--- license: cc-by-4.0 pretty_name: AET-Bench size_categories: - n<1K tags: - atomic-electron-tomography - electron-microscopy - haadf-stem - tomography - 3d-reconstruction - 4d-imaging - materials-science - benchmark --- # AET-Bench: Pixel-Accurate Atomic Tomography Is Not Atom-Accurate **NeurIPS 2026, Evaluations and Datasets Track** Zicheng Liu1, Wenzhuo Ma2, Jintao Chen3, Di Huang1,2,4,5,\* 1BCC Lab, Hangzhou International Innovation Institute, Beihang University, Hangzhou, China
2Shenzhen Loop Area Institute, Shenzhen, China
3Zhejiang University, Hangzhou, China
4Zhejiang Industrial Big Data and Robot Intelligent System Key Laboratory, Hangzhou Innovation Institute, Beihang University, Hangzhou, China
5State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing, China
\*Corresponding author: Di Huang (dhuang@buaa.edu.cn) [Paper (OpenReview)](https://openreview.net/forum?id=GmdQ25dIWq) · [Code (GitHub)](https://github.com/pone7/4D-ATE) · [Leaderboards](SUMMARY.md) ## Overview Atomic electron tomography (AET) aims to recover the 3D positions, chemical species, and temporal trajectories of individual atoms, yet reconstructions are usually scored with voxel metrics such as PSNR, SSIM, and FSC. AET-Bench is a dataset and benchmark for **atom-first** evaluation: it scores reconstructions by the atoms, species, and trajectories scientists actually use. On the Mono tier, the method with the best voxel similarity (MLEM, SSIM 0.946) is not the one that recovers the most atoms (FISTA-TV, 96.9% atom recall at SSIM 0.506), and 4D Gaussian reconstructions with SSIM 0.94 keep only about 15% of atom identities over time. ## Tiers | Tier | Samples | Data | Reference | Evaluation focus | |---|---|---|---|---| | Mono | 410 (50 canonical) | Analytical HAADF-style projections, single-element FCC nanoparticles | Atom coordinates | Localization | | Bi | 240 (30 canonical) | Analytical, binary-alloy nanoparticles | Coordinates + species | Chemistry | | 4D | 10 | Time-varying Cu trajectories, 10 phases each | Coordinates + atom IDs | Trajectories, identity persistence | | Multislice | 3 | abTEM multislice simulation (Cu, Pt) | Coordinates + species | Forward-model mismatch (pilot) | | Real | 26 | Experimental tilt series from published records | Reference atom models | Reference agreement | | Brownian-Pt | 66 | Dynamic Pt nanocrystal frames | Dynamic reference models | 4D tracking | The paper's leaderboards use the canonical base samples of Mono and Bi; the additional samples (ids with an `-sNN` suffix) are size and orientation variants for development and ablation. Each sample provides the tilt series, projection angles, physical calibration, metadata, and the available references (volume, atom coordinates, species labels, temporal identities). Within a tilt series, every seventh projection is reserved for validation (60 fitting / 11 validation projections for a 71-angle series); there is no sample-level train/test split, because every sample is an evaluation sample. ## Files | File | Content | |---|---| | `aet-bench-.tar` / `.tar.part-NNN` | Tier archives (mono, bi, sim4d, multislice, real, brownian_pt) | | `aet-bench-samples.csv` | One row per sample: tier, element, atom count, number of angles, paths inside the archive | | `results.jsonl` | Full benchmark result log (one JSON row per method run) | | `aet-bench-results.csv` | Flat CSV view of `results.jsonl` | | `SUMMARY.md` | Leaderboards for every tier, computed with the paper's aggregation rules | | `aet-bench-release-manifest.json` | File sizes and SHA256 for every uploaded file | | `croissant.json` | Croissant 1.0 metadata with Responsible AI fields | | `SHA256SUMS`, `ORIGINAL_ARCHIVE_SHA256SUMS` | Checksums of uploaded files and of reassembled archives | ### Reassembling split archives Some tier archives are split into parts to keep individual Hub files below the recommended size. Concatenate the parts in lexical order, check the SHA256, then extract: ```bash cat aet-bench-bi.tar.part-* > aet-bench-bi.tar sha256sum aet-bench-bi.tar # 6661b6b85c267e28010d479e854a95a84cf5bab8dde45cc0ed6a247b7a85b3e2 ``` ```bash cat aet-bench-mono.tar.part-* > aet-bench-mono.tar sha256sum aet-bench-mono.tar # 926adcfb2fc73c1d62b8c44befc9963540899f55a07669022da99c0376674602 ``` ```bash cat aet-bench-real.tar.part-* > aet-bench-real.tar sha256sum aet-bench-real.tar # 7e433949b0c53d739b594bf724d4505e7e348b7ca1113e0838da630ae242e43d ``` ## Using the benchmark The evaluation harness, method wrappers, and metrics are at : ```bash conda env create -f environment.yml && conda activate aet-bench python scripts/run_benchmark.py --subset mono --sample Cu-0165-00 --method wbp --task T1 ``` To reproduce the paper's tables and `SUMMARY.md` from `results.jsonl`: ```bash python scripts/paper/aggregate_results.py --results results.jsonl --out paper_results.json python scripts/paper/make_summary_md.py --agg paper_results.json --out SUMMARY.md ``` ## Provenance Simulated tiers (Mono, Bi, 4D, Multislice) are generated with ASE nanoparticle construction, an analytical Zγ forward model, and abTEM multislice simulation; the generating scripts are in the code repository. The Real tier is converted from published experimental records: | Source | DOI | |---|---| | Zr–ZrO₂ nanoparticles | [10.5281/zenodo.12742309](https://doi.org/10.5281/zenodo.12742309) | | MPEA2 high-entropy alloy | [10.5281/zenodo.17310133](https://doi.org/10.5281/zenodo.17310133) | | Pt Brownian tomography | [10.5281/zenodo.15531575](https://doi.org/10.5281/zenodo.15531575) | | RESIRE datasets | [10.5281/zenodo.7819857](https://doi.org/10.5281/zenodo.7819857) | | AlSiYb alloy | [10.5281/zenodo.2578866](https://doi.org/10.5281/zenodo.2578866) | | PdPt core–shell (Zhou group) | [10.5281/zenodo.7733503](https://doi.org/10.5281/zenodo.7733503) | | PdPt core–shell (Yang group) | [10.5281/zenodo.8245312](https://doi.org/10.5281/zenodo.8245312) | | NiPt multi-grained catalyst | [10.5281/zenodo.15321070](https://doi.org/10.5281/zenodo.15321070) | | AuPd icosahedral nanoparticles | [10.5281/zenodo.14677350](https://doi.org/10.5281/zenodo.14677350) | The Brownian-Pt tier uses synthetic projections of published Pt reconstructions. Please also cite the original records when you use the Real or Brownian-Pt tiers. ## Limitations * Real-tier scores measure agreement with experimentally derived reference models, not ground-truth accuracy; near-zero recall there is a reference-agreement failure, not proof that the atoms are unrecoverable. * The Multislice tier has three samples and is a pilot: it supports a direction-of-effect claim, not a ranking. * The analytical forward model does not capture dynamical scattering, and the data are biased toward inorganic metal and oxide nanoparticles imaged with HAADF-STEM. * All public samples carry their references; a held-out private evaluation split for a public leaderboard is planned. ## Citation ```bibtex @inproceedings{liu2026aetbench, title = {{AET-Bench}: Pixel-Accurate Atomic Tomography Is Not Atom-Accurate}, author = {Liu, Zicheng and Ma, Wenzhuo and Chen, Jintao and Huang, Di}, booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track}, year = {2026} } ``` ## License The dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The code is released under the MIT License. ## Acknowledgments This research was supported by the Ningbo Science and Technology Project (Grant No. 2025Z058), the Postdoctoral Innovative Talent Support Program of the China Postdoctoral Science Foundation (Grant No. BX2026462), the China Postdoctoral Science Foundation General Program (Grant No. 2026M794682), the Priority-Funded Postdoctoral Research Project of Zhejiang Province (Grant No. ZJ2026116), the Research Program of State Key Laboratory of Complex and Critical Software Environment, and the Fundamental Research Funds for the Central Universities.