---
license: cc-by-4.0
pretty_name: AET-Bench
tags:
- atomic-electron-tomography
- haadf-stem
- 3d-reconstruction
- 4d-imaging
- materials-science
---

# AET-Bench: A Comprehensive Benchmark for Atomic Electron Tomography

A NeurIPS 2026 Datasets & Benchmarks submission: 755 samples, 6 difficulty
tiers, 14 primary reconstruction methods, and 4 evaluation tasks spanning
static 3D, atom-level localization, element classification, and 4D dynamic
tracking of nanoparticles at the atomic scale.

AET-Bench introduces **atom identity persistence** and **trajectory RMSD**
— two atom-centric metrics that expose a striking failure mode:
state-of-the-art 4D Gaussian-splatting reconstructions reach SSIM 0.94 on
AET data yet track only 15 % of individual atoms across time.  The paper
argues that pixel-level metrics systematically mislead atomic-scale
evaluation because the HAADF-STEM $Z^\gamma$ forward model decouples density
from atom count.

## Highlights

| | |
|---|---|
| **Samples** | 755 across 6 tiers (analytical, binary alloys, 4D dynamics, multislice, real, Brownian Pt) |
| **Methods** | 14 primary + supplementary baselines from 5 algorithmic families |
| **Tasks** | T1 static 3D · T2 atom localization · T3 element classification · T7 4D dynamics |
| **Results** | 14,000+ rows in `data/aet_bench/results.jsonl` |
| **Findings** | 7 findings including 4D paradox, dose sensitivity, sim-to-real gap |

## Quickstart

```bash
# Install
conda env create -f environment.yml
conda activate aet-bench

# Run one method on one sample
python scripts/run_benchmark.py \
    --subset mono --sample Cu-0165-00 \
    --method wbp --task T1

# Regenerate paper figures & tables
python scripts/generate_paper_assets.py

# Audit the paper result log and record its SHA256/provenance
python scripts/validate_results.py data/aet_bench/results.jsonl
python scripts/generate_paper_assets.py --results data/aet_bench/results.jsonl

# Fill GPU-heavy missing baselines; defaults to one worker per visible GPU
# Uses the existing aet4dgs conda env unless PYTHON_BIN is set.
bash scripts/start_gpu_extras.sh

# 4D-paradox mechanism analysis (reproduces Appendix D)
python scripts/analysis/fourd_mechanism.py
```

## Repository Layout

```
data/aet_bench/           # benchmark data + results.jsonl (git-ignored)
scripts/
  run_benchmark.py        # unified entry point: --subset --method --task
  generate_paper_assets.py # regenerates every paper figure + LaTeX table
  bench_bootstrap.py      # 95 % bootstrap CI utility
  analysis/               # paradox / dose / ranking analyses
  metrics/                # FSC, Hungarian atom matching, trajectory metrics
  baselines/              # thin wrappers around each method
paper/aet_bench.tex       # NeurIPS submission
paper/croissant.json      # Croissant 1.0 dataset card
```

## Adding a New Method

Methods are registered in `scripts/run_benchmark.py` under `METHOD_REGISTRY`.
Each entry maps a method name to a callable `f(data, sample_id) -> volume`.
See `CONTRIBUTING.md` for the detailed contract and examples.

## Data

Simulated tiers are reproducible from scratch via
`scripts/generate_haadf_data.py` (analytical $Z^\gamma$),
`scripts/generate_abtem_multislice.py` (dynamical scattering), and
`scripts/generate_dose_tier.py` (dose-controlled Poisson noise).

Real experimental tier sources 26 nanoparticles from published Zenodo
records (DOIs listed in `paper/croissant.json`). The NeurIPS review
release bundle is hosted at
`https://huggingface.co/datasets/pone7/4D-ATE`.

## Citation

Submitted to NeurIPS 2026 Datasets & Benchmarks.  BibTeX will be added
after the review window.

## License

CC-BY-4.0 for the benchmark data and paper; MIT for the reference
reconstruction and evaluation code.

## Hugging Face Large-File Parts

Some tier archives are split into parts to keep individual Hub files below the recommended size limit.
Concatenate parts in lexical order before extracting the reconstructed tar archive:

```bash
cat aet-bench-bi.tar.part-* > aet-bench-bi.tar
sha256sum aet-bench-bi.tar
```
Expected SHA256 for `aet-bench-bi.tar`: `6661b6b85c267e28010d479e854a95a84cf5bab8dde45cc0ed6a247b7a85b3e2`.

```bash
cat aet-bench-mono.tar.part-* > aet-bench-mono.tar
sha256sum aet-bench-mono.tar
```
Expected SHA256 for `aet-bench-mono.tar`: `926adcfb2fc73c1d62b8c44befc9963540899f55a07669022da99c0376674602`.

```bash
cat aet-bench-real.tar.part-* > aet-bench-real.tar
sha256sum aet-bench-real.tar
```
Expected SHA256 for `aet-bench-real.tar`: `7e433949b0c53d739b594bf724d4505e7e348b7ca1113e0838da630ae242e43d`.

Uploaded-file checksums are in `SHA256SUMS`; reconstructed archive checksums are in `ORIGINAL_ARCHIVE_SHA256SUMS`.