Chemistrees: Data-Driven Identification of Reaction Pathways via Machine Learning.

Area of Science:

• Computational chemistry
• Machine learning applications
• Chemical reaction dynamics

Background:

• Molecular dynamics (MD) simulations generate vast datasets.
• Analyzing complex reaction pathways, like proton transfer, often relies on human interpretation.
• Identifying key molecular geometries driving state transitions is challenging.

Purpose of the Study:

• To develop and apply a supervised machine learning (ML) algorithm to analyze MD output.
• To identify geometric features correlating with state transitions in chemical reactions.
• To perform data-driven analysis without human chemical intuition bias.

Main Methods:

• Utilized a decision tree supervised ML algorithm for analyzing MD data.
• Employed *ab initio* MD simulations combined with path sampling for rare event sampling.
• Focused on proton exchange reactions in formic acid solvated by water clusters.

Main Results:

• The ML algorithm successfully identified predominant geometric features linked to proton transfer.
• Analysis revealed how these geometric variables change with the number of water molecules.
• The approach demonstrated objectivity in identifying reaction-driving features.

Conclusions:

• Supervised ML, specifically decision trees, is effective for analyzing MD simulations.
• This data-driven method can uncover key geometric insights into chemical reaction mechanisms.
• The findings provide a foundation for understanding solvation effects on proton transfer.