs_0000.h5 … obs_0011.h5
MANIFEST.txt  sha256, byte count and event count for each

The files are not a partition into subsets — there is no train/test split here. The division into twelve files is for size alone; how to group them, if at all, is yours to decide.

---

File contents

Three datasets, each of shape (n_events,) and variable-length: element i is a flat array holding event i.

dataset	vlen of	reshape to	meaning
coords	uint16	(N, 3)	voxel indices x, y, z
adc	uint16	(N,)	recorded charge in that voxel, ADC counts
particle	int32	(M, 3)	n_points, pid, semantic per particle

coords and adc are row-aligned: row r of coords is the voxel whose charge is adc[r]. N is the number of energised voxels in the event and M the number of particles.

Rows are per particle. Within one particle a given (x, y, z) appears at most once, but a voxel that several particles reached appears once for each of them, with that particle's own charge — so coordinates can repeat within an event (about 1% of rows share their coordinate with another row). Summing the rows that share a coordinate gives the total charge recorded at that voxel.

Which voxels belong to which particle

particle[:, 0] is a run length, not an index. Rows of coords/adc are contiguous by particle, in the order the particle table gives:

edges = np.concatenate([[0], np.cumsum(particle[:, 0])])   # length M+1
lo, hi = edges[k], edges[k + 1]                            # particle k owns
coords[lo:hi], adc[lo:hi]                                  #   these rows

sum(particle[:, 0]) == N holds in every event. Every particle has n_points >= 1.

Particle order within an event carries no information. Do not read anything into a particle's index.

Codes

pid and semantic are two independent labels, both given per particle. Counts below are over all 12,913,985 particles in the dataset.

pid		count		semantic		count
−1	none	217		0	shower	2,649,544
0	photon	1,766,203		1	track	5,727,389
1	electron	5,420,808		2	michel	594,027
2	muon	2,983,977		3	delta	3,943,025
3	pion	964,400
4	proton	1,778,380

pid can be −1. It is a sentinel meaning no particle type was assigned, not a sixth species. This is rare — 217 particles, 0.0017% — but those particles are otherwise ordinary: they carry a normal semantic (all 217 are 0 shower) and their voxels are unremarkable. Look codes up with dict.get() rather than dict[...] so the sentinel cannot raise. semantic takes only 0–3 and has no sentinel.

The two labels are not redundant and do not determine one another. Every combination that occurs, with its count:

pid	semantic	count
−1 none	0 shower	217
0 photon	0 shower	1,765,575
0 photon	1 track	628
1 electron	0 shower	883,751
1 electron	1 track	5
1 electron	2 michel	594,027
1 electron	3 delta	3,943,025
2 muon	0 shower	1
2 muon	1 track	2,983,976
3 pion	1 track	964,400
4 proton	1 track	1,778,380

Several combinations are vanishingly rare — one muon is labelled shower, five electrons are labelled track. Code that switches on either label should have a default branch rather than assume the common cases exhaust the data.

---

Geometry

A cube of liquid argon 230.4 cm on a side, divided into a 768 × 768 × 768 grid of 3 mm voxels. Indices run 0 … 767 on each axis.

The readout plane is the x = 0 face. Ionisation electrons drift along x toward it, so a voxel's index along x fixes how far its charge travelled before being recorded. Voxel (i, j, k) has its centre at

(i + 0.5, j + 0.5, k + 0.5) * 0.3 cm

and its drift distance to the readout plane is (i + 0.5) * 0.3 cm, between 0.15 cm and 230.25 cm.

drift_axis and anode_at_low_index in the file attributes state this convention; read them rather than assuming it.

---

Units

Charge is in ADC counts — a linear, integer-valued digitisation. No conversion to electrons or to energy is provided, and no detector conditions are recorded in these files.

adc_max is the largest value the digitiser can represent.

min_adc is the smallest value that actually appears. The underlying charge threshold does not fall on an ADC boundary, so the min_adc bin is only partially populated: do not read it as the threshold.

---

Edge cases

• 34 of the 1,199,200 events are empty. coords, adc and particle all have zero rows. Code that iterates over events has to tolerate this.
• pid can be −1. See the code tables above.
• adc is uint16, coords is uint16, particle is int32. An accumulation over voxels overflows uint16 immediately — cast before summing.
• Event counts differ per file; do not assume a uniform shard size.
• coords and adc are variable-length datasets. Indexing one event at a time re-reads the heap per row; read slabs for a full pass.

---

File attributes

Every file carries, at the root:

attribute
grid	768
pitch_cm	0.3
volume_cm	230.4
drift_axis	0 (x)
anode_at_low_index	true
adc_bits, adc_max	digitiser range
min_adc	smallest value present in this file
charge_units	ADC counts
particle_order	randomised
pid_codes, semantic_codes	the code tables above
run_id	file identifier
n_events, stat_particles, stat_voxels, stat_photons	totals for the file
created_utc
note	conventions in prose

Each dataset also carries its own note attribute describing its layout.

---

Integrity

MANIFEST.txt lists a sha256, a byte count and an event count for every file. Its paths are relative to the directory holding it, so check from there:

sha256sum -c <(awk 'NF==4 && !/^#/{print $1"  "$4}' MANIFEST.txt)

---

Source and citation

Derived from PILArNet-M (branch v3_extra, Apache-2.0) by simulating a liquid-argon TPC readout of it. The underlying particle simulation, the truth labels and the event content are theirs.

@misc{young2025particletrajectoryrepresentationlearning,
  title={Particle Trajectory Representation Learning with Masked Point Modeling},
  author={Sam Young and Yeon-jae Jwa and Kazuhiro Terao},
  year={2025}, eprint={2502.02558},
  archivePrefix={arXiv}, primaryClass={hep-ex},
  doi={10.48550/arXiv.2502.02558}
}

@misc{young2025pandaselfdistillationreusablesensorlevel,
  title={Panda: Self-distillation of Reusable Sensor-level Representations for
         High Energy Physics},
  author={Samuel Young and Kazuhiro Terao},
  year={2025}, eprint={2512.01324},
  archivePrefix={arXiv}, primaryClass={hep-ex}
}