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

The iterative neural variational solver develops a novel method to compute the committor function for under-damped Langevin dynamics, which is crucial for accurately characterizing rare transitions in real systems across varying friction coefficients. <ref:2610.20617#pg2>

How it works

The core of the method involves decomposing the infinitesimal generator of the underdamped Langevin dynamics into a self-adjoint and a skew-adjoint part, which allows for an iterative scheme where each subproblem has a variational form solvable by neural networks <ref:2610.20617#pg4>. This approach is fundamentally viewed as a generalized Concus-Golub-Widlund (CGW) splitting method, and each subproblem possesses a variational form that can be solved efficiently <ref:2610.20617#pg5>. The iterative scheme also offers the flexibility to introduce a preconditioning operator associated with an overdamped Langevin dynamics, which naturally leads to enhanced adaptive sampling for the subproblems <ref:2610.20617#pg5>.

The iterative scheme is formulated as:

((ω1I − γS)qk+1 = (ω1I + A)qk)

where qk+1(x, v) = 0, ∀x ∈ ∂A, v ∈ Rn and qk+1(x, v) = 1, ∀x ∈ ∂B, v ∈ Rn <ref:2610.20617#pg9>. This update qk+1 could be computed in a variational way since S is negative semi-definite with respect to the weighted inner product π <ref:2610.20617#pg9>. The primary bottleneck of the CGW method is the optimal selection of the shift parameter ω1 <ref:2610.20617#pg10>. To overcome this, a generalized CGW framework replaces the standard scalar shift with an operator-valued formulation, considering K = ω1I − ω2Kx <ref:2610.20617#pg10>. This combined operator K − γS is coercive and positive definite under the weighted product measure eµ ⊗ π <ref:2610.20617#pg10>.

Key Contributions and Enhancements

The work introduces several novel elements to address the challenges of computing the committor function in under-damped systems <ref:2610.20617#pg5>.

((Ue)

is introduced as an auxiliary potential to drive sampling directly toward transition regions, which is distinct from importance sampling because it acts as a preconditioning operator during training <ref:2610.20617#pg10>. This process concentrates training data at the steepest committor gradients, substantially accelerating convergence and dramatically improving the recovery of the q ≈ 0.5 isosurface <ref:2610.20617#pg10>.

((q(x, v) ≈ Q(s(x), s˙(x)) = Q(s(x), J T s (x)v))

is proposed as a novel reduction framework, establishing a theoretical foundation proving that the true committor function can be systematically approximated using only collective variables (CVs) and their associated phase-space velocities <ref:2610.20617#pg10>. This reduced parametrization captures the most dynamically relevant dependence on configuration and instantaneous coarse-grained velocity <ref:2610.20617#pg10>.

Validation and Results

The method was validated on benchmark systems, including a one-dimensional double-well potential, the Muller–Brown potential, and alanine dipeptide <ref:2610.20617#pg5>.

((Fig. 1)

shows that when γ is small, velocity significantly influences the committor; as γ → ∞, the committor approaches a function of x only, recovering the over-damped limit <ref:2610.20617#pg15>.

((Fig. 4)

compares the velocity-averaged committors, showing that both the FDM reference and our iterative neural solver pass precisely through the red reference points obtained from unbiased simulations, demonstrating that our method accurately resolves the critical transition boundary <ref:2610.20617#pg16>.

((Fig. 5)

compares results across different friction regimes, showing that at γ = 5, the reactive interface is relatively sharp, while increasing friction to γ = 25 produces a noticeably broader transition region <ref:2610.20617#pg18>.

((Fig. 7)

for alanine dipeptide in vacuum, the velocity-averaged committor agrees with established over-damped results, and direct shooting tests confirm the committor’s predictive accuracy <ref:2610.20617#pg19>.

Conclusion

In summary, the iterative neural variational solver successfully computes the committor function for under-damped Langevin dynamics by employing a generalized CGW splitting framework and a targeted auxiliary potential Ue <ref:2610.20617#pg5>. This method reliably resolves sharp transition interfaces and provides accurate estimates of reaction rates across various friction regimes <ref:2610.20617#pg5>. The framework offers a practical and extensible numerical tool for probing rare transition events in Langevin dynamics across varying damping coefficients <ref:2610.20617#pg5>. The proposed method is the first practical numerical solver for under-damped Langevin dynamics <ref:2610.20617#pg5>. The true committor function can be systematically approximated using only collective variables (CVs) and their associated phase-space velocities <ref:2610.20617#pg10>. The friction coefficient, γ, critically dictates transition kinetics and simulation efficiency <ref:2610.20617#pg5>.

How it works

The continuous limit of the iterative scheme is interpreted as a temporal discretization of a continuous evolution equation for an auxiliary field ˜q(x, v, t) <ref:2610.20617#pg10>. This leads to the discrete iteration (20), which provides a time-discrete approximation to the continuous flow generated by Ke−1L <ref:2610.20617#pg10>. The iterative scheme admits a natural interpretation as a temporal discretization of a continuous evolution equation for an auxiliary field ˜q(x, v, t) <ref:2610.20617#pg10>.

How it works

The final results show that the learned committor is nearly identical to the reference true solution obtained from the unbiased simulations <ref:2610.20617#pg5>. This high fidelity confirms that our approach successfully resolves the correct transition interface across the energy landscape <ref:2610.20617#pg5>. The iterative neural solver bypasses spectral ambiguities that cause severe topological failures near the transition interface <ref:2610.20617#pg5>.

How it works

The method's robustness in parameter selection is demonstrated by its consistent convergence behavior regardless of the ω1 value, insulating the training process from hyperparameter sensitivity <ref:2610.20617#pg10>. The empirical FDM implementations of our method fail to converge because their spectral radii are significantly greater than one, indicating that our generalized CGW formulation is only effective within the continuous neural network framework <ref:2610.20617#pg10>.

How it works

The two primary challenges remaining for future investigation are developing a fully adaptive, mathematically rigorous strategy for constructing the auxiliary measure Ue and establishing an automated, theoretically grounded protocol for selecting the optimal hyperparameters ω1 and ω2 <ref:2610.20617#pg5>. The spatial operator K effectively serves as a preconditioner for the system’s infinitesimal generator, L <ref:2610.20617#pg10>. The choice of sampling strategy significantly impacts the quality of the learned committor function <ref:2610.20617#pg10>.

How it works

The final results indicate that the reduced-coordinate ansatz captures the essential reactive features of alanine dipeptide, with velocity-averaged committor agreeing with established over-damped results <ref:2610.20617#pg5>. The modest velocity dependence observed after reduction highlights a practical trade-off between model reduction and explicit velocity effects <ref:2610.20617#pg10>. The learned committor successfully identifies the transition region and produces calibrated probability estimates through targeted shooting experiments <ref:2610.20617#pg5>.

The iterative neural variational solver introduces a generalized CGW splitting framework to compute the committor function for under-damped Langevin dynamics, which is crucial for accurately characterizing rare transitions in real systems across varying friction coefficients. <ref:2610.

1. This iterative neural solver can compute the committor function for under-damped Langevin dynamics across varying damping coefficients, providing a practical and extensible numerical tool.

2. The system can perform enhanced sampling by introducing an auxiliary potential Ue that acts as a preconditioning operator during training: concentrating training data at the steepest committor gradients substantially accelerates convergence.

3. It can resolve sharp transition interfaces more accurately than standard PINNs, as the paper states, Our method proves highly robust in challenging, high-dimensional regimes, successfully resolving sharp transition interfaces and overcoming PINN failures by avoiding eigenfunctions with an eigenvalue close to zero.

4. The AI can perform model reduction by parameterizing the committor using a reduced-variable ansatz of the form Q(s(x), s˙(x)) = Q(s(x), J T s(x)v), which dramatically reduces the function-approximation dimension for high-dimensional problems like alanine dipeptide.

5. The system can accurately estimate transition rates, as the learned committor allows comparison with asymptotic predictions, showing that the two estimates agree closely for large friction coefficients and validating the relation κAB = γ−1κ0 + O(γ−2).