GMD-26: Benchmarking Compositional Generalisation for Machine Learning Force Fields

GMD-26 is a systematic benchmark dataset for evaluating the compositional generalisation capabilities of Machine Learning Force Fields (MLFFs). Unlike standard benchmarks that train and test on the same molecular systems, GMD-26 explicitly separates training and test molecules, probing whether models have learned transferable physical principles or merely interpolate within their training distribution.

Paper: Benchmarking Compositional Generalisation for Learning Inter-atomic Potentials (submitted to NeurIPS 2026)

---

Dataset Overview

GMD-26 covers 118 molecules across ten fragment groups based on substituted linear alkyl carbon chains, yielding over 296,000 labelled geometries. Each molecule has two AIMD trajectories:

• Primary trajectory: ~2,013 snapshots — used for training and OOD test splits
• Secondary trajectory: 500 snapshots — used for in-distribution (ID) test splits (first 100 for hyperparameter tuning, remaining 400 for ID evaluation)

Energies and forces are labelled at the PBE/def2-TZVP + D3BJ level of theory using ORCA, making this dataset directly compatible with standard MLFF training pipelines.

---

Molecular Families

Functional Group	Example Molecules
Alkanes	Ethane – Tridecane
Primary Alcohols	Ethanol – Pentadecan-1-ol
Aldehydes	Ethanal – Pentadecanal
Carboxylic Acids	Ethanoic acid – Pentadecanoic acid
Primary Amines	Ethanamine – Octan-1-amine
Primary Amides	Butanamide – Octanamide
Diamines	Ethane-1,2-diamine – Decane-1,10-diamine
Dicarboxylic Acids	Ethanedioic acid – Hexadecanedioic acid
Amino Acids	2-Aminoethanoic acid – 10-Aminodecanoic acid
Complex Multifunctional	Various diols, triols, tetraols, diones, dialdehydes

---

Benchmark Tasks

GMD-26 defines four tasks that each require a distinct form of compositional generalisation. In all tasks, training and OOD test molecules are disjoint.

Task 1 — Fragment Chain Extension

Tests whether a model can extrapolate to longer carbon chains than seen during training.

• Base: Train on alkanes C2–C6; test on alkanes C7–C13
• Augmented: Train on short/long alcohols and medium carboxylic acids; test on the complementary length ranges of each family

Task 2 — Fragment Composition

Tests whether a model can generalise to a novel functional group that is a chemical composition of functional groups seen individually during training.

• Base: Train on alcohols and aldehydes; test on carboxylic acids (–COOH ≈ –OH + –CHO)
• Augmented: Training additionally includes amines and amides, providing an explicit demonstration of functional group composition

Task 3 — Fragment Duplication

Tests whether a model can generalise from a molecule with one occurrence of a functional group to the same molecule with two occurrences.

• Train on monocarboxylic acids (C5–C10); test on the corresponding dicarboxylic acids of identical chain lengths

Task 4 — Fragment Combination

Tests whether a model can generalise to asymmetrically functionalised molecules when trained exclusively on symmetrically functionalised analogues.

• Train on diamines and dicarboxylic acids (C2–C9); test on the corresponding amino acids (one amine + one carboxylic acid)

Task Splits at a Glance

Task	Training Set	OOD Test Set
1 Base	C2–C6 alkanes	C7–C13 alkanes
1 Augmented	C2–C3 & C9–C15 alcohols; C4–C8 carboxylic acids	C4–C8 alcohols; C2–C3 & C9–C15 carboxylic acids
2 Base	C4–C10 alcohols, aldehydes; C7–C11 complex carbonyls/alcohols	C4–C10 carboxylic acids
2 Augmented	As above + C4–C10 amines and amides	C4–C10 carboxylic acids
3	C5–C10 monocarboxylic acids	C5–C10 dicarboxylic acids
4	C2–C9 diamines; C2–C9 dicarboxylic acids	C2–C9 amino acids

---

Generation Pipeline

Each trajectory was generated via a four-stage pipeline:

1. Conformer generation — RDKit generates an initial 3D geometry from a SMILES string
2. Trajectory sampling — FlashMD (PET-MAD / OMat-PES checkpoint) propagates molecular dynamics in vacuum for 2×10⁵ steps at 16 fs timestep, Langevin thermostat at 300 K, recording every 100 steps (~2,000 snapshots)
3. DFT relabelling — Each snapshot is relabelled at PBE/def2-TZVP + D3BJ using ORCA (!EnGrad PBE D3BJ def2-TZVP RI TightSCF); snapshots where SCF convergence fails are excluded
4. Packaging — Outputs stored in extended XYZ format (compatible with ASE) alongside pre-computed task splits

---

Data Format

Data is stored in ASE trajectory format (.traj), readable natively with the Atomic Simulation Environment. Each frame contains:

Field	Description
positions	Atomic coordinates (Å)
numbers	Atomic numbers
energy	Total PBE/def2-TZVP+D3BJ energy (eV)
forces	Per-atom force vectors (eV/Å)

Pre-computed split files for all four tasks are included so that reproducing the evaluation protocol does not require re-running any upstream generation stages.

---

Loading the Dataset

# Using huggingface_hub (raw files)
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="AmirMasoud/GMD-26",
    repo_type="dataset",
    local_dir="./GMD-26"
)

# Using ASE to read individual trajectory files
from ase.io.trajectory import Trajectory

traj = Trajectory("path/to/molecule.traj")
energies = [atoms.get_potential_energy() for atoms in traj]
forces   = [atoms.get_forces() for atoms in traj]

---

GOAL Training Framework

Alongside the dataset we release GOAL (available upon paper acceptance), a PyTorch Lightning-based framework that:

• Exposes GMD-26 as a benchmark module with a unified data-loading interface
• Supports from-scratch training of arbitrary MLFF architectures via a single forward(graph) -> dict method
• Wraps foundation models (MACE-MP-0 small/medium/large, UMA-small) through thin adapters for fine-tuning under identical conditions
• Supports DDP, FSDP, and FSDP2 for multi-GPU and multi-node training (tested on NVIDIA GH200)
• Configures all experiments through Hydra for full reproducibility
• Exposes trained models as ASE calculators for downstream MD simulation

Extending the Benchmark

New molecular families can be added by providing SMILES strings and a YAML task specification:

task: chain_extension
fragment: alcohol
train_lengths: [2, 3, 4, 5, 6]
ood_lengths: [7, 8, 9, 10, 11, 12, 13]
n_snapshots_per_trajectory: 1000

The four task templates are defined over abstract fragment sets and reusable without modification.

---

Citation

If you use GMD-26 in your research, please cite:

@inproceedings{nourollah2026gmd26,
  title     = {Benchmarking Compositional Generalisation for Learning Inter-atomic Potentials},
  author    = {Nourollah, Amir Masoud and Khalid, Irtaza and Leoni, Stefano and Schockaert, Steven},
  booktitle = {The Fourteenth International Conference on Learning Representations (ICLR 2026)},
  year      = {2026},
  url       = {https://openreview.net/forum?id=WxTlAbRUE6}
}

Note: The paper is currently under review. If accepted, the citation will be updated with the final publication details.

---

Licence

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence.