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

Area of Science:

• Computational quantum chemistry
• Materials science
• Drug discovery

Background:

• First-principles methods like DFT are accurate but computationally expensive.
• Machine learning interatomic potentials (MLIPs) offer a computationally efficient alternative.
• Accurate MLIPs require large, high-quality datasets with energy and force labels.

Purpose of the Study:

• Introduce PubChemQCR, a large-scale dataset of molecular relaxation trajectories.
• Provide a benchmark for MLIP models.
• Facilitate the development of accurate and transferable MLIPs.

Main Methods:

• Curated molecular relaxation trajectories from PubChemQC project outputs.
• Included DFT-based calculations for small organic molecules.
• Labeled each conformation with total energy and atomic forces.

Main Results:

• PubChemQCR is the largest public dataset of DFT-based relaxation trajectories.
• Contains ~3.5 million trajectories and over 300 million conformations.
• Benchmarked nine representative MLIP models on the dataset.

Conclusions:

• PubChemQCR enables training and evaluation of MLIPs for DFT-level accuracy.
• Facilitates efficient large-scale atomistic simulations.
• Publicly available resource for the scientific community.