WS22 database, Wigner Sampling and geometry interpolation for configurationally diverse molecular datasets.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning models require extensive and diverse databases for quantum chemical properties.
• Exploring molecular configurational space is crucial for accurate property prediction.

Purpose of the Study:

• To create the WS22 database, a comprehensive resource for quantum mechanical (QM) properties.
• To provide a dataset that challenges machine learning models with broader configurational space sampling.

Main Methods:

• Composed the WS22 database with 1.18 million geometries for ten flexible organic molecules (up to 22 atoms).
• Sampled geometries from Wigner distributions for equilibrium and non-equilibrium states, including interpolated structures.
• Analyzed dataset diversity using dimensionality reduction and statistical property comparisons.

Main Results:

• The WS22 database covers potential energies, forces, dipole moments, polarizabilities, and HOMO/LUMO energies.
• Demonstrated dataset diversity through geometric distribution visualization and statistical analysis.
• The sampling strategy covers a wider QM distribution than classical molecular dynamics.

Conclusions:

• The WS22 database offers a valuable, diverse resource for developing robust machine learning models in quantum chemistry.
• The dataset's broad sampling presents a significant challenge, pushing the boundaries of current machine learning capabilities in molecular modeling.