Chasing Collective Variables Using Autoencoders and Biased Trajectories.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Free energy biasing methods accelerate molecular simulations by modifying sampling.
• These methods often require prior knowledge of slow degrees of freedom, known as collective variables (CVs).
• Machine learning (ML) and dimensionality reduction algorithms can identify these essential CVs.

Purpose of the Study:

• To introduce a novel iterative method for learning collective variables (CVs) in molecular simulations.
• To enhance free energy biasing techniques by integrating autoencoder-based CV discovery.
• To ensure convergence and optimize learning through a reweighting scheme.

Main Methods:

• Developed Free Energy Biasing and Iterative Learning with AutoEncoders (FEBILAE).
• Employed autoencoders for iterative learning of CVs.
• Integrated a reweighting scheme for consistent loss optimization and CV convergence.
• Utilized the extended adaptive biasing force method for free energy adaptive biasing.

Main Results:

• Demonstrated the FEBILAE algorithm's effectiveness on alanine dipeptide and solvated chignolin systems.
• Showcased the method's ability to iteratively learn and refine CVs.
• Validated the convergence properties of the proposed reweighting scheme.

Conclusions:

• FEBILAE offers an efficient, data-driven approach to identify CVs for molecular simulations.
• The iterative learning and reweighting scheme ensure robust CV discovery and simulation acceleration.
• This method advances the application of ML in computational chemistry for studying molecular conformational changes.