A Benchmark for Quantum Chemistry Relaxations via Machine Learning Interatomic Potentials.

Cong Fu, Yuchao Lin, Zachary Krueger, Wei-Kuan Yu, Xiaoning Qian, Byung-Jun Yoon, Raymundo Arróyave, Xiaofeng Qian, Toshiyuki Maeda, Maho Nakata, Shuiwang JiPubMed 2025

TL;DR

This paper introduces PubChemQCR, the largest publicly available dataset of molecular relaxation trajectories, to address the need for high-quality data in training machine learning interatomic potentials (MLIPs). They benchmark nine MLIP models on this dataset, providing a crucial resource for advancing accurate and efficient atomistic simulations.

What Problem Does This Paper Address?

Accurately simulating molecular behavior at the atomic scale is fundamental to fields like drug discovery and materials science. While first-principles methods like Density Functional Theory (DFT) offer high accuracy, they are computationally prohibitive for large-scale simulations. Machine Learning Interatomic Potentials (MLIPs) offer a promising alternative by learning from DFT data to predict energies and forces much faster. However, developing accurate and generalizable MLIPs hinges on large, high-quality datasets that capture not just stable molecular structures but also the intermediate, non-equilibrium conformations encountered during simulations. Existing datasets often lack sufficient diversity, force information, or comprehensive trajectory data, hindering MLIP development.

Methods

The core contribution is the creation of the PubChemQCR dataset. This dataset is curated from the raw geometry optimization outputs of the PubChemQC project. The original PubChemQC data involved a multi-stage optimization process:

  1. Initial 3D structure generation: Using OpenBabel with rule-based approaches and predefined ring templates.
  2. MMFF94 optimization: A preliminary optimization using a molecular mechanics force field.
  3. Conformer search: A weighted rotor conformer search.
  4. Semi-empirical optimization: Using the PM3 method for a fast, approximate refinement.
  5. Hartree-Fock optimization: Using the STO-6G basis set for higher quantum accuracy.
  6. DFT optimization: The most accurate stage, using B3LYP functional with the 6-31G* basis set, employing Firefly/SMASH for initial refinement and GAMESS for the final, rigorous optimization.

The PubChemQCR dataset extracts atomic numbers, coordinates, total energies, and atomic forces from each snapshot along these trajectories, with a particular focus on the DFT-optimized segments for high-fidelity labels. The authors also computed isolated atomic energies for each element to enable normalization of energies, which helps in training MLIPs by reducing systematic biases.

Datasets

  • PubChemQCR: The primary contribution, a new, large-scale dataset comprising approximately 3.5 million molecular relaxation trajectories and over 300 million molecular conformations. It includes data from PM3, Hartree-Fock, and DFT optimization stages, with a focus on the DFT data for MLIP training and evaluation.
  • PubChemQCR-S: A smaller, curated subset of PubChemQCR, containing 40,979 trajectories and 1.5 million DFT-calculated snapshots, specifically designed for ablation studies, hyperparameter tuning, and preliminary evaluations.
  • PubChemQC: The source project from which PubChemQCR was derived.
  • Molecule3D: A dataset previously curated from PubChemQC, but PubChemQCR provides more detailed trajectory data.

The paper also references and compares against several existing datasets, including QM9, QM7-X, ANI-1x, GEOM, MD17, MD22, OC20, OC22, MPTrj, and ∇²DFT, highlighting PubChemQCR's advantage in scale and comprehensive trajectory data.

Key Insights and Ideas

  • The critical need for trajectory data: MLIPs need to generalize not just to stable, relaxed geometries but also to intermediate, non-equilibrium conformations. Datasets containing full relaxation trajectories are essential for this.
  • Normalization of energies is key: Subtracting isolated atomic energies (atomization energies) from total energies significantly reduces the variance and centers the distribution, making the learning task for MLIPs easier and more efficient.
  • DFT data is paramount for high-fidelity MLIPs: While multi-stage optimization is used for efficiency, the DFT-level data provides the most accurate energy and force labels, making it the most critical for training robust MLIPs.
  • Benchmarking provides a baseline: The paper establishes a benchmark by evaluating nine representative MLIP models, offering valuable insights into their performance on this new, challenging dataset.

Key Results

  • PubChemQCR is the largest dataset of its kind: It contains ~3.5 million trajectories and over 300 million conformations, with ~105 million from DFT calculations.
  • MLIP performance varies significantly: On the PubChemQCR-S subset, Equiformer achieved the best overall performance in geometry optimization, with PaiNN showing strong energy and force prediction.
  • Larger datasets improve performance: When trained on the full PubChemQCR dataset, PaiNN and FAENet showed improved performance compared to training on the smaller PubChemQCR-S subset, underscoring the importance of data scale.
  • Geometry optimization is challenging: Even with models trained on near-optimal data, achieving high accuracy in geometry optimization (especially force convergence) remains difficult, with Equiformer showing the best results among the tested models.

Limitations and Future Work

  • Limited element coverage: The dataset currently includes only 25 chemical elements due to the compatibility of the chosen DFT functional and basis set. Future work could expand this coverage.
  • Inconsistent accuracy across stages: The sequential optimization process means that labels from different stages (PM3, HF, DFT) have varying levels of accuracy and are not directly comparable. The authors recommend using only DFT data for training.
  • Near-equilibrium focus: The reliance on DFT data restricts the dataset to the near-equilibrium region of the potential energy surface, where atomic forces are relatively small. This limits the MLIPs' ability to generalize to broader conformational spaces with larger force magnitudes.
  • Future work suggestions: The authors envision PubChemQCR as a resource for developing more accurate and transferable MLIPs, enabling efficient atomistic simulations, and facilitating downstream tasks like property prediction and 3D molecular generation. They also suggest improving label consistency by using a single, uniform DFT functional and basis set for the entire optimization process.

Summary updated

Abstract

Computational quantum chemistry plays a critical role in drug discovery, chemical synthesis, and materials science. While first-principles methods, such as density functional theory (DFT), provide high accuracy in modeling electronic structures and predicting molecular properties, they are computationally expensive. Machine learning interatomic potentials (MLIPs) have emerged as promising surrogate models that aim to achieve DFT-level accuracy while enabling efficient large-scale atomistic simulations. The development of accurate and transferable MLIPs requires large-scale, high-quality datasets with both energy and force labels. Critically, MLIPs must generalize not only to stable geometries but also to intermediate, non-equilibrium conformations encountered during atomistic simulations. In this work, we introduce PubChemQCR, a large-scale dataset of molecular relaxation trajectories curated from the raw geometry optimization outputs of the PubChemQC project. PubChemQCR is the largest publicly available dataset of DFT-based relaxation trajectories for small organic molecules, comprising approximately 3.5 million trajectories and over 300 million molecular conformations computed at various levels of theory. Each conformation is labeled with both total energy and atomic forces, making the dataset suitable for training and evaluating MLIPs. To provide baselines for future developments, we benchmark nine representative MLIP models on the dataset. Our resources are publicly available at https://huggingface.co/divelab.
BibTeX
@inproceedings{fu2025benchmark,
  author = {Cong Fu and Yuchao Lin and Zachary Krueger and Wei-Kuan Yu and Xiaoning Qian and Byung-Jun Yoon and Raymundo Arróyave and Xiaofeng Qian and Toshiyuki Maeda and Maho Nakata and Shuiwang Ji},
  title = {{A Benchmark for Quantum Chemistry Relaxations via Machine Learning Interatomic Potentials.}},
  booktitle = {PubMed},
  year = {2025},
}
Figure