Neural Transport Nested Sampling

TL;DR - Neural Transport Nested Sampling combines nested sampling with flow-guided, Metropolis–Hastings-corrected Langevin dynamics to sample molecular Boltzmann distributions and estimate partition functions. It improves sampling accuracy and cost on challenging particle systems while recovering temperature-dependent phase structure from one run.

• Requires only evaluations of the target energy function.
• Scales to Lennard–Jones clusters containing 55 interacting particles.
• Reduces interatomic-distance and energy Wasserstein errors by over an order of magnitude versus the strongest neural baselines, at lower wall-clock cost.
• Produces calibrated, temperature-resolved partition-function estimates at a scale the authors report as unprecedented for neural samplers.