PaperlayerView exact revision on arXiv

Source-linked AI summary

A Unified Approach to Enhanced Sampling

Michele Invernizzi, Pablo Miguel Piaggi, Michele Parrinello

arXiv:2007.03055v3physics.comp-ph

TL;DR

Atomistic simulations can produce poor results when rare events and metastable-state transitions are not adequately sampled. The paper adapts OPES to sample expanded ensembles through collective-variable-based biases. It demonstrates applications across several ensemble types and combinations, including temperature-related fluctuations with a system-specific order parameter.

  • Problem

    Atomistic simulations face a sampling challenge because kinetic bottlenecks and rare transitions can leave phase space inadequately explored, limiting result quality even for accurate models.

    Source: §I · p. 1

  • Method

    The paper adapts OPES to use expansion collective variables to define biases for expanded-ensemble target distributions.

    Source: §I · p. 1, §VIII · p. 13

  • Results

    The method is demonstrated for common expanded ensembles and for enhancing temperature-related fluctuations together with a system-specific order parameter.

    Source: §VIII · pp. 13–14

  • Takeaways & Limitations

    The approach is presented as a general, reliable method with few robust adjustable parameters and a reweighting procedure for recovering statistics at desired λ values.

    Source: §I · p. 1, B. Reweighting · p. 5

Abstract

from arXiv · show

The sampling problem lies at the heart of atomistic simulations and over the years many different enhanced sampling methods have been suggested towards its solution. These methods are often grouped into two broad families. On the one hand methods such as umbrella sampling and metadynamics that build a bias potential based on few order parameters or collective variables. On the other hand, tempering methods such as replica exchange that combine different thermodynamic ensembles in one single expanded ensemble. We instead adopt a unifying perspective, focusing on the target probability distribution sampled by the different methods. This allows us to introduce a new class of collective-variables-based bias potentials that can be used to sample any of the expanded ensembles normally sampled via replica exchange. We also provide a practical implementation, by properly adapting the iterative scheme of the recently developed on-the-fly probability enhanced sampling method [Invernizzi and Parrinello, J. Phys. Chem. Lett. 11.7 (2020)], which was originally introduced for metadynamics-like sampling. The resulting method is very general and can be used to achieve different types of enhanced sampling. It is also reliable and simple to use, since it presents only few and robust external parameters and has a straightforward reweighting scheme. Furthermore, it can be used with any number of parallel replicas. We show the versatility of our approach with applications to multicanonical and multithermal-multibaric simulations, thermodynamic integration, umbrella sampling, and combinations thereof.

I. INTRODUCTION

Atomistic simulations can fail to produce high-quality results when rare events and metastable-state transitions are not adequately sampled. The paper unifies enhanced-sampling methods by their target distributions and introduces a collective-variable-based approach for sampling expanded ensembles.

I. INTRODUCTION

  • Rare events and kinetic bottlenecks create ergodicity problems, so even accurate atomistic models may fail to produce high-quality results without proper phase-space sampling.Enhanced sampling raises the probability of rare events in a modified ensemble.Source: §I · p. 1

II. A UNIFIED APPROACH

  • Bias-based methods such as umbrella sampling and metadynamics accelerate slow modes by biasing selected collective variables.The bias is a function of the chosen collective variables.Source: §II · p. 2
  • Expanded ensembles combine overlapping distributions, for example across temperatures or Hamiltonians, and can be sampled with replica exchange or single-replica approaches.Multiple-window umbrella sampling and thermodynamic integration can also combine separated ensembles through post-processing.Source: §II · p. 2
  • Classifying methods by their target distribution suggests the two traditional enhanced-sampling families are not fundamentally distinct, though methods within a family can still use different strategies.The paper notes that each strategy has its own strengths and weaknesses.Source: §II · p. 2
  • The proposed framework uses expansion collective variables to define expanded-ensemble targets and biases, while also covering uniform and well-tempered targets used in metadynamics-like sampling.It is intended to combine the target types as well as sample them separately.Source: §II · p. 3

III. ON-THE-FLY PROBABILITY ENHANCED SAMPLING

OPES modifies the probability distribution along collective variables toward a chosen target. The paper extends this approach to expanded ensembles, defining target distributions and biases through expansion collective variables.

III. ON-THE-FLY PROBABILITY ENHANCED SAMPLING

  • OPES biases collective variables to transform their physical probability distribution toward a specified target distribution.Its well-tempered target lowers free-energy barriers along the collective variables by a factor γ; the uniform target is recovered as γ approaches infinity.Source: §III · p. 3
  • The expanded-ensemble target and bias extend OPES beyond its original well-tempered and uniform targets, with the target bias built from free-energy differences along λ.Those free-energy differences are estimated on the fly by reweighting.Source: §III · p. 3, §IV · p. 4

IV. TARGETING EXPANDED ENSEMBLES

  • For expanded ensembles, the method selects λ-points across a parameter range whose adjacent ensembles overlap, then targets their combined distribution.The paper considers non-weighted expanded targets, while noting that λ-dependent weights are also possible.Source: §IV · pp. 3–4
  • Expansion collective variables ∆uλ(x) determine the bias, free-energy differences, and expanded target distribution.The bias depends on coordinates through these quantities.Source: §IV · p. 4

A. Iterative OPES Scheme

The expanded-ensemble OPES scheme iteratively estimates free-energy differences from a biased trajectory and updates the applied bias. As sampling approaches the target, the estimates improve and also provide free-energy information across λ.

A. Iterative OPES Scheme

  • Reweighting a biased trajectory provides a more accurate estimate of ∆F(λ) than estimating it from an unbiased trajectory with very small overlap.The method uses this estimate to define the bias, so it updates the estimate and bias iteratively.Source: A. Iterative OPES Scheme · p. 4
  • As the bias approaches convergence, sampling approaches the target and the ∆F(λ) estimates become more accurate, yielding both the target bias and free-energy estimates across λ.The scheme is similar in spirit to integrated tempering sampling, but differs in implementation and applications.Source: A. Iterative OPES Scheme · p. 4
  • The explicit iterative equations underpin sampling different expanded ensembles by specifying different expansion collective variables.Once those variables are chosen, the iteration stride is the method’s only free parameter and should limit correlation between consecutive steps.Source: A. Iterative OPES Scheme · pp. 4–5
  • Multiple replicas can share the bias and contribute to the ensemble average, and their number need not equal the number of λ-points.This differs from replica exchange in the required relationship between parallel simulations and λ-points.Source: A. Iterative OPES Scheme · p. 5
  • Cumulant expansions can be inaccurate over broad λ-ranges and may introduce artificial barriers that hinder visits to relevant configurations.They can provide a good estimate close to the reference λ = 0.Source: A. Iterative OPES Scheme · p. 5

B. Reweighting

The reweighting procedure recovers averages at desired λ values from the biased simulation. Effective sample size and weighted block averaging support diagnostics and uncertainty estimation.

B. Reweighting

  • Standard umbrella-sampling reweighting retrieves the average of an observable at any desired λ from the biased simulation.The observable is any function of the atomic coordinates.Source: B. Reweighting · p. 5
  • Reweighting assumes a static or quasi-static bias, so the rapidly changing initial transient should be excluded from analysis.OPES generally assigns that transient very low weight, so the discard choice is not expected to significantly affect results.Source: B. Reweighting · p. 5
  • Effective sample size provides a diagnostic for reweighting, and its ratio to the total sample count can be monitored across λ.The expected efficiency is roughly neff(λ)/n ∝ 1/N{λ} for a minimal set of λ-points covering the target range.Source: B. Reweighting · p. 5
  • Uncertainty estimates for weighted samples require care, for which the paper uses weighted block averaging.The paper describes the procedure in Appendix A and explains the role of effective sample size.Source: B. Reweighting · p. 5

V. LINEARLY EXPANDED ENSEMBLES

Linearly expanded ensembles use a parameterized family of reduced potentials to sample related thermodynamic states together. In the multicanonical example, biasing potential energy enables repeated transitions between alanine-dipeptide basins, while a dihedral-based well-tempered run converges the single-temperature free-energy difference about ten times faster.

V. LINEARLY EXPANDED ENSEMBLES

  • Linearly perturbing the reduced potential, uλ(x) = u0(x) + λ∆u(x), produces ensembles including multicanonical, multibaric, and alchemical transformations.These ensembles share a practical automatic heuristic for choosing λ-points from effective sample size.Source: §V · p. 5

A. Multicanonical Ensemble

  • The multicanonical target covers configurations relevant to canonical simulations across a temperature range, using potential energy U as the collective variable.Here ∆u(x) = β0U(x), with λ spanning the target inverse-temperature range relative to the thermostat temperature.Source: A. Multicanonical Ensemble · p. 6
  • For temperature-expanded ensembles, the bias depends only on the thermodynamic conjugate variable ∆u, although the method does not require a single-CV bias.The free energy along λ must be estimated to define this bias.Source: A. Multicanonical Ensemble · p. 6

Example: Alanine Dipeptide

  • Biasing potential energy in the multicanonical ensemble increases transition-state visits and yields multiple basin transitions, allowing convergence of their free-energy difference.The potential energy is a poor CV for distinguishing the two basins, which have roughly the same energy.Source: Example: Alanine Dipeptide · p. 6
  • About 10 times faster at a single temperature, the well-tempered run biasing the dihedral angles converges the free-energy difference than the multicanonical run.The paper attributes this to focusing on the relevant degrees of freedom.Source: Example: Alanine Dipeptide · p. 6

B. Multithermal-Multibaric Ensemble

A multithermal-multibaric target combines temperatures and pressures in one expanded ensemble, represented using potential energy and volume. In the chignolin example, effective sample sizes are of the same order across the retained target region, and the estimated stability diagram has no striking qualitative features.

B. Multithermal-Multibaric Ensemble

  • A two-parameter expansion combines temperature and pressure, with potential energy and volume used as collective variables.The target spans specified temperature and pressure ranges and estimates free-energy differences across their grid.Source: B. Multithermal-Multibaric Ensemble · p. 6

Example: Chignolin

  • The chignolin simulation covers 270–800 K and 1–4000 bar, while excluding 91 temperature-pressure points to avoid conditions where water could evaporate.The simulation used 40 walkers and ran for 300 ns, with roughly 10 ns needed for bias convergence.Source: Example: Chignolin · p. 7
  • Effective sample sizes are of the same order of magnitude across the chignolin target region, though not perfectly uniform.The target excludes high-temperature, low-pressure conditions to avoid water vaporization.Source: Example: Chignolin · p. 7, Figure 2 · p. 8
  • The chignolin stability diagram has no striking features, qualitatively agreeing with Ref. 62.The folded fraction is defined using Cα-RMSD cutoffs of 0.1 nm for folded and 0.4 nm for unfolded configurations.Source: Example: Chignolin · p. 7

C. Thermodynamic Integration

Thermodynamic integration connects two systems through intermediate reduced potentials and integrates an ensemble average over the interpolation parameter. The paper adapts this to a single simulation that samples all parameter values and applies it to TIP4P water relative to a Lennard-Jones reference.

C. Thermodynamic Integration

  • A single simulation samples all λ values, enabling reweighting at each value and numerical evaluation of the thermodynamic-integration integral.The method uses expansion collective variables ∆uλ = λ∆u for linear interpolation.Source: C. Thermodynamic Integration · p. 8
  • The method also applies to interpolation schemes other than linear, provided the expansion collective variables are defined accordingly.Source: C. Thermodynamic Integration · p. 8

Example: TIP4P Water to Lennard-Jones

  • After an initial bias-optimization transient of about 3 ns, ∆U diffuses freely; the reported free-energy estimate uses values of ⟨∆U⟩λ across the λ range.The target uses 30 equally spaced λ-points selected from a short unbiased run.Source: Example: TIP4P Water to Lennard-Jones · p. 9
  • The TIP4P-to-Lennard-Jones free-energy difference is 7.00(1) NkBT units, in agreement with the reported reference result.The calculation uses TIP4P water at 443 K and a Lennard-Jones reference system.Source: Example: TIP4P Water to Lennard-Jones · p. 9

VI. BEYOND LINEARITY: MULTIUMBRELLA ENSEMBLE

The multiumbrella ensemble samples umbrella windows in one simulation using a global potential and obtains the free-energy surface by reweighting. In the double-well example, the method tolerates a suboptimal collective variable, though convergence is slower and broad umbrellas can oversmooth direct estimates.

VI. BEYOND LINEARITY: MULTIUMBRELLA ENSEMBLE

  • Multiumbrella sampling combines all umbrella windows in one simulation with a single global potential, then reconstructs the free-energy surface by reweighting.Unlike multiple-window sampling, it needs no further post-processing such as WHAM.Source: §VI · p. 9
  • Umbrella width σ controls the target: very narrow umbrellas flatten its CV marginal, while very broad umbrellas approach the original distribution; σ should match the smallest FES features of interest.Umbrella centers are spaced no farther apart than σ to ensure overlap; broad widths can produce small FES artifacts.Source: §VI · p. 10

Example: Double-well Model

  • A suboptimal x coordinate slows convergence in the double-well model but does not prevent sampling the barrier or require extra handling.The standard single-window approach would struggle at x = 0 because of a barrier along y.Source: Example: Double-well Model · p. 10
  • In the double-well example, changing σ changes the target distribution, and broad σ can oversmooth direct FES estimates, so reweighting is preferred.The reported reference scale for σ is the unbiased basin standard deviation, σ0 ≈ 0.18.Source: Example: Double-well Model · p. 10, Figure 5 · p. 11
  • Despite different biasing strategies, multiumbrella OPES and metadynamics or well-tempered OPES produce similar target distributions that allow FES reconstruction.Multiumbrella OPES keeps sampling within a chosen CV region, whereas metadynamics and well-tempered OPES push the system away from visited regions.Source: Example: Double-well Model · p. 10

A. Combining Thermodynamic and Order Parameter Expansions

The approach combines thermodynamic ensembles with order-parameter sampling through a bias potential. In the sodium example, reweighting yields phase and free-energy surfaces consistent with prior results, while biasing crystallinity samples the transition region and converges quickly despite a first-order transition.

A. Combining Thermodynamic and Order Parameter Expansions

  • OPES combines thermodynamic generalized ensembles with order-parameter sampling; it matches the target distribution of multidimensional replica exchange without requiring a minimum number of replicas.Compared with flexible VES, OPES uses fewer free parameters and is simpler to set up.Source: A. Combining Thermodynamic and Order… · p. 11
  • For first-order phase transitions, changing temperature may alter phase stability while leaving the barrier high along the coexistence line, motivating an order-parameter bias.The paper also identifies barriers orthogonal to an alchemical transformation as a setting where combining these approaches may help.Source: A. Combining Thermodynamic and Order… · p. 11

Example: Sodium

  • The sodium application uses the environment similarity collective variable as a crystallinity measure based on comparing atomic local environments with a reference.The variable is also described as applicable to a variety of phase transitions.Source: Example: Sodium · p. 11
  • The reweighted sodium phase diagram and CV free-energy surfaces agree with Ref. 40, while biasing the crystallinity CV samples the transition region and converges quickly despite the first-order transition.The transition-state probability remains extremely small across the studied range without active CV biasing.Source: Example: Sodium · p. 12

VII. ABOUT THE OPTIMAL TARGET DISTRIBUTION

The paper proposes effective sample size as a criterion for choosing an expanded target distribution. For reweighting across a parameter range, it favors targets yielding uniform effective sample size, with fewer intermediate parameter points closer to that criterion in the studied systems.

VII. ABOUT THE OPTIMAL TARGET DISTRIBUTION

  • The paper presents effective sample size as a criterion for judging whether an expanded target distribution is optimal.The section asks whether the non-weighted expanded target is optimal and notes that other target choices are possible.Source: §VII · p. 12
  • For reweighting across the full parameter range, the optimal target is the one yielding the highest uniform effective sample size; in studied systems, fewer intermediate λ points came closer to this criterion.The authors do not further explore this target or the definition of effective sample size.Source: §VII · p. 12
  • For a subset of λ-states, the proposed criterion is to maximize effective sample size for those states while preserving ergodicity.The authors cite their earlier argument that a well-tempered target can outperform a uniform target on this criterion by avoiding unimportant high-free-energy regions.Source: §VII · p. 12

VIII. CONCLUSION

The paper presents a unified framework that uses OPES and expansion collective variables to sample expanded ensembles. Its examples cover combined thermodynamic and order-parameter sampling, while the target distribution omits velocities and the authors identify higher-dimensional scaling as future work.

VIII. CONCLUSION

  • The framework uses OPES and expansion collective variables, Δuλ(x), to characterize non-weighted expanded targets, estimate free-energy differences, and define the target bias.The authors adapt OPES’s iterative scheme, originally introduced for metadynamics-like sampling, to sample expanded ensembles typically handled by another enhanced-sampling family.Source: §VIII · p. 13
  • The examples include simultaneous enhancement of temperature-related fluctuations and a system-specific order parameter.The conclusion also points to applications across common expanded ensembles and their combinations.Source: §VIII · pp. 13–14
  • The target distribution includes positional degrees of freedom but not atomic velocities, so its velocity-space sampling differs from replica exchange.The authors state that this difference does not affect statistical averages of coordinate-only observables.Source: §VIII · p. 14
  • Combining expanded targets with well-tempered-like distributions is proposed as future work because it can scale better with higher dimensionality.The authors also identify weighted expanded targets as a possible direction.Source: §VIII · p. 14

Appendix A: Weighted Block Average

The appendix describes uncertainty estimation for biased, weighted samples and an optional initialization safeguard for multiumbrella targets.

Appendix A: Weighted Block Average

  • The uncertainty protocol accounts for both time correlation and the effective sample size of weighted samples.Time correlation reduces effective sample size, so using the ordinary sample count can underestimate uncertainty.Source: Appendix A · p. 14
  • The weighted block procedure combines block estimates and computes uncertainty using the effective block size, Meff, rather than the total block count, M.The reported error is σO/√Meff; using M can considerably underestimate uncertainty when M is small or block weights are unbalanced.Source: Appendix A · p. 14
  • Table I pairs common expanded ensembles with their expansion collective variables and free-energy differences; target biases can use one or two CVs and can be combined.Source: Table I · p. 15

Appendix B: Improving Robustness for the Multiumbrella Target

  • For multiumbrella initialization, an optional barrier parameter caps initial free-energy estimates to prevent an excessively strong initial bias.The cap can be set to an estimate of the free-energy barrier to overcome; the initial estimates otherwise may become very large and cause the simulation to fail during its first biased steps.Source: Appendix B · p. 15
Loading 2007.03055v3…