--- license: cc-by-4.0 pretty_name: "MaterialsSaddles" size_categories: - 10M A high-throughput library of converged transition states for solid-state and > surface chemistry. > Hub URL: > Released by SciLM: **34,135,597 fully converged transition states** computed by massively-parallel saddle searches on top of public materials and catalysis datasets, using the [SaddleMill](https://github.com/SciLM-ai/SaddleMill) package and Meta's [`uma-s-1p2`](https://github.com/facebookresearch/fairchem) machine-learning interatomic potential. Each entry in a file is a single structure. **Three consecutive entries form one transition-state event**: reactant minimum, transition state (first-order saddle), product minimum. Endpoints are converged to 0.02 eV/Å (max\|F\|), saddles to 0.05 eV/Å. Each saddle row also stores its **eigenmode** — an `(N, 3)` per-atom displacement field giving the direction along which the saddle is unstable. --- ## Quick stats | | | |---|---| | Transition states (total) | **34,135,597** | | Files (`.aselmdb`) | 416 | | Rows per TS | 3 (reactant, saddle, product), in order | | Saddle search method | Dimer (lemat / oc20 / oc22), NEB-CI (mp20bat) | | Calculator | `fairchem` `uma-s-1p2` | | Endpoint convergence (max\|F\|) | 0.02 eV/Å | | Saddle convergence (max\|F\|) | 0.05 eV/Å | | License | CC-BY-4.0 | ### Breakdown by source dataset | Subset | Source | Method | # transition states | Files | |-----------|---------------------------------------------|--------|---------------------|-------| | `lemat/` | LeMat-Bulk | Dimer | 31,346,419 | 256 | | `oc20/` | Open Catalyst 2020 (OC20) | Dimer | 2,587,101 | 96 | | `oc22/` | Open Catalyst 2022 (OC22) | Dimer | 167,335 | 32 | | `mp20bat/`| Materials Project battery structures | NEB-CI | 34,742 | 32 | --- ## Try it: minimal example notebook A self-contained Jupyter notebook ([`example.ipynb`](./example.ipynb)) demonstrates loading the dataset, converting ASE-LMDB rows to ASE `Atoms` objects (including a non-obvious `atoms.info` round-trip), walking the (R, S, P) triplet layout, visualizing a reaction, loading the train/val/test split manifests, and reproducing two small panels of Fig. 1 of the accompanying paper. It auto-installs its dependencies and downloads roughly 1 GB of sample shards (one each from `lemat`, `oc20`, `oc22`, and `mp20bat`); end-to-end runtime is roughly five minutes on a typical laptop connection. To run locally: ` ` `bash hf download SciLM/MaterialsSaddles example.ipynb \ --repo-type dataset --local-dir . jupyter notebook example.ipynb ` ` ` You can also open the notebook directly in the Hugging Face web UI; it renders inline. --- ## Directory structure ``` . ├── README.md (this file) ├── DATASHEET.md (Datasheet for Datasets, Gebru et al. 2018) ├── lemat/ │ ├── lemat_dimer_000.aselmdb │ ├── ... │ └── lemat_dimer_255.aselmdb (256 files) ├── mp20bat/ │ ├── mp20bat_neb_000.aselmdb │ ├── ... │ └── mp20bat_neb_031.aselmdb (32 files) ├── oc20/ │ ├── oc20_dimer_000.aselmdb │ ├── ... │ └── oc20_dimer_095.aselmdb (96 files) └── oc22/ ├── oc22_dimer_000.aselmdb ├── ... └── oc22_dimer_031.aselmdb (32 files) ``` The split into multiple `.aselmdb` shards is purely for file-size convenience (each shard is at most a few GB). Shards within a subset are interchangeable and can be processed in any order. --- ## What is in each row? Each `.aselmdb` is an ASE-LMDB database whose rows are stored in **triplets**: ``` row 0 reactant (Dimer: side = -1 ; NEB: image_type = 'endpoint') row 1 saddle (Dimer: side = 0 ; NEB: image_type = 'climbing') ← TS row 2 product (Dimer: side = 1 ; NEB: image_type = 'endpoint') row 3 reactant ... ``` The rich per-row metadata lives in `row.data['info']`. After `row.toatoms()`, **`atoms.info` is empty** — you have to copy `row.data['info']` over yourself (see [Loading](#loading-the-data)). Each row also exposes a small set of **searchable scalar key/value pairs** via `row.key_value_pairs` — at minimum `task_name` and `ms_id`, plus `src_index`, `side`, and `status` where applicable. These are convenient for `db.select(...)`-style filtering, but note that ASE's aselmdb backend performs **linear scans**: queries are O(N) per shard, not indexed. The keys in `row.data['info']` vary by source dataset and saddle-search method. **Only `task_name` and `ms_id` are guaranteed on every row** — for anything else, the table below documents *which subsets typically have it*. When in doubt, inspect `row.data['info'].keys()` for a few rows of your target subset. | key (in `info` dict) | who has it | what it is | |---|---|---| | `side` | dimer rows | -1 / 0 / 1 (reactant / saddle / product) | | `image_type` | NEB rows | `'endpoint'` (reactant/product) or `'climbing'` (saddle) | | `image_idx`, `subband_idx`, `nimages` | NEB rows | NEB band geometry | | `image_converged`, `band_converged`, `band_converged_CI` | NEB rows | per-image / band-level convergence flags | | `effective_fmax` | NEB rows | per-image max-force (NEB-modified) at convergence | | `converged` | dimer endpoint rows | endpoint converged flag | | `eigenmode` | saddle rows | `(N, 3)` float64 — the eigenmode at the saddle, one 3D displacement vector per atom | | `curvature` | dimer saddles | eigenvalue along that eigenmode (eV/Ų) | | `barrier`, `dE` | NEB saddles | reactant→TS and reactant→product energy differences (eV) | | `is_reaction`, `n_formed_bonds`, `n_broken_bonds`, `formed_bonds`, `broken_bonds` | dimer rows | bond changes detected via ASE neighbor lists | | `is_ads_reaction`, `n_ads_*`, `ads_*_bonds` | dimer rows | adsorbate-restricted bond changes | | `parent_ts_index` | dimer endpoints | upstream SaddleMill identifier of the parent saddle (an internal pointer; the saddle in the same triplet is the immediately adjacent row in this `.aselmdb` file) | | `src_index` | all rows | SaddleMill internal ID (file-local) | | `ms_id` | all rows | global, contiguous row index across the **entire dataset** (0-indexed, 0..102,406,790). Unique across all 416 shards and all four subsets. Triplets occupy three consecutive `ms_id`s: `3k` (reactant), `3k+1` (saddle), `3k+2` (product). Used by the `splits/` manifests below. | | `task_name` | all rows | UMA task head used during the saddle search. Varies by subset: `"omat"` for `lemat` and `mp20bat`, `"oc20"` for `oc20`, `"oc22"` for `oc22`. | | `status` | all rows | SaddleMill run status (always a `converged*` value here) | | `orig_info` | all rows | nested dict; carries the **source-dataset identifiers** (see below) | `row.data` may carry additional internal bookkeeping fields (e.g. `traj_path`) that are artifacts of the upstream production pipeline and have no scientific or downstream value. Read from `row.data['info']`; ignore anything else. ### Where to find the source-dataset identifiers The location depends on whether the input went through one or two stages of the SaddleMill pipeline before this release: | Subset | Path to source IDs in the row's `info` dict | Example fields | |-----------|---------------------------------------------|----------------| | `lemat` | `info['orig_info']['orig_info']` | `immutable_id` (e.g. `'agm005964602'`), `chemical_formula_*`, `functional`, `entalpic_fingerprint` | | `oc20` | `info['orig_info']['orig_info']` | `source_file` (e.g. `'random1176828.extxyz.xz'`) | | `oc22` | `info['orig_info']['orig_info']` | `sid`, `id`, `nads`, `natoms` | | `mp20bat` | `info['orig_info']` | `discharge_id` / `charge_id` (Materials Project IDs, e.g. `'mp-1006112'`), `working_ion`, `removed_ion_idxs` | For Dimer subsets `info['orig_info']` itself is a SaddleMill-internal dict (`attempt_id`, `reaction_type`, etc.) and the upstream identifiers live one level deeper. For the NEB subset there's only one level of nesting. --- ## Intended uses This dataset was built with three downstream uses in mind: 1. **Training generative models for transition-state prediction.** Each triplet gives reactant + product (conditioning) and saddle (target). The eigenmode and bond-change annotations make it easy to filter for chemically meaningful events. 2. **Generating DFT labels to fight MLIP barrier softening.** ML interatomic potentials systematically *under-predict* activation barriers. Computing single-point energies/forces on these saddles + endpoints with DFT yields targeted training data that pushes MLIPs toward correct barrier heights without requiring full DFT saddle searches. 3. **Warm-starting DFT saddle searches.** The ML-relaxed saddles are usually close enough to the DFT minimum that running a Dimer/NEB at DFT level converges in a small number of force evaluations. --- ## Loading the data ### Requirements ```bash pip install "ase>=3.26.0" ase_db_backends ``` `ase_db_backends` registers the `aselmdb` backend so `ase.db.connect(path, type="aselmdb")` works directly. **No fairchem-core install is required to read the data.** If you happen to have fairchem-core in your environment already, an `import fairchem.core.datasets` before `connect` works as a fallback registration on older stacks. ### ⚠ The `atoms.info` reconstruction trap ASE's `aselmdb` backend does **not** round-trip `atoms.info`. Calling `row.toatoms()` returns an `Atoms` object whose `.info` is empty — the full original `info` dict (every metadata key documented above, including nested `orig_info` and the `eigenmode` ndarray) lives in `row.data["info"]`. Always use the canonical reader helper: ```python def row_to_atoms(row): atoms = row.toatoms() atoms.info.update(row.data["info"]) # restore the original info dict return atoms ``` ### Minimal example ```python from ase.db import connect db = connect("lemat/lemat_dimer_000.aselmdb", type="aselmdb") for row in db.select(limit=10): atoms = row_to_atoms(row) print(atoms.get_chemical_formula(), "side=", atoms.info.get("side"), "image_type=", atoms.info.get("image_type")) ``` ### Walking the rows in triplets ```python from ase.db import connect db = connect("lemat/lemat_dimer_000.aselmdb", type="aselmdb") print(len(db), "rows ->", len(db) // 3, "transition states") batch = [] for row in db.select(): batch.append(row_to_atoms(row)) if len(batch) == 3: reactant, saddle, product = batch print(saddle.get_chemical_formula(), "eigenmode", saddle.info.get("eigenmode").shape if "eigenmode" in saddle.info else None, "curvature", saddle.info.get("curvature"), "barrier", saddle.info.get("barrier")) batch = [] ``` The same loop works without modification on every subset (`lemat`, `oc20`, `oc22`, `mp20bat`). --- ## Train / val / test splits A precomputed **stratified 90 / 5 / 5** train / val / test split is published alongside the data under [`splits/`](./splits/). The split is: - **Stratified by subset** — each of `lemat`, `oc20`, `oc22`, `mp20bat` is independently split 90 / 5 / 5, so the global split preserves subset proportions exactly. - **Triplet-level** — the three rows of a transition-state event (reactant, saddle, product) always land in the same split. - **Deterministic** — produced with NumPy seed 42 from the global `ms_id` enumeration. Re-running the build script reproduces the same assignment. ### Layout ``` splits/ ├── lemat/ train.parquet val.parquet test.parquet ├── oc20/ train.parquet val.parquet test.parquet ├── oc22/ train.parquet val.parquet test.parquet └── mp20bat/ train.parquet val.parquet test.parquet ``` Each parquet file has a single `ms_id` (uint32) column listing all `ms_id`s that belong to that `(subset, split)` bucket, sorted ascending. Every triplet contributes its three consecutive ms_ids (`3k`, `3k+1`, `3k+2`). ### Counts | subset | total | train | val | test | |-----------|------------:|------------:|-----------:|-----------:| | `lemat` | 31,346,419 | 28,211,777 | 1,567,321 | 1,567,321 | | `oc20` | 2,587,101 | 2,328,391 | 129,355 | 129,355 | | `oc22` | 167,335 | 150,602 | 8,367 | 8,366 | | `mp20bat` | 34,742 | 31,268 | 1,737 | 1,737 | ### Using the splits Train on a single subset: ```python import polars as pl train_ms_ids = pl.read_parquet("splits/lemat/train.parquet")["ms_id"].to_numpy() ``` Train on the union of all subsets: ```python import polars as pl train_ms_ids = pl.concat([ pl.read_parquet(f"splits/{s}/train.parquet") for s in ("lemat", "oc20", "oc22", "mp20bat") ])["ms_id"].to_numpy() ``` Route rows from an aselmdb shard at iteration time. Because each parquet column is sorted, `np.searchsorted` is the fastest membership test: ```python import numpy as np from ase.db import connect import fairchem.core.datasets # registers aselmdb backend train_ms_ids = ... # loaded as above; sorted uint32 array db = connect("lemat/lemat_dimer_000.aselmdb", type="aselmdb") for row in db.select(): ms_id = row.data["info"]["ms_id"] idx = np.searchsorted(train_ms_ids, ms_id) if idx < len(train_ms_ids) and train_ms_ids[idx] == ms_id: # row is in the training split — feed it to your trainer ... ``` The same parquet files can be loaded with `datasets`: ```python from datasets import load_dataset ds = load_dataset( "SciLM/MaterialsSaddles", data_files={"train": "splits/lemat/train.parquet", "val": "splits/lemat/val.parquet", "test": "splits/lemat/test.parquet"}, ) ``` ### Reproducing the split The 12 parquet files are produced deterministically: shuffle each subset's triplet indices with NumPy seed 42, take the first 90% as train, next 5% as val, last 5% as test, then expand each triplet `t` to its three consecutive `ms_id`s `{3t, 3t+1, 3t+2}`. The per-subset triplet counts in the table above are sufficient to regenerate every file under `splits/` byte-identically. --- ## How the data was produced We took fully relaxed structures from four public datasets (LeMat-Bulk, OC20, OC22, and Materials Project battery structures) and ran high-throughput saddle searches against each one using the [SaddleMill](https://github.com/SciLM-ai/SaddleMill) package, with Meta's **`uma-s-1p2`** universal interatomic potential ([fairchem-core](https://github.com/facebookresearch/fairchem)) as the calculator. | Subset | Method | SaddleMill entrypoint | |-----------|----------|-----------------------------------| | `lemat` | Dimer | `SaddleMill.dimeropt` | | `oc20` | Dimer | `SaddleMill.dimeropt` | | `oc22` | Dimer | `SaddleMill.dimeropt` | | `mp20bat` | NEB-CI | `SaddleMill.nebopt` (climbing image) | Initialization protocol (per-subset displacement modes such as `vacancy`, `hop_insert`, `kickout_*`, `ring`, `adsorbate_atom`, `diffusion`, `rotation`, …), eigenmode refinement, and post-search filtering are documented in the companion paper. After saddle convergence, every TS was validated by **DoubleMinimization** — displacing along the eigenmode in both directions and relaxing — and only triplets where the resulting endpoints actually correspond to two distinct basins (i.e. a real reaction occurred) are kept here. Anything that errored, hit a step limit, desorbed, or failed the reaction check is excluded. --- ## Known limitations - **MLIP, not DFT.** All saddles and endpoints in this release were converged with the `uma-s-1p2` MLIP rather than DFT. ML interatomic potentials systematically *under-predict* activation barriers, so the geometries here should be treated as approximate transition states. For DFT-level accuracy, run a single-point or short DFT saddle/NEB starting from these structures. - **`atoms.info` is not auto-restored by `row.toatoms()`.** See the [trap callout above](#-the-atomsinfo-reconstruction-trap). Always use the `row_to_atoms` helper. - **`row.key_value_pairs` queries are linear scans.** ASE's aselmdb backend has no secondary indices; `db.select(side=0)` reads every row. For large filters, iterate the rows yourself in shard order. - **Multi-shard cursors are user code.** Each shard is independent; if you need to iterate the entire subset (or use the [`splits/`](./splits/) manifests), open shards in sequence and route rows by `row.data["info"]["ms_id"]`. - **Schema varies by source.** Only `task_name` and `ms_id` are guaranteed on every row. NEB-derived rows (e.g. `mp20bat`) have a different `info` schema than dimer rows (e.g. no `side`, but `image_type` / `image_idx` / `barrier` instead). Inspect `row.data["info"].keys()` if you need to discover what's actually there for a given subset. --- ## Citation If you use this dataset, please cite: ```bibtex @misc{materialssaddles2026, title = {{MaterialsSaddles}: 34 Million Transition States and a Flow-Matching Saddle-Point Predictor for Materials}, author = {Baghishov, Ilgar and Jung, Sung Hoon and Henkelman, Graeme}, year = {2026}, note = {Submitted to the Datasets \& Benchmarks Track, NeurIPS 2026}, howpublished = {\url{https://huggingface.co/datasets/SciLM/MaterialsSaddles}} } ``` …and the upstream sources you actually used: ```bibtex @article{chanussot2021oc20, title = {Open Catalyst 2020 (OC20) Dataset and Community Challenges}, author = {Chanussot, Lowik and Das, Abhishek and Goyal, Siddharth and others}, journal = {ACS Catalysis}, volume = {11}, pages = {6059--6072}, year = {2021}, doi = {10.1021/acscatal.0c04525} } @article{tran2023oc22, title = {The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts}, author = {Tran, Richard and Lan, Janice and Shuaibi, Muhammed and others}, journal = {ACS Catalysis}, volume = {13}, pages = {3066--3084}, year = {2023}, doi = {10.1021/acscatal.2c05426} } @article{jain2013mp, title = {Commentary: The {Materials Project}: A materials genome approach to accelerating materials innovation}, author = {Jain, Anubhav and Ong, Shyue Ping and Hautier, Geoffroy and others}, journal = {APL Materials}, volume = {1}, number = {1}, pages = {011002}, year = {2013}, doi = {10.1063/1.4812323} } @misc{lemat-bulk, title = {{LeMat-Bulk}: A unified, deduplicated dataset of bulk crystal structures}, author = {{Entalpic} and {Hugging Face}}, year = {2024}, note = {\url{https://huggingface.co/datasets/LeMaterial/LeMat-Bulk}} } @misc{uma2025, title = {{UMA}: A Family of Universal Models for Atoms}, author = {{Meta FAIR Chemistry}}, year = {2025}, note = {\url{https://github.com/facebookresearch/fairchem} -- model {\tt uma-s-1p2}} } @article{ase, title = {The atomic simulation environment---a {Python} library for working with atoms}, author = {Larsen, Ask Hjorth and Mortensen, Jens J{\o}rgen and Blomqvist, Jakob and others}, journal = {Journal of Physics: Condensed Matter}, volume = {29}, pages = {273002}, year = {2017}, doi = {10.1088/1361-648X/aa680e} } ``` > Several of the entries above are placeholders or trimmed; please verify > the canonical version against the publisher before submission. ## License This dataset is released under [Creative Commons Attribution 4.0 International (CC-BY-4.0)](https://creativecommons.org/licenses/by/4.0/). The upstream datasets retain their own licenses; consult them before any redistribution that combines this dataset with theirs. ## Changelog - **v1** — initial public release. ## Contact Website: . Issues / questions: open a discussion on the [Hugging Face Hub page](https://huggingface.co/datasets/SciLM/MaterialsSaddles).