The AQCat25 models and dataset are for non-commercial use. If you want to work with us or discuss obtaining a commercial license, please contact us directly at aqcat25@sandboxaq.com.

AQCat25: Unlocking spin-aware, high-fidelity machine learning potentials for heterogeneous catalysis

This repository contains AQCat-EV2 model checkpoints and the AQCat25 dataset. The AQCat25 dataset provides a large and diverse collection of 13.5 million DFT calculation trajectories, encompassing approximately 5K materials and 47K intermediate-catalyst systems. It is designed to complement existing large-scale datasets by providing calculations at higher fidelity and including critical spin-polarized systems, which are essential for accurately modeling many industrially relevant catalysts.

The AQCat-EV2 models provided here are designed to solve the challenge of training on this mixed-physics (spin-on/off) and mixed-fidelity (high/low) data. The models released currently are based on the EquiformerV2 (EV2) architecture, wherein scalar activations are additively modulated by a Feature-wise Linear Modulation (FiLM) network that explicitly conditions on the underlying DFT settings (i.e., "is this calculation high-fidelity?" and "is spin on?").

Stay tuned for additional models and complete dataset splits that will be released in the near future.

Please see our website and paper for more details about the impact of the dataset and models.

1. Model Installation and Usage (EV2-FiLM)

This section details how to install and run the EquiformerV2-FiLM (EV2-FiLM) model.

Step 1.1: Create Environment

First, create and activate a new micromamba (or conda) environment with Python 3.10.

micromamba create -n aqcat-ev2 python=3.10
micromamba activate aqcat-ev2

Step 1.2: Install Dependencies

Before installing fairchem, install PyTorch and all other required libraries.

pip install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu121 --no-input
pip install torch_geometric --no-input
pip install torch_sparse -f https://data.pyg.org/whl/torch-2.4.0+cu121.html --no-input
pip install torch_scatter -f https://data.pyg.org/whl/torch-2.4.0+cu121.html --no-input
pip install e3nn submitit torchtnt hydra-core pymatgen ase orjson wandb tensorboard lmdb huggingface-hub numba datasets pandas tqdm requests --no-input

Step 1.3: Log in to Hugging Face

To download the model files and dataset, you must log in to your Hugging Face account.

Create an Access Token: Navigate to your Settings -> Access Tokens page or click here. Create a new token with at least read permissions. Copy this token.

Log in via the Command Line: Open your terminal, run the following command, and paste your token when prompted.

hf auth login

Step 1.4: Download the AQCat Model Files

Next, download the necessary model checkpoints and scripts from this Hugging Face repository.

Save the following code as download_files.py. You will need to edit it to add your Hugging Face token.

from huggingface_hub import snapshot_download

MY_TOKEN = "hf_YOUR_TOKEN_HERE" 

print("Downloading model and code files...")

snapshot_download(
    repo_id="SandboxAQ/aqcat25",
    repo_type="dataset",
    allow_patterns=[
        "scripts/*", 
        "checkpoints_aqcat_ev2/*", 
        "ev2_film/*",
        "patched_code/*",
    ],
    local_dir="./aqcat25",
    token=MY_TOKEN
)

print("Download complete.")

Now, run the script. This will create a new folder named aqcat25 containing all the necessary files.

python download_files.py

Step 1.5: Clone, Patch, and Install `fairchem`

Finally, we will clone the fairchem repo, check out the correct V1 code, copy our custom files into it, and install the modified version.

git clone git@github.com:facebookresearch/fairchem.git
cd fairchem
git fetch --all --tags
git checkout -b aqcat-ev2 tags/fairchem_core-1.10.0
cp ../aqcat25/ev2_film/equiformer_v2_film.py packages/fairchem-core/src/fairchem/core/models/equiformer_v2/
cp ../aqcat25/patched_code/ase_utils.py packages/fairchem-core/src/fairchem/core/common/relaxation/ase_utils.py
pip install -e packages/fairchem-core --no-deps --no-input

Step 1.6: Model Checkpoints

This repository provides the following checkpoints.

Trained-from-scratch (EV2-in+midFiLM): A generally well-rounded model that provides the best performance on practical catalysis discovery tasks. It was jointly trained from scratch on both AQCat25 and 20M examples from OC20. This model particularly excels on the most challenging material subclasses, such as non-metals and organics. See the paper for more details.
Cotuned (EV2-inFiLM): This model uses the pre-trained EV2-31M (OC20 All+MD) checkpoint as its starting point and is then fine-tuned on AQCat25 while replaying 20M examples from OC20. It offers strong performance particularly on metal-only systems.
Directly Tuned (Baselines): These are the pre-trained EV2-31M and EV2-153M (OC20 All+MD) checkpoints that have been directly fine-tuned on AQCat25 with no OC20 replay. They serve as baselines for comparison to show the effects of co-tuning.

Step 1.7: Model Usage Example

Setup is complete. Your aqcat-ev2 environment now has the patched fairchem v1.10.0 installed.

Here is a full example of relaxing a carbon monoxide molecule on a cobalt slab.

The patched_calc provided in the installation guide defaults to the high-fidelity context and will detect if spin polarization is needed based on the elements in your system (e.g., Co, Fe, Ni), though you can also toggle these flags manually as shown below.

import numpy as np

from ase.build import hcp0001, molecule, add_adsorbate
from ase.constraints import FixAtoms
from ase.optimize import LBFGS
from ase.io import write

from fairchem.core.common.relaxation.ase_utils import patched_calc

CHECKPOINT_PATH = "aqcat25/checkpoints_aqcat_ev2/ev2-in+midFiLM-AQCat25+OC20-20M_20251008_223220.pt"

calc = patched_calc(checkpoint_path=CHECKPOINT_PATH)

slab = hcp0001('Co', size=(3, 4, 4), orthogonal=True)
co_molecule = molecule('CO') 
add_adsorbate(slab, co_molecule, 3.0, 'ontop')
slab.center(vacuum=10.0, axis=2)
slab.set_pbc(True)

slab_symbols = slab.get_chemical_symbols()
is_slab_atom = np.array([sym == 'Co' for sym in slab_symbols])
slab_z_coords = slab.get_positions()[is_slab_atom][:, 2]
unique_slab_z = np.unique(slab_z_coords)
unique_slab_z.sort()
top_layer_z = unique_slab_z[-1] 
mask = slab.get_positions()[:, 2] < top_layer_z - 0.1
num_fixed = mask.sum() 
slab.set_constraint(FixAtoms(mask=mask))

# Run with spin polarization context
slab_spin_on = slab.copy()
slab_spin_on.info['is_spin_off'] = False
slab_spin_on.calc = patched_calc(checkpoint_path=CHECKPOINT_PATH)
optimizer_on = LBFGS(slab_spin_on, trajectory='co_co_spin_on.traj')
optimizer_on.run(fmax=0.05)
final_energy_spin_on = slab_spin_on.get_potential_energy()

# Run without spin polarization context
slab_spin_off = slab.copy()
slab_spin_off.info['is_spin_off'] = True
slab_spin_off.calc = patched_calc(checkpoint_path=CHECKPOINT_PATH)
optimizer_off = LBFGS(slab_spin_off, trajectory='co_co_spin_off.traj')
optimizer_off.run(fmax=0.05)
final_energy_spin_off = slab_spin_off.get_potential_energy()


print("\n--- Final Comparison ---")
print(f"Spin-On Energy:    {final_energy_spin_on:.4f} eV")
print(f"Spin-Off Energy: {final_energy_spin_off:.4f} eV")
print(f"Energy Difference (Off - On): {(final_energy_spin_off - final_energy_spin_on):.4f} eV")

Understanding the example

You should observe a difference between the two final adsorption energies. The spin-unpolarized run will likely have a lower (more negative) final energy, which indicates a stronger, more stable binding.

This is the correct behavior for the model, as it has learned the underlying physics of magnetic systems from the AQCat25 and OC20 datasets.

As explained in the paper "Spin Effects in Chemisorption and Catalysis" (ACS Catal. 2023, 13, 3456-3462), for 3d magnetic metals like Co, Fe, and Ni, the true spin-polarized ground state (is_spin_off=False) actually results in weaker adsorbate bonding compared to a hypothetical non-spin-polarized state (is_spin_off=True). This is because the energetic stabilization from the spin-down (minority-spin) d-states does not fully compensate for the destabilization from the spin-up (majority-spin) d-states.

By toggling the is_spin_off flag, you are telling the model to apply a different physical context, and the model predicts a different, more stable energy for the (hypothetical) non-spin-polarized system.

2. AQCat25 Dataset Details

This repository uses a hybrid approach, providing lightweight, queryable Parquet files for each split alongside compressed archives (.tar.gz) of the raw ASE database files. More details can be found below.

Queryable Metadata (Parquet Files)

A set of Parquet files provides a "table of contents" for the dataset. They can be loaded directly with the datasets library for fast browsing and filtering. Each file contains the following columns:

Column Name	Data Type	Description	Example
`frame_id`	string	Unique ID for this dataset. Formatted as `database_name::index`.	`data.0015.aselmdb::42`
`adsorption_energy`	float	Key Target. The calculated adsorption energy in eV.	-1.542
`total_energy`	float	The raw total energy of the adslab system from DFT (in eV).	-567.123
`fmax`	float	The maximum force magnitude on any single atom in eV/Å.	0.028
`is_spin_off`	boolean	`True` if the system is non-magnetic (VASP ISPIN=1).	`false`
`mag`	float	The total magnetization of the system (µB).	32.619
`slab_id`	string	Identifier for the clean slab structure.	`mp-1216478_001_2_False`
`adsorbate`	string	SMILES or chemical formula of the adsorbate.	`*NH2N(CH3)2`
`is_rerun`	boolean	`True` if the calculation is a continuation.	`false`
`is_md`	boolean	`True` if the frame is from a molecular dynamics run.	`false`
`sid`	string	The original system ID from the source data.	`vadslabboth_82`
`fid`	integer	The original frame index (step number) from the source VASP calculation.	0

Understanding `frame_id` and `fid`

Field	Purpose	Example
`fid`	Original Frame Index: This is the step number from the original VASP relaxation (`ionic_steps`). It tells you where the frame came from in its source simulation.	`4` (the 5th frame of a specific VASP run)
`frame_id`	Unique Dataset Pointer: This is a new ID created for this specific dataset. It tells you exactly which file (`data.0015.aselmdb`) and which row (`101`) to look in to find the full atomic structure.	`data.0015.aselmdb::101`

Downloadable Data Archives

The full, raw data for each split is available for download in compressed .tar.gz archives. The table below provides direct download links. The queryable Parquet files for each split can be loaded directly using the datasets library as shown in the "Example Usage" section. The data currently available for download (totaling ~11.1M frames, as listed in the table below) is the initial dataset version (v1.0) released on September 10, 2025. The 13.5M frame count mentioned in our paper and the introduction includes additional data used to rebalance non-magnetic element systems and add a low-fidelity spin-on dataset. These new data splits will be added to this repository soon.

Split Name	Structures	Archive Size	Download Link
In-Domain (ID)
Train	`7,386,750`	`23.8 GB`	`train_id.tar.gz`
Validation	`254,498`	`825 MB`	`val_id.tar.gz`
Test	`260,647`	`850 MB`	`test_id.tar.gz`
Slabs	`898,530`	`2.56 GB`	`id_slabs.tar.gz`
Out-of-Distribution (OOD) Validation
OOD Ads (Val)	`577,368`	`1.74 GB`	`val_ood_ads.tar.gz`
OOD Materials (Val)	`317,642`	`963 MB`	`val_ood_mat.tar.gz`
OOD Both (Val)	`294,824`	`880 MB`	`val_ood_both.tar.gz`
OOD Slabs (Val)	`28,971`	`83 MB`	`val_ood_slabs.tar.gz`
Out-of-Distribution (OOD) Test
OOD Ads (Test)	`346,738`	`1.05 GB`	`test_ood_ads.tar.gz`
OOD Materials (Test)	`315,931`	`993 MB`	`test_ood_mat.tar.gz`
OOD Both (Test)	`355,504`	`1.1 GB`	`test_ood_both.tar.gz`
OOD Slabs (Test)	`35,936`	`109 MB`	`test_ood_slabs.tar.gz`

3. Dataset Usage Guide

This guide outlines the recommended workflow for accessing and querying the AQCat25 dataset files.

(Note: Helper scripts were already downloaded in Step 1.4)

Step 3.1: Download Desired Dataset Splits

Data splits may be downloaded directly via the Hugging Face UI, or via the download_split.py script (found in aqcat25/scripts/).

python aqcat25/scripts/download_split.py --split val_id

This will download val_id.tar.gz and extract it to a new folder named aqcat_data/val_id/.

Step 3.2: Query the Dataset

Use the query_aqcat.py script to filter the dataset and extract the specific atomic structures you need. It first queries the metadata on the Hub and then extracts the full structures from your locally downloaded files.

Example 1: Find all CO and OH structures in the test set:

python aqcat25/scripts/query_aqcat.py \
    --split test_id \
    --adsorbates "*CO" "*OH" \
    --data-root ./aqcat_data/test_id

Example 2: Find structures on metal slabs with low adsorption energy:

python aqcat25/scripts/query_aqcat.py \
    --split val_ood_both \
    --max-energy -2.0 \
    --material-type nonmetal \
    --magnetism magnetic \
    --data-root ./aqcat_data/val_ood_both \
    --output-file low_energy_metals.extxyz

Example 3: Find CO on slabs containing both Ni AND Se with adsorption energy between -2.5 and -1.5 eV with a miller index of 011

python aqcat25/scripts/query_aqcat.py \
    --split val_ood_ads \
    --adsorbates "*COCH2OH" \
    --min-energy -2.5 \
    --max-energy -1.5 \
    --contains-elements "Ni" "Se" \
    --element-filter-mode all \
    --facet 011 \
    --data-root ./aqcat_data/val_ood_ads \
    --output-file COCH2OH_on_ni_and_se.extxyz

4. How to Cite

If you use the AQCat25 dataset or the models in your research, please cite the following paper:

Omar Allam, Brook Wander, & Aayush R. Singh. (2025). AQCat25: Unlocking spin-aware, high-fidelity machine learning potentials for heterogeneous catalysis. arXiv preprint arXiv:XXXX.XXXXX.

BibTeX Entry

@article{allam2025aqcat25,
  title={{AQCat25: Unlocking spin-aware, high-fidelity machine learning potentials for heterogeneous catalysis}},
  author={Allam, Omar and Wander, Brook and Singh, Aayush R},
  journal={arXiv preprint arXiv:2510.22938},
  year={2025},
  eprint={2510.22938},
  archivePrefix={arXiv},
  primaryClass={cond-mat.mtrl-sci}
}

Downloads last month: 219

You need to agree to share your contact information to access this dataset