Adaptive Tensor Train Metadynamics for High-Dimensional Free Energy Exploration

Area of Science:

• Computational Chemistry
• Molecular Dynamics Simulations
• Free Energy Calculations

Background:

• Efficient exploration of free energy landscapes is crucial for molecular dynamics (MD) simulations.
• Standard methods like metadynamics become computationally expensive with increasing collective variables (CVs).

Purpose of the Study:

• Introduce TT-Metadynamics, a novel method for enhanced sampling in MD simulations.
• Address the computational scaling limitations of traditional metadynamics for high-dimensional systems.

Main Methods:

• Developed TT-Metadynamics, compressing metadynamics bias potentials into a tensor train (TT) format.
• Implemented a sketching algorithm for linear scaling in constructing the TT representation with respect to CVs.

Main Results:

• TT-Metadynamics demonstrates efficient memory usage and prevents computational cost from increasing with simulation time.
• Achieved accuracy comparable to or exceeding standard metadynamics in systems with up to 14 CVs.
• Showed particular effectiveness in systems with high energy barriers.

Conclusions:

• TT-Metadynamics offers a scalable and effective approach for computing free energies across multiple collective variables.
• Represents a significant advancement for molecular dynamics simulations of complex systems.