[Paper Review] Accelerated Markov Chain Monte Carlo Simulation via Neural Network-Driven Importance Sampling

Michael Kim, Wei Cai|arXiv (Cornell University)|Jan 31, 2026

Machine Learning in Materials Science0 citations

TL;DR

Proposes an importance sampling framework for accelerating MCMC of rare transitions by learning a neural-network–parametrized bias potential (and optional importance function) to bias transition probabilities, with a BRW variance reduction strategy and a generalized finite-region formulation.

ABSTRACT

Atomistic simulations provide valuable insights into the physical processes governing material behavior. However, their applicability is fundamentally constrained by the limited time scales accessible to brute-force simulations. This bottleneck often stems from complex energy landscapes where the systems stay trapped in metastable states for long periods of time. Yet, the long-term evolution is controlled by the transitions between the metastable states, which are rare events and difficult to observe. We present an importance sampling method designed to accelerate the time scale of Markov chain Monte Carlo (MCMC) simulations. By employing a bias potential, our approach enhances the sampling of rare transition events while preserving the relative probabilities of distinct transition pathways. The bias potential is represented by a neural network which enables the flexibility needed for high-dimensional systems. We propose a rigorous formulation to obtain the original transition rates between metastable states using transition paths obtained from the biased simulation. We further use a branching random walk (BRW) technique to enhance efficiency and to reduce variance. The proposed methodology is validated on 2-dimensional and 14-dimensional systems, demonstrating its accuracy and scalability.

Motivation & Objective

• Address time-scale limitations in atomistic simulations due to metastable-state trapping.
• Develop an importance sampling framework that preserves relative pathway probabilities while accelerating transitions.
• Introduce a neural network representation of bias potentials to handle high-dimensional systems.
• Provide a robust, continuum-friendly formulation with finite-region failures/successes to stabilize discretization.
• Incorporate variance reduction via branching random walks and a practical training scheme for bias potentials.

Proposed method

• Formulate importance sampling for rare transitions by modifying jump probabilities with an importance function I(i) while ensuring normalization.
• Introduce an extended domain with auxiliary states F and S to define failure and success, enabling a generalized, discretization-robust framework.
• Define the optimality condition as a right eigenvector relation Iopt(i) for the modified transition matrix, linking I(i) to the discrete committor function.
• Estimate transition rates rFS via mean first passage time from F to S, using pS(F) and average failure time tFF, with Iopt(F) guiding the rate calculation.
• Employ a neural-network–parameterized bias potential Eb(i;θ) to circumvent underflow and stabilize training; optimize using a loss Lθ derived from normalization constraints, with adaptive sampling.
• Use Branching Random Walk (BRW) to efficiently sample paths and control estimator variance, with stochastic rounding for path replication.

Experimental results

Research questions

• RQ1 How can importance sampling be formulated to accelerate rare state transitions while preserving the relative probabilities of competing pathways?
• RQ2 Can a neural network–driven bias potential (or its logarithmic form) stabilize and scale the importance sampling approach in high-dimensional systems?
• RQ3 How can transition rates between metastable states be accurately recovered from biased simulations via BRW and finite-region formulations?
• RQ4 Does the proposed framework remain effective when grid refinement increases dimensionality, and can coarse-grid training transfer to finer grids?
• RQ5 What practical training scheme (adaptive sampling, annealing) ensures convergence of the bias potential and accurate rate estimates?

Key findings

• The method reformulates rare-event sampling into an optimization problem for the optimal importance function, enabling biasing of transitions while preserving pathway probabilities.
• A generalized, discretization-robust formulation uses auxiliary states F and S to stabilize the boundary definitions and ensure convergence as grid spacing changes.
• A neural-network–parameterized bias potential Eb(i;θ) enables stable training and unbiased rate estimates, with a loss that enforces normalization-like conditions.
• Branching Random Walk provides variance reduction and scalable path sampling, improving efficiency in high dimensions.
• The framework yields transition-rate estimates rFS from mean first passage times with correction via pS(F) and Eb, maintaining proper dependence on temperature and grid spacing (ν0 scaling).
• The neural-network bias can be trained on coarse grids and applied to finer grids without retraining, preserving efficiency across resolutions.