Neural Non-Equilibrium Hamiltonian Monte Carlo for Corrected Boltzmann Sampling
Introduction of Neural Non-Equilibrium Hamiltonian Monte Carlo (NHMC), a "train-then-correct" learned sampler for unnormalized Boltzmann densities. Utilization of stochastic Hamiltonian-style paths where statistical correction is applied via non-equilibrium work derived from forward-reverse path probability ratios. Application of multiple correction methods including path-SNIS, path-IMH, and shared-bridge round-trip Metropolis kernels to preserve the Boltzmann target. Validation on double-well p
Analysis
TL;DR
- Introduction of Neural Non-Equilibrium Hamiltonian Monte Carlo (NHMC), a "train-then-correct" learned sampler for unnormalized Boltzmann densities.
- Utilization of stochastic Hamiltonian-style paths where statistical correction is applied via non-equilibrium work derived from forward-reverse path probability ratios.
- Application of multiple correction methods including path-SNIS, path-IMH, and shared-bridge round-trip Metropolis kernels to preserve the Boltzmann target.
- Validation on double-well potentials and lattice $\phi^4$ models, demonstrating success with sufficient path overlap but failure modes under poor overlap.
- Feasibility study reported for molecular internal coordinates using MD priors and learned-force path proposals.
Why It Matters
This work bridges machine learning and statistical physics by offering a scalable method for sampling complex energy landscapes where traditional MCMC struggles due to high barriers or high dimensionality. For researchers in computational chemistry and physics, it provides a framework to leverage neural networks for global exploration while maintaining rigorous statistical correctness through non-equilibrium thermodynamic identities.
Technical Details
- Core Architecture: NHMC learns stochastic paths from a tractable base distribution to the target Boltzmann distribution using a neural network to parameterize the dynamics.
- Correction Mechanism: After training, parameters are fixed. The method records non-equilibrium work, defined as the log-ratio of forward to reverse path probabilities, to compute importance weights or Metropolis acceptance ratios.
- Sampling Variants: The framework supports path-space Self-Normalized Importance Sampling (SNIS), path-space Independent Metropolis-Hastings (IMH), and a shared-bridge round-trip Metropolis kernel acting directly on configurations.
- Training Objective: Minimizing the mean non-equilibrium work during training reduces path-space KL divergence and bounds endpoint mismatch.
- Empirical Results: Tested on double-well and finite-volume lattice $\phi^4$ targets; performance degrades significantly when path overlap is poor, leading to weight degeneracy and high autocorrelation.
Industry Insight
- Hybrid Modeling: Combining physical priors (like MD simulations) with learned neural dynamics can enhance sampling efficiency in high-dimensional spaces common in molecular modeling.
- Limitations of Learned Samplers: Practitioners must rigorously assess path overlap; learned proposals are not universally robust and require careful validation against standard MCMC baselines to avoid biased estimates.
- Theoretical Rigor: The "train-then-correct" paradigm ensures that despite using approximate neural dynamics, final estimates remain unbiased, making this approach viable for scientific computing applications requiring precise thermodynamic quantities.
Disclaimer: The above content is generated by AI and is for reference only.