Scientific Background

The code in this repository is best understood as a practical implementation of the method family developed across the AIMMD papers.

The central object is the committor \(q(x)\), the probability that a trajectory started from configuration \(x\) reaches the product state \(B\) before the reactant state \(A\):

\[q(x) = \Pr\!\left[\, \tau_B < \tau_A \,\mid\, x(0) = x \,\right],\]

where \(\tau_A\) and \(\tau_B\) are the first-passage times to each state. The committor is the ideal reaction coordinate: the dividing surface \(q(x) = 1/2\) is the transition state. AIMMD learns a model of \(q\) from sampled trajectories and uses it to bias shooting-point selection toward the transition-state region, where new trajectories are most informative.

AIMMD as a Learning-Guided Path-Sampling Loop

H. Jung, R. Covino, A. Arjun, C. Leitold, C. Dellago, P. G. Bolhuis, and G. Hummer, Machine-Guided Path Sampling to Discover Mechanisms of Molecular Self-Organization, Nat. Comput. Sci. 2023, 3, 334-345. DOI: https://doi.org/10.1038/s43588-023-00428-z

This paper introduces the central AIMMD idea: rare-event path sampling and a learned mechanistic model form a closed loop. Newly generated trajectories are used to build and validate a committor-like description, and that description then guides where the next trajectories should start. This forms an autonomous mechanism-discovery workflow for rare molecular events.

In the repository, that logic appears most directly in:

  • aimmd.worker._shoot for guided two-way shooting,

  • aimmd.worker._train for iterative model updates,

  • aimmd.launcher for coordinating concurrent workers,

  • and aimmd.network.fit for learning the committor surrogate.

Reweighting Paths into Free Energies, Rates, and Mechanisms

G. Lazzeri, H. Jung, P. G. Bolhuis, and R. Covino, Molecular Free Energies, Rates, and Mechanisms from Data-Efficient Path Sampling Simulations, J. Chem. Theory Comput. 2023, 19, 9060-9076. DOI: https://doi.org/10.1021/acs.jctc.3c00821

This paper supplies the statistical interpretation layer: once AIMMD has generated a mixture of free trajectories, excursions, and transition paths, the ensemble can be reweighted to recover equilibrium observables such as free energy profiles and rate-related quantities.

That is the scientific rationale behind:

although the details of reweighting have been updated in the 2025 paper.

Rejection-Free Path Sampling and reformulated reweighting

G. Lazzeri, P. G. Bolhuis, and R. Covino, Optimal Rejection-Free Path Sampling, arXiv, 2025. Link: https://arxiv.org/html/2503.21037v1

This shifts the focus from sampling a restricted path ensemble to sampling the distribution of shooting points directly. The key idea is that if shooting points are selected with a bias proportional to the inverse free energy along a good reaction coordinate, the reactive region can be sampled much more uniformly, and the resulting algorithm can be rejection-free.

In AIMMD, that perspective is implemented as:

  • the rfps chain type,

  • the adaptive shooting-point selection pool,

  • the interplay between bins, densities, and network values,

  • a reformulated reweighting scheme (via the drop-method).

How the Repository Encodes the Theory

The repository decomposes the method into clear implementation units:

aimmd.Params

Encodes the scientific parameters of a run: states, descriptors, values, network, engine details, and sampling controls.

aimmd.Worker.shoot()

Implements the enhanced sampling step by selecting shooting points and generating new two-way trajectories.

aimmd.Worker.train()

Updates the learned committor surrogate, recomputes values, and refreshes the adaptive discretization of value space.

aimmd.PathEnsemble.reweight()

Converts the biased sampled data back into equilibrium-weighted estimates.

aimmd.analysis.utils

Supplies support routines for binning, uncertainty estimation, and related post-processing steps.