Comparison of the Performance of Machine Learning Models in Representing High-Dimensional Free Energy Surfaces and...

Area of Science:

• Computational Chemistry
• Machine Learning
• Statistical Mechanics

Background:

• Enhanced sampling methods generate free energy surfaces using collective variables (CVs).
• High-dimensional free energy surfaces are challenging for traditional methods.
• Machine learning (ML) can aid in reconstructing these surfaces.

Purpose of the Study:

• Compare the performance of various ML models for learning high-dimensional free energy surfaces.
• Assess ML models' ability to generate accurate ensemble averages for observable properties.
• Evaluate ML model performance with varying amounts of sampled training data.

Main Methods:

• Applied driven adiabatic free-energy dynamics/temperature-accelerated molecular dynamics for sampling.
• Utilized regression ML methods: neural networks, kernel ridge regression, support vector machines, and weighted neighbor schemes.
• Trained ML models on sampled data from oligopeptide systems in gas and aqueous phases (2-10 dimensions).

Main Results:

• Demonstrated the capability of ML methods to learn and represent high-dimensional free energy surfaces.
• Found that ML model performance varies with the amount of training data.
• Confirmed that trained ML models can accurately generate ensemble averages for observable properties.

Conclusions:

• Machine learning regression techniques are effective for reconstructing high-dimensional free energy surfaces.
• The choice of ML model and data quantity significantly impacts the accuracy of surface representation and property prediction.
• This combined approach offers a powerful strategy for analyzing complex chemical and physical systems.