Reactive Flux Matching: Mechanism Discovery and Adaptive Sampling of Rare Events

Abstract: Path sampling methods generate ensembles of reactive trajectories connecting metastable states, but extracting mechanistic insight from these data remains nontrivial. We introduce Flux Matching, a framework that learns two complementary objects directly from reactive trajectory data: a current velocity $u(z)$, whose streamlines trace the dominant reaction pathways, and a scalar potential $h(z)$, obtained from a weighted Helmholtz-Hodge decomposition of the reactive current, that serves as a data-driven reaction coordinate. Both minimize quadratic functionals over the reactive path ensemble, analogous to the flow matching loss in generative modeling, and require no knowledge of the underlying dynamics or stationary distribution. Unlike committor-based methods, $u$ and $h$ remain well-defined under projection onto non-Markovian collective variables, and their level sets in turn provide adaptive interfaces for improved sampling with enhanced sampling methods. Flux Matching is validated through the generation of current velocity trajectories and rate constant calculations on molecular systems.

• The paper introduces a regression-based variational framework that directly learns reactive current and scalar potential from trajectory data.
• It demonstrates that the learned flow fields and adaptive collective variables enhance sampling efficiency and mechanistic interpretability in high-dimensional, non-Markovian systems.
• Empirical validations on molecular systems show improved transition rate estimation and reaction pathway elucidation compared to traditional committor-based methods.

Flux Matching: A Variational Framework for Mechanism Discovery and Adaptive Sampling of Rare Events

Problem Context and Motivation

Accurately characterizing rare transitions between metastable states in high-dimensional stochastic systems is central to progress in biomolecular simulation, chemical physics, and climate modeling. Standard molecular dynamics or Monte Carlo simulations are computationally inefficient for rare events due to the timescale gap between typical and reactive episodes. Path ensemble methods—transition path sampling (TPS), transition interface sampling (TIS), forward flux sampling (FFS), weighted ensemble (WE), and milestoning—address this by directly sampling reactive trajectories that shuttle between specified reactant and product states. However, extracting interpretable mechanistic insight and constructing rigorous reaction coordinates from these ensembles, especially under non-Markovian projections, remains a major challenge. Traditionally, this analysis centers on the committor function, but its reliance on the Markov property limits its applicability when dynamics are projected onto collective variables.

Theoretical Framework

The paper introduces Flux Matching, a regression-based variational framework that circumvents the limitations of committor-centric approaches by learning the reactive current and its associated potential directly from sampled trajectory data. The fundamental objects are:

• Current velocity $u(z)$: The local, empirical mean velocity of reactive trajectories at point $z$. Its streamlines delineate dominant reaction pathways that probabilistically mediate the system's transitions.
• Scalar potential $h(z)$: Obtained from a weighted Helmholtz-Hodge decomposition of the reactive current, $h(z)$ acts as a data-driven, globally optimal reaction coordinate. For dynamics obeying detailed balance, $h$ reduces to $\log \left[ q/(1-q)\right]$, with $q$ as the committor.

Both $u$ and $h$ are defined as unique minimizers of quadratic functionals over the reactive path ensemble, requiring only empirical path data—no knowledge of the underlying equations of motion, force field, or stationary distribution is necessary. The framework remains valid for non-Markovian projected dynamics, yielding exact marginal currents and potentials, a property unattainable for approximate committors in reduced representations.

Variational Principles and Loss Definitions

The variational characterization is central: $u$ minimizes a quadratic loss analogous to the flow matching loss used in generative modeling, but with the sampling measure defined via the reactive path ensemble, not a probability coupling. Similarly, $z$0 minimizes a Benamou-Brenier-type loss. These connections unify the theory with recent advances in generative modeling and optimal transport, but uniquely position it for nonequilibrium, rare event analysis. Critically, the construction ensures that the empirical estimators for $z$1 and $z$2 can be built from finite collections of reactive trajectories, independent of the timescale separation or underlying process Markovianity.

Mechanistic and Sampling Implications

Deterministic integration of the learned velocity field $z$3 produces flow lines representing the most probable transition routes, enabling mechanistic pathway analysis in both low- and high-dimensional representations. The monotonically increasing scalar potential $z$4 serves as an adaptive, learned collective variable (CV). Its level sets furnish interfaces or bin boundaries for a variety of enhanced sampling algorithms (TIS, FFS, WE, milestoning), enabling closed-loop schemes where improved mechanistic understanding informs improved sampling design, which then yields better data and better mechanistic inference in subsequent iterations.

Critically, since both $z$5 and $z$6 remain theoretically well-defined under arbitrary non-Markovian projections, the approach avoids uncontrolled approximations that plague committor-based methods in reduced-variable settings. This is validated rigorously in Section 2.6 and Appendix B of the paper.

Empirical Validation and Numerical Results

The framework is instantiated using neural network approximators for $z$7 and $z$8, trained via stochastic gradient descent on triplets or pairs of observed configurations, and evaluated on established molecular systems:

• Müller-Brown potential (2D): Flux Matching recovers the dominant transition channels under both overdamped and underdamped regimes, with learned flow fields that closely track the underlying stochastic path ensemble.
• Alanine Dipeptide (ADP, $z$9 and dihedral representations): In dihedral angle space, the framework achieves completion rates exceeding 98% and lower Torsional Wasserstein-2 ($h(z)$0-$h(z)$1) distances compared to path sampling reference ensembles. The monotonically varying $h(z)$2 captures the principal reaction coordinate, as supported by smooth flow lines and improved rate constant estimation in WE simulations.
• AIB9 peptide (backbone dihedrals): Accurate flow fields and reaction potentials are learned in a far higher-dimensional space, supporting many successful transitions and yielding interpretable mechanistic insight. Again, $h(z)$3 outperforms standard hand-crafted CVs in driving rare event sampling.

Notably, the results highlight that dihedral-based representations offer superior statistical performance and lower distributional discrepancy. Flux Matching retains exactness even under high-dimensional, non-Markovian projection, an important property for practical applications in large systems where dimensionality reduction is unavoidable.

Contradictory to committor-based pipelines, the method requires no assumption that the committor can be represented as a function of the reduced variables; all learning is performed via mechanically meaningful, regression-driven flows.

Algorithmic and Architectural Considerations

The models for $h(z)$4 and $h(z)$5 are realized as MLPs or transformers, depending on input features (Cartesian coordinates or internal dihedrals) and system complexity. The learning algorithm is stable due to careful handling of time-lag-induced bias-variance tradeoffs and regularization of boundary terms in the $h(z)$6 loss (via a cross-entropy surrogate for the boundary penalty). Full training and inference details, including flow line generation, post-processing, and validation, are specified; these contribute to the practical reliability and reproducibility of the approach.

Broader Implications and Future Directions

Flux Matching decouples the learning of reaction mechanisms and reaction coordinates from the need for Markovianity and explicit committor estimation, expanding interpretability and quantitative analysis in rare event simulation. Practically, this enables:

• Data-driven, system-specific construction of adaptive CVs for feedback into enhanced sampling workflows.
• Quantitative mechanism elucidation in non-Markovian, coarse-grained, or projected systems where traditional methods break down.
• Possibility of iterative refinement between sampling and mechanism discovery, enabling increasingly efficient exploration of high-dimensional transition landscapes.

The principal limitation for large-system deployment is the neural approximator’s scalability and the quality/diversity of the available trajectory ensemble. Addressing these (e.g., through symmetry-aware architectures or improved path sampling heuristics) is a clear direction.

Theoretically, the variational structure presents connections to flow-based generative modeling, optimal transport, and stochastic processes, potentially inspiring new crossover methods in statistical physics, chemical kinetics, and generative machine learning.

Conclusion

Flux Matching constitutes a rigorous, regression-driven framework for extracting mechanistic pathways and adaptive reaction coordinates from ensembles of rare, reactive trajectories, subsuming and generalizing traditional committor-based approaches. Its core strengths—exactness under projection, agnosticism to underlying dynamics, robust empirical performance, and direct connection to enhanced sampling methodologies—make it foundational for both future algorithmic development and application to complex rare event systems. Its potential for integration with deep generative models and adaptive sampling schemes positions it as a pivotal tool for the characterization of rare events in high-dimensional stochastic systems.

Reference: "Reactive Flux Matching: Mechanism Discovery and Adaptive Sampling of Rare Events" (2606.06295)