MLSolvA: solvation free energy prediction from pairwise atomistic interactions by machine learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Machine learning (ML) advances enable sophisticated structure-property relationship models for chemical properties.
• Solvation free energy is a critical property for understanding chemical behavior and reactions.
• Existing models may lack generalizability or detailed physicochemical insights.

Purpose of the Study:

• To introduce a novel ML-based model for predicting solvation free energy.
• To develop a model based on pairwise atomistic interactions.
• To provide both accurate predictions and deeper physicochemical understanding of solvation.

Main Methods:

• A novel ML architecture employing two encoding functions to extract atomic feature vectors.
• Calculation of atomistic interactions via the inner product of feature vectors.
• Validation against 6239 experimental measurements of solvation free energy.

Main Results:

• The ML model achieved outstanding performance and transferability across diverse chemical structures.
• The solvent-non-specific nature of the model facilitates the enlargement of training data.
• Analysis of the interaction map revealed potential for group contributions to solvation energy.

Conclusions:

• The proposed ML model accurately predicts solvation free energy.
• The model's architecture offers valuable physicochemical insights beyond simple property prediction.
• This approach has potential for advancing the understanding of solvation phenomena.