Kinetic reconstruction of free energies as a function of multiple order parameters

Yagyik Goswami, Srikanth Sastry

TL;DR

This paper introduces a novel method to reconstruct multi-dimensional free energy landscapes from unconstrained, steady-state simulation trajectories. By relating macroscopic fluxes and steady-state sampling probabilities to equilibrium states, the authors successfully map complex phase transitions—like the crystallization of supercooled silicon—without the need for the artificial biasing forces required by traditional methods.

What Problem Does This Paper Address?

In thermodynamics, a "free energy surface" (FES) is essentially a map that tells us which states a system prefers and how difficult it is to move between them (the barriers). For "rare events," such as a liquid turning into a crystal, the barriers are often so high that standard simulations get "stuck" in one state for a long time.

To solve this, researchers usually use enhanced sampling (like Umbrella Sampling), which applies artificial biases to push the system over barriers. However, these methods can be computationally expensive and sensitive to how the bias is applied. A simpler alternative is kinetic reconstruction, which uses the "natural" flow of a system to infer the landscape. Until now, kinetic reconstruction was largely limited to a single dimension (one order parameter). This paper tackles the challenge of extending this approach to multiple dimensions, allowing researchers to study complex phenomena where one coordinate isn't enough to describe the physics.

Methods

The authors propose a framework based on Steady-State Fluxes. Instead of waiting for a system to reach equilibrium, they set up a steady-state flow: particles are "injected" at a starting point (the reactant) and "absorbed" at a distant boundary (the product).

Key components include:

  • The Flux-Probability Relation: They derive a mathematical relationship where the equilibrium probability $P_{eq}$ is proportional to the steady-state probability $P_{st}$ multiplied by a correction factor: the ratio of forward flux to backward flux between two points in the order parameter space.
  • Mean First Passage Time (MFPT): They use the time it takes for a system to first reach a boundary to help define the 1D backbone of the free energy profile.
  • Sectioned Reconstruction: To improve accuracy across large landscapes, they divide the space into overlapping sections, calculating local fluxes and "stitching" the free energy pieces together.
  • Unconstrained MD: Unlike umbrella sampling, the Molecular Dynamics (MD) runs are "unconstrained," meaning the system evolves according to its natural physics without added biasing potentials.

Datasets

The authors validated their method using two primary systems:

  1. Synthetic Toy Model: Independent overdamped Brownian random walkers moving on a 2D potential energy surface defined by four Gaussian functions. This allowed for a direct comparison between the "true" known potential and the reconstructed one.
  2. Supercooled Liquid Silicon: Simulations using the Stillinger-Weber potential. This is a classic "real-world" test case for polyamorphism (the existence of two different liquid phases). The order parameters used were the size of the largest crystalline cluster ($n_{max}$) and the system density ($\rho$).

Key Insights and Ideas

  • Flux as a Hidden Variable: The "aha moment" is that even when a system is far from equilibrium (due to an absorbing sink), the underlying equilibrium landscape is still encoded in the ratio of how often the system moves forward versus backward.
  • Correcting the Steady State: The paper demonstrates that steady-state sampling ($P_{st}$) deviates significantly from Boltzmann sampling when barriers are low. Their flux-based correction factor successfully "undoes" this deviation to recover the true free energy.
  • Capturing Polyamorphism: The method successfully distinguished between High-Density Liquid (HDL) and Low-Density Liquid (LDL) states in silicon, showing that the LDL state becomes more stable as the temperature drops—a finding that matches much more complex biased sampling studies.

Key Results

  • High Accuracy: For the toy models, the error in the reconstructed free energy was less than 0.1 $k_B T$ in the metastable regions, representing an error of roughly 1%.
  • Silicon Consistency: The method reconstructed the density-dependent free energy of silicon at 975K, 985K, and 995K. The results were quantitatively consistent with previous Umbrella Sampling data, capturing the shift in the metastable minimum from $\sim$2.45 gcc⁻¹ to $\sim$2.35 gcc⁻¹ as temperature decreased.
  • Efficiency: The reconstruction was achieved using a relatively small ensemble of trajectories (e.g., 400–600 independent runs), which is computationally competitive with traditional biased methods.

Limitations and Future Work

  • The "Beyond the Barrier" Problem: Accuracy drops significantly once the system crosses the main barrier toward the absorbing boundary. This is because "backward flux" (moving from the product back toward the reactant) becomes extremely rare, leading to poor statistical sampling.
  • Boundary Sensitivity: The assumption that transition rates are unaffected by the absorbing boundary holds only if the boundary is placed sufficiently far from the regions of interest.
  • Future Directions: The authors suggest that this method could be extended to driven systems (systems with external work being done) or non-equilibrium steady states where traditional equilibrium thermodynamics don't apply, potentially opening a new window into active matter and driven chemical reactions.

Summary updated

Abstract

A vast array of phenomena, ranging from chemical reactions to phase transformations, are analyzed in terms of a free energy surface defined with respect to a single or multiple order parameters. Enhanced sampling methods are typically used, especially in the presence of large free energy barriers, to estimate free energies using biasing protocols and sampling of transition paths. Kinetic reconstructions of free energy barriers of intermediate height have been performed, with respect to a single order parameter, employing the steady state properties of unconstrained simulation trajectories when barrier crossing is achievable with reasonable computational effort. Considering such cases, we describe a method to estimate free energy surfaces with respect to multiple order parameters from a steady state ensemble of trajectories. The approach applies to cases where the transition rates between pairs of order parameter values considered is not affected by the presence of an absorbing boundary, whereas the macroscopic fluxes and sampling probabilities are. We demonstrate the applicability of our prescription on different test cases of random walkers executing Brownian motion in order parameter space with an underlying (free) energy landscape and discuss strategies to improve numerical estimates of the fluxes and sampling. We next use this approach to reconstruct the free energy surface for supercooled liquid silicon with respect to the degree of crystallinity and density, from unconstrained molecular dynamics simulations, and obtain results quantitatively consistent with earlier results from umbrella sampling.
BibTeX
@article{goswami2023kinetic,
  author = {Yagyik Goswami and Srikanth Sastry},
  title = {{Kinetic reconstruction of free energies as a function of multiple order parameters}},
  journal = {The Journal of Chemical Physics},
  year = {2023},
  doi = {10.1063/5.0144338},
}

Loading related content...

Figure