Clusterome: A Comprehensive Data Set of Atmospheric Molecular Clusters for Machine Learning Applications

Area of Science:

• Atmospheric Chemistry
• Climate Science
• Computational Chemistry

Background:

• Atmospheric molecular clusters influence global climate and introduce uncertainty into climate models.
• Current quantum chemical methods for studying cluster formation are computationally intensive, limiting system size and scope.
• Accurate modeling of aerosol particle formation is essential for understanding climate dynamics.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting atmospheric molecular cluster structures.
• To create a large database of atmospheric relevant cluster structures for training ML models.
• To assess the extrapolation and transferability capabilities of ML models for atmospheric clusters.

Main Methods:

• Compiled a database of approximately 250,000 atmospheric relevant cluster structures.
• Trained a kernel ridge regression (KRR) ML model using the FCHL19 molecular representation.
• Evaluated the ML model's performance in extrapolating to larger clusters, different molecules, and new interactions.

Main Results:

• The KRR ML model demonstrated strong ability to extrapolate to larger cluster sizes.
• The model accurately transferred acid and base interactions with mean absolute errors below 1 kcal/mol.
• The ML approach showed potential for predicting out-of-equilibrium structures and systems with new interactions.

Conclusions:

• Kernel ridge regression models offer a computationally efficient alternative to quantum chemical methods for studying atmospheric clusters.
• Integrating ML into configurational sampling can significantly reduce computational expense.
• This ML-driven approach enables the study of a wider range of atmospheric compounds with higher accuracy, improving climate model reliability.