Accelerating Rare Event Sampling in Molecular Dynamics

320 concepts242 papers1 proposal

At the microscopic scale, nature moves to the rhythm of atomic vibrations, which occur on the order of femtoseconds ($10^{-15}$ seconds). Yet, the macroscopic phenomena we care about—such as protein folding, chemical catalysis, and crystal nucleation—unfold over milliseconds, hours, or even years. This vast separation of timescales creates the rare event problem in molecular dynamics (MD).

As discussed by Magnus Petersen et al. in PINN-MEP: Continuous Neural Representations for Minimum Energy Path Discovery in Molecular Systems, standard brute-force simulations spend virtually all their computational budget trapped within stable, low-energy local minima (metastable states). Escaping these wells requires crossing high free-energy barriers. Spontaneous crossings are driven by rare, exceptionally large thermal fluctuations, making it computationally intractable to simply wait for them to happen.

To bridge this timescale gap, researchers have developed enhanced sampling methods. The field is currently transitioning from classical techniques that rely on human intuition to modern machine learning frameworks that discover transition pathways automatically.

The Collective Variable Bottleneck

The classical solution to the rare event problem is to inject external energy into the simulation to help the system escape its local minimum. For instance, Metadynamics deposits history-dependent "bias potentials" (like dropping sand into a valley) to discourage the system from revisiting already-explored configurations, as reviewed in Enhancing Important Fluctuations: Rare Events and Metadynamics from a Conceptual Viewpoint.

However, adding a bias potential directly to thousands of atomic coordinates is highly inefficient. Instead, simulations must project the high-dimensional configuration space down to a few critical degrees of freedom called Collective Variables (CVs) or reaction coordinates (e.g., a specific dihedral angle or a distance between two atoms).

This introduces a severe bottleneck: if the chosen CVs fail to capture a slow, hidden physical motion, the biasing force will push the system in the wrong direction, leading to hysteresis and inaccurate kinetics. As Maksim Kulichenko et al. point out in Uncertainty-driven dynamics for active learning of interatomic potentials, finding these informative regions traditionally requires a "chicken-and-egg" loop where you need prior intuition or successful transition trajectories just to design the bias that finds them.

To alleviate this, Dan Mendels et al. developed Harmonic Linear Discriminant Analysis (HLDA) in Path collective variables without paths. HLDA constructs effective CVs purely by analyzing the local, equilibrium fluctuations within the initial and final metastable states, eliminating the need for a pre-existing transition path. Similarly, Z. Faidon Brotzakis demonstrated in Enhanced Sampling of Protein Conformational Transitions via Dynamically Optimized Collective Variables that CVs can be variationally optimized on the fly, allowing the simulation to adaptively discover the lowest-energy escape routes.

Bypassing CVs with Stochastic Optimal Control

Rather than trying to find a low-dimensional projection of the energy landscape, a powerful modern alternative treats rare event sampling as a Stochastic Optimal Control problem over the entire trajectory space.

In standard Langevin dynamics, a molecule's motion is governed by the potential energy gradient plus random thermal noise. In the optimal control framework, we add a time-dependent "drift" force $\mathbf{u}(x, t)$ to the equation of motion. The goal is to find a control force that guarantees the system travels from state A to state B within a specified time, while minimizing the "extra work" done by the controller.

Standard Dynamics:    dX_t = -∇V(X_t)dt + √(2D)dW_t  (Trapped in State A)
Controlled Dynamics:  dX_t = [-∇V(X_t) + u(X_t,t)]dt + √(2D)dW_t  (Guided to State B)

Lars Holdijk et al. implemented this concept in PIPS: Path Integral Stochastic Optimal Control for Path Sampling in Molecular Dynamics. By parameterizing the control force $\mathbf{u}$ as a neural network, they optimize it using a path-integral loss function.

Why does this work? The mathematical formulation ensures that the optimal control force naturally guides the system along the dynamical instanton—the most statistically probable transition trajectory through the high-dimensional landscape, as characterized by Valentina Ros et al. in Dynamical instantons and activated processes in mean-field glass models. This completely bypasses the need to define CVs, as the neural network learns the exact forces required to clear the barrier efficiently.

A closely related approach leverages generative modeling. In Collective Variable Free Transition Path Sampling with Generative Flow Network, Kiyoung Seong et al. used Generative Flow Networks (GFlowNets) to sample trajectories. By training the model to match the probability of a path to its underlying physical action, they generate a diverse ensemble of transition pathways without getting stuck in a single mechanism.

Learning the Committor Function

The gold standard for understanding a rare event is the committor function, $q(x)$. For any given molecular configuration $x$, the committor is the probability that a trajectory initiated from $x$ will reach the product state B before returning to the reactant state A.

  • Configurations deep inside state A have $q(x) = 0$.
  • Configurations deep inside state B have $q(x) = 1$.
  • The transition state or "mountain pass" is defined by the isosurface $q(x) = 0.5$.

Knowing $q(x)$ provides the perfect reaction coordinate because it measures the true progress of the reaction. However, calculating $q(x)$ requires solving a high-dimensional partial differential equation (the backward Kolmogorov equation), which suffers from the curse of dimensionality.

To solve this, Haoya Li et al. introduced a semigroup method in A semigroup method for high dimensional committor functions based on neural network. Instead of dealing with unstable second-order derivatives, they reformulate the problem into an integral equation that can be optimized using neural networks.

Complementing this, Andrew Mitchell developed a framework in Committor guided estimates of molecular transition rates that pairs a specialized logarithmic loss function with an adaptive "swarm" sampling strategy. The simulation spawns local trajectories where the committor uncertainty is highest, rapidly mapping out the transition state ridge and calculating exact macroscopic transition rates.

Continuous Neural Path Representations

Instead of running sequential, step-by-step simulations through time, another paradigm optimizes the entire transition path simultaneously as a geometric object.

In PINN-MEP: Continuous Neural Representations for Minimum Energy Path Discovery in Molecular Systems, Magnus Petersen et al. represent a transition path as a continuous neural function $\gamma(\tau)$, which maps a normalized scalar parameter $\tau \in [0, 1]$ directly to a full high-dimensional molecular configuration.

  τ = 0 (Reactant)  ───────►  Neural Network γ(τ)  ───────►  τ = 1 (Product)
                                      │
                           Optimized via Physics Loss
                                      ▼
                             Minimum Energy Path

By utilizing Physics-Informed Neural Networks (PINNs), the authors optimize the network's weights so that the path minimizes the perpendicular forces acting along it. Because the path is continuous and fully differentiable, the network can leverage exact analytical gradients from the molecular potential energy surface. The neural path behaves like an elastic band that automatically snaps onto the Minimum Energy Path (MEP), providing a clear view of the transition mechanism without the immense computational cost of traditional brute-force sampling.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 246 papers
  2. Version 1 · 2026-10-04 · 255 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_rare_event_sampling_e62e3025,
  title = {rare event sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/rare-event-sampling/rdr_e62e302520bb4e0e8a66442198ad7f48},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_rare_event_sampling_e62e3025_v2,
  title = {rare event sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/rare-event-sampling/rdr_e62e302520bb4e0e8a66442198ad7f48/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake