Neural mode jump Monte Carlo

Area of Science:

• Computational Physics
• Statistical Mechanics
• Machine Learning

Background:

• Markov chain Monte Carlo (MCMC) methods are essential for simulating complex systems.
• Slow convergence due to large energy barriers in metastable states is a significant challenge.
• Existing methods struggle to efficiently sample configurations across multiple distinct energy wells.

Purpose of the Study:

• To develop a novel computational method for accelerating MCMC convergence in systems with numerous metastable states.
• To directly connect disparate metastable regions within the configuration space.
• To improve the efficiency of sampling equilibrium configurations in complex systems.

Main Methods:

• Utilizing generative neural networks to propose new configurations for the Markov chain.
• Optimizing the acceptance probability for large transitions between different metastable states.
• Developing a comprehensive theoretical framework and a training scheme for the neural network.

Main Results:

• Demonstrated significant increases in convergence speed on example systems.
• Successfully connected previously isolated metastable regions.
• The proposed method effectively enhances the exploration of the configuration space.

Conclusions:

• Generative neural networks offer a powerful approach to overcome convergence limitations in MCMC simulations.
• This novel method provides a robust solution for sampling complex systems with multiple metastable states.
• The approach has broad applicability in various scientific domains requiring advanced simulation techniques.