The QCML dataset, Quantum chemistry reference data from 33.5M DFT and 14.7B semi-empirical calculations

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Machine learning (ML) models predict chemical properties, reducing the need for expensive ab initio calculations.
• The accuracy of ML predictions relies heavily on the quality and comprehensiveness of the training data.
• Existing datasets may not sufficiently cover the diverse chemical space required for robust ML model development.

Purpose of the Study:

• Introduce the Quantum Chemistry Machine Learning (QCML) dataset, a novel resource for training ML models in quantum chemistry.
• Systematically explore chemical space with small molecules, diverse elements, and varied electronic states.
• Provide a foundation for developing accurate and efficient ML-based computational chemistry tools.

Main Methods:

• Generated 3D molecular structures from chemical graphs using conformer search and normal mode sampling.
• Calculated molecular properties for equilibrium and off-equilibrium structures using semi-empirical methods (14.7 billion entries) and density functional theory (33.5 million entries).
• Included properties such as energies, forces, multipole moments, and Kohn-Sham matrices.

Main Results:

• The QCML dataset systematically covers chemical space for molecules up to 8 heavy atoms, including a wide range of elements and electronic states.
• A vast number of quantum mechanical properties were computed, providing a rich source for ML model training.
• Demonstrated the dataset's utility by training ML force fields and performing molecular dynamics simulations.

Conclusions:

• The QCML dataset is a valuable resource for advancing ML in quantum chemistry.
• The dataset facilitates the development of accurate ML models for predicting molecular properties.
• The successful application in molecular dynamics simulations highlights the practical utility of the QCML dataset.