Fast and Interpretable Machine Learning Modeling of Atmospheric Molecular Clusters | JoVE Visualize

Area of Science:

Atmospheric Chemistry
Computational Chemistry
Climate Science

Background:

Understanding atmospheric molecular cluster formation is crucial for accurate climate modeling and predicting new aerosol particle formation.
Current quantum chemistry methods provide high accuracy but are computationally expensive, limiting large-scale studies.
Developing efficient computational models is essential for advancing atmospheric chemistry research.

Purpose of the Study:

To present a fast, interpretable, and accurate alternative to computationally intensive methods for studying atmospheric molecular clusters.
To evaluate the performance of k-nearest neighbor (k-NN) regression models using chemically informed distance metrics.
To demonstrate the scalability and predictive power of k-NN models for atmospheric systems.

Main Methods:

Employed k-nearest neighbor (k-NN) regression models.
Utilized chemically informed distance metrics, including kernel-induced and metric learning for kernel regression (MLKR) metrics.
Compared k-NN performance against kernel ridge regression (KRR) using the FCHL19 molecular descriptor and other descriptors.
Applied models to QM9 benchmark data and large datasets of sulfuric acid-water and sulfuric acid-multibase-base clusters.

Main Results:

k-NN models achieved accuracy comparable to KRR models but with orders of magnitude reduction in computational time.
Models demonstrated near-chemical accuracy on both benchmark and large atmospheric cluster datasets (>250,000 entries).
The k-NN approach showed minimal error when extrapolating to larger, unseen clusters, often nearing 1 kcal/mol.

Conclusions:

k-nearest neighbor (k-NN) regression offers a computationally efficient and accurate method for studying atmospheric molecular clusters.
The developed k-NN models, with built-in interpretability and uncertainty estimation, can accelerate discovery in atmospheric chemistry.
This work positions k-NN as a powerful tool for improving climate models and understanding aerosol formation processes.