QMe14S: A Comprehensive and Efficient Spectral Data Set for Small Organic Molecules.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning models for molecular simulations require extensive and high-quality datasets.
• Existing datasets may not cover the breadth of elements and functional groups needed for comprehensive simulations.

Purpose of the Study:

• Introduce the QMe14S dataset, a large and diverse collection of small organic molecules.
• Provide a benchmark for developing and evaluating machine learning protocols in molecular simulations.
• Investigate structure-property relationships using advanced computational methods.

Main Methods:

• Generated QMe14S dataset with 186,102 molecules across 14 elements and 47 functional groups.
• Employed density functional theory (DFT) at the B3LYP/TZVP level for geometry optimization and property calculations.
• Conducted ab initio molecular dynamics (AIMD) simulations for dynamic configurations and nonequilibrium properties.
• Utilized an E(3)-equivariant message-passing neural network (DetaNet) for model training and performance evaluation.

Main Results:

• QMe14S dataset includes optimized geometries and calculated properties (energy, charge, force, moments, polarizability, Hessian) for diverse molecules.
• Harmonic IR, Raman, and NMR spectra were computed at the B3LYP/TZVP level.
• AIMD simulations yielded dynamic configurations and nonequilibrium properties.
• Machine learning models trained on QMe14S demonstrated improved accuracy in simulating molecular spectra compared to models trained on QM9S.

Conclusions:

• The QMe14S dataset is a valuable resource for advancing machine learning in molecular simulations.
• It provides a robust benchmark for assessing model performance and exploring structure-property relationships.
• The dataset facilitates the development of more accurate and efficient computational tools for chemical research.