PaperlayerView exact revision on arXiv

Source-linked AI summary

Iterative Neural Variational Solver of Committor Function in Under-damped Langevin Dynamics

Tiejun Li, Zhaoqi Liu, Xiang Zhou

arXiv:2610.20617v1physics.comp-phcs.LG

TL;DR

Computing committors for under-damped Langevin dynamics remains challenging because the backward Kolmogorov equation lacks the variational structure available in the over-damped case. The paper develops an iterative neural variational solver, with an optional sampling preconditioner, and reports accurate recovery of transition interfaces while examining friction-dependent dynamics.

  • Problem

    Computing committors for under-damped Langevin dynamics remains challenging because its backward Kolmogorov equation lacks a variational structure, limiting quantitative study of finite-friction transitions.

    Source: §1 · pp. 2–3

  • Method

    The method decomposes the under-damped generator into self-adjoint and skew-adjoint parts and solves iterative variational subproblems with neural networks, optionally using an auxiliary-potential preconditioner.

    Source: §1 · p. 4, §2.2 · p. 7

  • Results

    The solver yields accurate committor estimates and recovers the sharp q ≈ 0.5 isosurface, while the study reports that friction affects transition kinetics and simulation efficiency.

    Source: §1 · p. 4

  • Takeaways & Limitations

    The framework provides a practical numerical approach for probing rare transitions in under-damped Langevin dynamics across friction regimes.

    Source: §1 · pp. 3–4

    The reduced-coordinate model simplifies learning and identifies transition regions, but may suppress explicit velocity effects relevant in more weakly damped or strongly inertial regimes.

    Source: §3.3 · p. 19

Abstract

from arXiv · show

The committor function describes the probability that a stochastic dynamical system, while transitioning between two metastable states, will first reach one state rather than the other. It is a central object in transition path theory, providing the accurate reaction rate and the characterization of the transition mechanism through reactive probability flux and the transition state ensemble. While the numerical methods for this function are well studied for over-damped Langevin dynamics where the friction damping coefficient goes to infinity, its computation in the general under-damped case remains an open challenge due to the lack of a variational structure in the underlying backward Kolmogorov equation. This restricts the quantitative investigation of the impact of the finite damping coefficient on rare transitions in real systems. In this work, by decomposing the infinitesimal generator of the underdamped Langevin dynamics into a self-adjoint and a skew-adjoint part, we develop an iterative scheme in which each subproblem has a variational form and can be solved efficiently using neural networks. Fundamentally, this approach can be viewed as a generalized Concus-Golub-Widlund (CGW) splitting method and each subproblem has a variational form, which can be solved efficiently. This iterative scheme also has the flexibility of introducing a preconditioning operator associated with a new overdamped Langevin dynamics, which naturally leads to the design of enhanced adaptive sampling for the subproblems. We validate the efficiency and accuracy of our method on benchmark systems including a one-dimensional double-well potential, the Muller-Brown potential, and alanine dipeptide. The proposed iterative variational framework provides a practical and extensible numerical tool for probing rare transition events in Langevin dynamics across varying damping coefficients.

1. Introduction

The paper addresses the difficulty of computing committors for under-damped Langevin dynamics, where finite friction introduces velocity-dependent behavior and the backward Kolmogorov equation lacks the variational structure used in over-damped methods. It proposes an iterative neural variational solver and related tools for sampling transition regions, reducing coordinates, and studying friction-dependent transitions.

1.1. Related works

  • Prior neural Deep Ritz methods address over-damped committors, while other over-damped approaches include enhanced sampling, time-lag and semigroup formulations, and alternative machine-learning models.The cited over-damped variational functional's optimal value gives the transition rate between metastable sets.Source: §1 · p. 3, §1.1 · pp. 4–5

1. Introduction

  • For large friction γ, the over-damped committor is the leading-order approximation to the under-damped committor, while moderate or small friction makes velocity-dependent corrections significant.The committor is used to characterize transition probabilities, rates, dividing surfaces, and transition-state ensembles.Source: §1 · pp. 2–3
  • The generalized CGW scheme splits the under-damped problem into variational subproblems solvable with the Deep Ritz method, targeting accurate recovery of the sharp q ≈ 0.5 transition interface.The method also uses an auxiliary potential to concentrate training samples in the transition region.Source: §1 · p. 4
  • The auxiliary potential acts as a training preconditioner, concentrating data at steep committor gradients to accelerate convergence and improve recovery of the q ≈ 0.5 isosurface without importance-sampling reweighting.Source: §1 · p. 4
  • The work proposes a collective-variable reduction framework that systematically approximates the under-damped committor using collective variables and their associated phase-space velocities.Source: §1 · p. 4
  • The authors report that friction critically affects transition kinetics and simulation efficiency, and compare over-damped and under-damped committors and transition rates.Source: §1 · p. 4

2. Main Method

The method addresses the lack of a direct variational formulation for the under-damped committor by splitting its generator into self-adjoint and skew-adjoint components. It builds variational neural-network subproblems, adds flexible sampling and convergence strategies, and describes a reduced-variable representation for systems governed by a few collective variables.

2.1. Committor function of the under-damped Langevin dynamics

  • The under-damped generator’s decomposition does not make the committor equation a first-derivative variational problem, so the overdamped Deep Ritz approach does not extend directly.The decomposition separates the generator into component operators, including a self-adjoint, negative-semidefinite part.Source: §2.1 · pp. 5–6

2.2. Iterative Neural Variational Scheme

  • The generalized CGW iteration replaces a scalar shift with a spatial operator and makes each update a variational minimization that can be solved with the Deep Ritz method.The operator-valued shift is intended to address CGW shift sensitivity and improve convergence and robustness.Source: §2.2 · pp. 7–9

2.3. Enhanced sampling strategy

  • Choosing an auxiliary potential freely lets the method bias sampling toward transition regions, and its variational forms require samples rather than an explicit formula for that potential.The sampling formulation can combine committor-gradient and PDE-residual information to focus on informative regions.Source: §2.3 · pp. 9–10

2.4. Continuous Limit of the Iterative Scheme

  • The fixed-point iteration (20) is a time-discrete approximation to the continuous flow generated by e K−1L; under the spectral-gap condition, the semigroup decays exponentially at least at rate r, and iteration converges when ∆t respects the chosen integrator’s stability constraints.Larger ω1 and ω2 lead to more stable optimization but a lower convergence rate. A fully rigorous automated parameter-selection strategy remains an open problem.Source: §2.4 · pp. 10–11

2.5. Committor Function in Low-Dimensional Descriptors

  • When the dynamics are governed by a few collective variables, a reduced neural representation in CV positions and velocities can lower approximation dimension while retaining interpretable inertial corrections.If the chosen CVs respect physical symmetries, the reduced representation inherits them.Source: §2.5 · p. 12

3. Numerical Experiments

The experiments evaluate the iterative neural solver on double-well, Müller-Brown, and alanine-dipeptide systems across friction regimes. They assess committor accuracy, transition-rate estimates, robustness, and the effects of reducing the molecular system to collective coordinates.

3.1. Double well potential

  • In the double-well system, velocity dependence is pronounced at γ = 0.2 and diminishes toward the overdamped limit; at γ = 0.2, underdamped and overdamped rate estimates diverge substantially.PINNs lose accuracy in the weakly damped regime, while the iterative solver maintains fidelity; its generalized CGW configuration also converges consistently across tested ω1 values.Source: §3.1 · pp. 13–15, Figure 2 · p. 15
  • Standard finite-difference implementations of the generalized CGW method fail to converge in the tested scenarios, whereas the authors attribute localized convergence to neural-network optimization regularization.The reported spectral radii are 386.85, 82.58, 7.48, and 5.72.Source: §3.1 · p. 14

3.2. M¨uller-Brown potential

  • For the Müller-Brown potential, the iterative solver closely matches unbiased-simulation references at γ = 5 and 25, while PINN misses the transition region and its friction-dependent broadening.The solver also aligns the velocity-averaged qθ = 0.5 boundary with reference points at kBT = 5 and γ = 5.Source: §3.2 · pp. 16–17, Figure 4 · p. 16, Figure 5 · p. 17

3.3. Alanine dipeptide

  • The alanine-dipeptide model reduces a nominal 132-dimensional phase space to four coordinates and projected velocities, simplifying the committor approximation.The reduced model shows modest velocity-slice variation, and targeted shooting yields empirical committor distributions close to Gaussian and centered near 0.5.Source: §3.3 · pp. 18–19
  • The alanine-dipeptide reduction captures the reported reactive features, but may suppress explicit velocity effects relevant in more weakly damped or strongly inertial regimes.The learned velocity-averaged committor agrees qualitatively with overdamped results, and the expected near-linear ϕ–θ anticorrelation is recovered.Source: §3.3 · p. 19

4. Conclusion

The paper introduces a generalized CGW splitting framework for computing under-damped Langevin committors, while identifying unresolved challenges in its adaptive sampling and preconditioner selection.

4. Conclusion

  • The generalized CGW splitting framework computes committor functions for under-damped Langevin dynamics, with an auxiliary potential shifting sampling toward the reactive transition region.The authors report that the method resolves the sharp q≈0.5 interface and that standard PINNs can fail near it in complex energy landscapes.Source: §4 · p. 20
  • The authors identify two open refinements: rigorous adaptive construction of the auxiliary measure and automated, theoretically grounded selection of the preconditioner hyperparameters ω1 and ω2.The auxiliary-potential design currently relies on a heuristic focus on the transition zone; the spatial operator K empirically stabilizes optimization.Source: §4 · p. 20

CRediT authorship contribution statement

The contribution statement assigns conceptualization, methodology, validation, analysis, and writing responsibilities to three authors.

CRediT authorship contribution statement

Appendix A.1. Double well potential

The double-well appendix evaluates the neural variational solver across friction regimes and compares its learned committor with PINNs and a finite-difference reference.

Appendix A.1. Double well potential

  • Training uses neural networks for iterative subproblems, Adam optimization, and an L2 residual loss; the PINN baseline uses the same network architecture.The reference solution is computed by finite differences with Δx1 = 0.05 and Δv = 0.01√kBT.Source: Appendix A.1 · p. 21
  • The method produces a thinner, sharper transition interface than the PINN baseline, closely matching the finite-difference reference.The comparison covers γ = 0.2, 1, 5, and 25; the plotted planes show x1 and v1 with other variables fixed.Source: Appendix A.1 · p. 21, Figure A.9 · p. 22, Figure A.10 · p. 22
  • As γ increases, the q≈0.5 isosurface becomes less tilted, indicating reduced velocity influence on the committor, consistent with asymptotic analysis.Source: Appendix A.1 · p. 21, Figure A.9 · p. 22, Figure A.10 · p. 22

Appendix A.2. M¨uller-Brown potential

The Müller-Brown appendix compares learned committors with PINNs and brute-force simulations across friction values, and reports how friction changes the transition interface.

Appendix A.2. Müller-Brown potential

  • The appendix trains neural-network subproblems on the Müller-Brown spatial domain using composite sample weights and excludes points with potential energy above zero.The reported baseline PINN training runs for 100,000, 100,000, and 200,000 epochs at successively lower learning rates.Source: Appendix A.2 · pp. 22–23
  • As γ increases from 1 to 25, the q≈0.5 transition interface widens at fixed velocity, consistent with faster velocity relaxation and weaker initial-velocity influence.Source: Figure A.11 · p. 23, Appendix A.2 · p. 24
  • Figure A.11 compares the proposed method and PINNs with brute-force committor analysis at kBT = 5 and γ = 1, 5, and 25.The brute-force reference uses 100 independent BAOAB simulations with a time step of 2 × 10^-5.Source: Appendix A.2 · pp. 23–24, Figure A.11 · p. 23

Appendix A.3. Alanine dipeptide

The alanine dipeptide appendix describes a two-stage sampling workflow for learning the under-damped committor and compares its biased-simulation data with metadynamics data. The biased-simulation dataset yields a more sharply resolved transition interface.

Appendix A.3. Alanine dipeptide

  • A two-stage workflow pre-trains an overdamped committor q0(x), then uses it to bias a 12 ns simulation and retains transition-pathway configurations.The bias is V = −λ log(∥∇xq0∥2 + ϵ), with λ = 3.5 reported as the most efficient sampling choice.Source: Appendix A.3 · p. 24
  • The under-damped committor is represented by a [4, 8, 256, 256, 1] neural network trained with Adam; Table A.5 tabulates subproblem tuning parameters and epoch counts.The tuning parameters include the spatial operator K and learning rate.Source: Appendix A.3 · p. 24, Table A.5 · p. 25
  • Figure A.12 plots the committor projected onto ϕ-θ space, with separate planes showing fixed velocities.Source: Figure A.12 · p. 25
  • Biased-simulation data learn a smooth, sharply defined transition interface, whereas metadynamics data fail to resolve the sharp boundary adequately.The comparison holds hyperparameters fixed and uses ϕ and ψ as metadynamics collective variables.Source: Appendix A.3 · p. 25
Loading 2610.20617v1…