Well-Tempered Metadynamics in Practice: A CeO2 Electrode in Saline Water Case Study

    This article walks through a complete well-tempered metadynamics (WT-MTD) workflow for studying Cl–O coupling and HOCl formation on a CeO₂(111) electrode in NaCl solution. It covers the three-stage methodology—base AIMD relaxation, unbiased collective variable (CV) probing, and biased sampling—explaining how Gaussian widths, bias factors, Hill heights, deposition intervals, and boundary walls are calibrated from unbiased CV fluctuations. It also addresses convergence checks, reweighting, common pitfalls, and large-scale job orchestration. Readers will learn not just how to configure WT-MTD inputs, but why CV selection and width calibration—not the software setup—determine success.

    1. Background and the Core Question

    In neutral chloride-containing aqueous solutions, the chlorine evolution reaction (CER) on metal oxide anodes is an important anodic process. It competes with the oxygen evolution reaction for the same key intermediate: the surface-adsorbed chlorine atom. This intermediate has two possible fates:

    • Coupling with another adsorbed chlorine to form chlorine gas—the CER pathway;
    • Coupling with a surface oxygen species to form a Cl–O bond, evolving toward chlorine-oxygen products such as hypochlorous acid (HOCl) or hypochlorite.

    In early chlor-alkali operations, oxygen-involving Ru-based anodes exhibited rising oxygen-species coverage and falling current efficiency—essentially the competition between chlorine and oxygen evolution. If Cl–O coupling becomes the dominant channel under certain conditions, it amounts to adding a new branch at the fork between the two reactions. Contested access to this pathway is a key piece of the chlorine electrocatalysis selectivity puzzle.

    The specific questions addressed here are: in a system consisting of a CeO₂(111) electrode and NaCl aqueous solution, can spontaneous Cl–O bond formation be observed? Can hypochlorous acid be obtained along this pathway? And can the free energy barrier of this pathway relative to competing routes be quantified?

    Answering such questions requires satisfying two conditions simultaneously:

    1. Sufficient electronic-structure accuracy. Bond formation and breaking happen at the first-principles level; semiempirical potentials cannot reliably describe charge transfer, surface hydroxyls, and lattice oxygen behavior.
    2. Sufficiently broad sampling. Spontaneous coupling is a classic rare event; plain ab initio molecular dynamics (AIMD) almost never observes it within a feasible time budget.

    Well-tempered metadynamics (WT-MTD) is one of the mainstream enhanced sampling methods for the second problem. Its core idea:

    • Deposit Gaussian-shaped bias potentials (Hills) repeatedly along chosen collective variables (CVs);
    • Progressively fill in the free energy basins of already-sampled regions, forcing the system over barriers;
    • Rely on the well-tempered mechanism to make the bias “slow down” automatically, eventually converging to a free energy surface (FES) in CV space.

    A common misconception should be cleared up first: the success of WT-MTD depends far less on how the software is installed or how the input is written than on whether the CVs are well chosen and whether the Gaussian widths are set sensibly. These two points are not intrinsic to WT-MTD itself; below, “metadynamics (MTD)” refers to this method family generically. Setting widths by gut feeling leads, at best, to spurious FES barriers or missing intermediates—and at worst, to a sampling direction distorted by the bias. Using a concrete CeO₂–brine system as an example, this article walks through the full three-stage pipeline: relax first, probe second, bias last.

    2. Method Essentials: Understand the Bias Before Setting Parameters

    Metadynamics adds a time-evolving bias potential along the CV (shown here as one-dimensional). At time t, the bias at CV value s is:

    V(s, t) = ∫ W(s’, t) · exp[ −(s − s’)² / (2σ²) ] ds’

    where W(s’, t) is the height of the Hill deposited at s’ at time t, and σ is the Hill width (the standard deviation of the Gaussian).

    In classical metadynamics, every Hill has a constant height W₀ and the bias grows roughly linearly with time, eventually “over-filling” free energy basins so the trajectory oscillates in CV space without converging. WT-MTD improves on this by making each new Hill’s height decay with the current bias:

    W(s, t) = W₀ · exp[ −V(s, t) / (k_B·ΔT) ]

    In the long-time limit, the bias approaches V(s, t → ∞) = −(1 − 1/γ)·F(s) + C, where:

    • γ = (T + ΔT) / T is the bias factor;
    • F(s) is the free energy surface.

    The biased trajectory evolves with a distribution ∝ exp[−F(s)/(k_BT) − V(s,t)/(k_BΔT)]. Because the bias cancels most of the free energy landscape, CV sampling is “flattened out.”

    Two conclusions follow that directly guide parameter tuning.

    Conclusion 1: Width is the spatial resolution; height sets the timescale for filling basins.

    • σ is the spatial extent of the Gaussian in CV space, equivalent to a smoothing scale of the FES: any free energy variation on scales smaller than σ gets smeared out by the Hills themselves;
    • The Hill height, via the factor exp(−Δ²/(2σ²)), controls how quickly the bias decays with distance from the deposition point, and thus how fast the overall bias builds up.

    This decay factor is steep. For a Gaussian kernel, the value is about 0.61 at one σ from the center, about 0.14 at two σ, and below 0.012 at three σ. If σ is clearly larger than the unbiased fluctuation amplitude of the CV, configurations sampled between two consecutive deposits are nearly indistinguishable—the bias treats them as the same region and piles up quickly. If σ is much smaller than the fluctuation amplitude, each deposit affects only a tiny neighborhood, the bias rises very slowly, and thousands of Hills are needed to fill a single basin.

    Conclusion 2: The well-tempered mechanism does not automatically fix a badly chosen width.

    Hill heights decay with bias depth, but once the width is chosen it remains a fixed smoothing scale throughout the simulation, never adapting as the system evolves. This is the fundamental reason “Gaussian widths cannot be set arbitrarily,” and why Stage 2 must exist as a separate step.

    One often-overlooked detail: if a program scales Hill heights proportionally to k_BT, then W₀/(k_BΔT) = T/ΔT is temperature-independent. In other words, under conventional usage, temperature does not change how the WT-MTD bias is constructed—σ and γ are the knobs that actually require careful calibration.

    3. Overall Workflow: Three Stages

    The workflow has three stages:

    • Stage 1 | Base sampling and structural relaxation: Use plain AIMD to obtain a well-relaxed, structurally stable model, and learn the system’s basic dynamics in the absence of bias.
    • Stage 2 | Unbiased CV probing: With bias off, observe the natural fluctuations of candidate CVs to obtain quantitative inputs for width selection and CV screening.
    • Stage 3 | Biased metadynamics sampling: With well-calibrated bias parameters, explore new coordination states progressively and converge the free energy surface.

    The key to Stage 2 is switching off the Hill deposition flag in the bias sampling settings: the system then evolves on the physical potential energy surface, equivalent to plain MD. This is the baseline for all calibration. The flag is not unique to WT-MTD—every metadynamics simulation needs it; WT-MTD simply adds the bias factor so that the bias decays with depth automatically.

    4. Stage 1: Model Construction and Base AIMD

    4.1 Building the Model

    CeO₂ adopts the fluorite structure; the (111) facet is the most exposed and thermodynamically most stable in experiments, making it a standard representative surface for oxide electrode modeling. When building the model, pay attention to:

    • Enough atomic layers so that the two surfaces do not couple to each other;
    • A vacuum layer to isolate periodic image interactions;
    • A dipole correction to cancel the net dipole introduced by asymmetric configurations along the surface normal (e.g., OH/H coverage).

    The liquid environment is a water slab thick enough to hold several layers of water molecules. Its thickness should clearly exceed water’s structural correlation length, to avoid artificial coupling between the two interfaces through the water film. In this example, 60 water molecules plus one NaCl pair (one Na⁺ and one Cl⁻) are added to build a neutral electrolyte solution.

    The neutral system is built deliberately, to avoid net charge and the associated uniform background charge in subsequent AIMD. Neutrality also better resembles open-circuit or weakly polarized initial conditions, letting us first examine the system’s most “intrinsic” reaction tendencies.

    The charge state deserves attention: in neutral aqueous solution, chlorine exists as Cl⁻. Forming HOCl requires oxidation first—losing electrons. Open-circuit AIMD with no applied potential and no net charge cannot directly capture that process. Metadynamics here answers the following: when the system locally develops highly oxidized sites, holes, or strongly oxidizing intermediates, is Cl–O coupling and HOCl formation favorable in free energy, how high is the barrier, and does the system prefer to stay in an HOCl configuration? This caveat must be stated in the scope of any conclusions.

    4.2 Base AIMD

    Run plain AIMD before any bias, to obtain thermally equilibrated structures and a liquid environment, and to accumulate baseline data for later analysis. An illustrative configuration follows (key–value syntax is for illustration only, not the grammar of any specific code):

    1. # Illustrative configuration: base AIMD (key-value syntax, not a specific program's grammar)
    2. run_type = md
    3. ensemble = nvt
    4. temperature = 300.0 # units: K
    5. thermostat = nose_hv
    6. thermostat_time = 200.0 # thermostat relaxation time, in fs
    7. timestep = 0.5 # fs; recommended scale for hydrogen-containing systems
    8. md_steps = 20000 # 0.5 fs x 20000 = 10 ps
    9. dcd_freq = 10 # write a frame every 10 steps
    10. print_forces = .false. # no need to print forces per frame during equilibration; saves I/O
    • Timestep: ~0.5 fs is typical for systems containing hydrogen;
    • Trajectory output frequency: too dense and I/O becomes the bottleneck; too sparse and very short-lived structural events may be missed. A useful reference is the ability to resolve changes on the scale of O–H stretching vibrations;
    • Equilibration criteria: whether temperature, potential energy, and the Cl–water and Cl–surface coordination numbers stabilize over time.

    4.3 System Checks

    After equilibration, confirm the model is “usable,” focusing on four areas:

    1. Surface stability: CeO₂ surface oxygens remain intact with no anomalous reconstruction; the ratio of lattice oxygen to surface hydroxyls matches expectations.
    2. Water behavior at the surface: water partially dissociates on oxide surfaces to form surface hydroxyls—normal physical chemistry, but it affects the later definition of “surface oxygen.”
    3. Ion behavior: whether Cl⁻ spontaneously enters the inner Helmholtz layer and its average residence time; whether ion pairs form; whether the local solution density is reasonable.
    4. Pseudopotentials and functional: whether the pseudopotential choice matches the reported data, and whether strong-correlation corrections are included.

    A concrete recommendation: with a closed-shell electronic-structure setup, verify that the spin-polarization setting matches the system’s actual electronic state. If electronic SCF steps fail to converge or occupancies change abruptly, the system likely has open-shell tendencies. A charge-neutral supercell is not necessarily “nonmagnetic” or a singlet—a point especially common in transition-metal oxide systems.

    Also ensure Stages 1, 2, and 3 use exactly the same structural model, functional settings, pseudopotentials, and cutoff energies. Any inconsistency among the three invalidates the widths calibrated in Stage 2.

    5. Stage 2: Unbiased Sampling and CV Diagnostics

    5.1 Principles for Choosing CVs

    A good CV should:

    • Have clear physical meaning, directly corresponding to the key bonding changes of the target event;
    • Be continuous and smooth, ensuring stable bias and force calculations;
    • Offer high resolution: initial, transition, and final states should be as separated as possible in CV space;
    • Be low-dimensional: sampling cost grows exponentially with dimension; if one dimension suffices, do not add a second.

    Hard-cutoff coordination numbers are discontinuous at the cutoff radius and should not be used directly as CVs. In practice, smooth switching functions are used, e.g.:

    f(r) = [1 − (r/r₀)ⁿ] / [1 − (r/r₀)ᵐ] (n < m; commonly n = 6, m = 12)

    As r ≪ r₀, f → 1; as r ≫ r₀, f → 0, with a smooth transition near r₀. The coordination number of one species around a central atom is the sum of the f values over all relevant atoms.

    For this system, a set of candidate CVs can be designed:

    ID Definition Physical meaning Region of interest
    CV1 Coordination number of Cl with surface metal sites, r₀ ≈ 2.8 Å Whether Cl moves from its hydrated state into inner-sphere surface coordination, and how coordination evolves Small variations around 0.1–0.2
    CV2 Cl–O_min distance, where O_min is the nearest oxygen (surface or water) within Cl’s neighborhood How close Cl gets to oxygen species; a precursor coordinate for Cl–O bond formation Fast fluctuations due to solvent exchange
    CV3 Coordination number of Cl with surface oxygen, r₀ ≈ 3.0 Å Direct measure of Cl–O interaction strength Whether it jumps from near 0 toward 1
    CV4 Coordination number of Cl’s first hydration shell, r₀ ≈ 3.3 Å Degree of Cl desolvation Ligand-exchange process

    Note that when coordination numbers serve as CVs, r₀ is not arbitrary: it is generally taken near the first minimum of the corresponding radial distribution function—the boundary between the first and second coordination shells. This is consistent with the n, m recommendations above.

    5.2 What Unbiased Sampling Should Do

    In this stage, set the Hill deposition flag to off and run a medium-length trajectory with the same model and settings as Stage 1—say 10–20 ps, extended to 50 ps if necessary—and output the CV time series at every step.

    There is only one criterion: the sampling length must be significantly longer than the CV’s correlation time. If a CV completes a full fluctuation cycle every 1 ps, 10 ps suffices to characterize its distribution; if the correlation time is tens of ps, a 10 ps sample has almost no statistical value.

    5.3 A Realistic Example

    In this example, CV1 fluctuates only slightly between 0.1 and 0.2 throughout the unbiased run. This means that under equilibrium conditions, the interaction between Cl and the surface metal sites (the “Ru” in the original study should be read as an illustrative surface metal site) barely changes—Cl does not spontaneously enter inner-sphere coordination. CV3 stays close to 0 the whole time: under pure thermal fluctuations, Cl–O contact events essentially never happen. That is precisely the signal that bias is needed.

    1. import numpy as np
    2. def cv_stats(path, col=1, dt=0.0005):
    3. """Read a CV time series; return mean, standard deviation, and integrated autocorrelation time (in ps)."""
    4. s = np.loadtxt(path)[:, col]
    5. x = s - s.mean()
    6. n = len(x)
    7. # Unbiased autocorrelation estimate
    8. acf = np.correlate(x, x, mode="full")[n - 1:] / (np.arange(n, 0, -1) * x.var())
    9. # The first lag where the ACF crosses zero serves as the integration cutoff,
    10. # avoiding accumulated bias from the noisy tail
    11. k = int(np.argmax(acf < 0)) or n
    12. tau = dt * (1 + 2 * acf[1:k].sum())
    13. return dict(mean=s.mean(), std=s.std(),
    14. p5=np.percentile(s, 5), p95=np.percentile(s, 95), tau=tau)
    15. print(cv_stats("colvar.dat"))

    5.4 From Fluctuations to Parameters: Four Deliverables of Stage 2

    • Lower bound and order of magnitude for Gaussian widths: For CV1, the unbiased fluctuation is roughly 0.1–0.2. Gaussian widths are typically set to 0.5–1.5 times the unbiased fluctuation (standard deviation) and fine-tuned through convergence tests in the production run. If the CV is normalized to the [0, 1] interval (commonly via a tanh-form distance function), the fluctuation amplitude shrinks considerably, and the width should be correspondingly smaller.
      • The width has a physical upper bound: it must not span the separation between neighboring coordination states. If the coordination-0 and coordination-1 states differ by only ~0.8 in CV space, a width of 0.5 makes the Gaussian tails of the two states overlap heavily, degenerating the FES into a smooth curve and erasing the “double well.” Hence 0.1–0.2 is a natural starting point for width selection.
      • Too large a width places part of the bias in regions near—but not yet reached by—the updated CV, effectively “pre-biasing” unexplored configurations. Combined with the non-Markovian nature of Gaussian basis functions, this introduces systematic FES errors; too small a width slows convergence.
    • Screening candidate CVs: If a CV barely changes over the entire unbiased run, it may not be a suitable coordinate for the target event and needs redefinition. If it fluctuates excessively across multiple regimes, it is “dirty” under equilibrium conditions—biasing it wastes many Hills on regions that never needed filling. The ideal case: the CV dwells in a few well-defined regimes separated by clear gaps.
    • Which CVs need biasing: Here, CV1 varies only in a narrow window and CV3 stays near zero—both classic cases of “does not change spontaneously, needs bias to change.” Conversely, if CV2 shuttles back and forth between 2–4 Å, hydration-shell exchange happens rapidly on this timescale; making it an independent biased dimension would only inject noise. It is better used as an auxiliary analysis quantity.
    • Configuration checks: At moments when the CV reaches characteristic values, extract several frames and verify that the corresponding chemical pictures match expectations (is Cl still surrounded by water, are surface hydroxyls intact, any anomalous bond breaking?). CV statistics must serve the physical picture, not the other way around.

    6. Stage 3: Biased WT-MTD Sampling

    6.1 Key Parameters and Their Calibration

    Parameter Role Recommended approach
    Hill width σ Smoothing scale of the FES 0.5–1.5 times the CV’s unbiased fluctuation, confirmed by convergence tests
    Bias factor γ Controls bias decay rate and effective temperature 5–20; increase for high-barrier systems, decrease when fine FES structure matters
    Initial Hill height W₀ Rate of bias buildup Match to the unbiased fluctuation; typically a few k_BT, scaled down proportionally as σ increases
    Deposition interval Number of Hills per unit time Far shorter than the CV’s round-trip correlation time; typically hundreds to thousands of steps
    Buffer period Equilibration before production statistics Deposit but do not accumulate statistics until the system re-equilibrates under the new bias
    Boundary walls Prevent CV escape into unphysical regions Place outside the target interval with ~1–2σ of margin
    1. Width: 0.5–1.5 times the unbiased standard deviation, finalized via convergence tests.
    2. Bias factor: 5–20; larger for high barriers, smaller when fine free energy structure must be resolved.
    3. Hill height: The initial height should match the width so adjacent Hills overlap sufficiently; a few k_BT is the usual reference scale. At room temperature, 1 k_BT ≈ 2.5 kJ/mol. If the width doubles, scale the height by roughly 1/2 to keep the deposited “volume rate” approximately constant.
    4. Deposition interval: Must be far shorter than the CV’s round-trip correlation time; commonly one Hill every few hundred to a few thousand MD steps.
    5. Buffer period and boundaries: Before production, run a deposit-only equilibration period, and apply upper/lower boundary walls to the CV. Place the walls outside the target interval with about 1–2 widths of margin, so Gaussian tails have essentially decayed before reaching the wall. Otherwise, structures pushed toward the boundary contaminate the free energy estimate.

    6.2 Illustrative Configuration

    1. # Illustrative configuration: WT-MTD (key-value syntax, not a specific program's grammar)
    2. &METADYN
    3. DO_HILLS = .true. # Hill deposition flag: .false. gives the unbiased MD of Stage 2
    4. NT_HILLS = 4000
    5. WELL_TEMPERED = .true.
    6. BIAS_FACTOR = 12.0 # gamma = 12
    7. SIGMA = 0.15 # CV1 fluctuates 0.1-0.2; take a value near its standard deviation
    8. HILL_HEIGHT = 5.0 # units: kJ/mol, about 2 k_BT
    9. HILL_SPACING = 500 # deposit one Hill every 500 steps (0.5 fs x 500 = 0.25 ps)
    10. TEMPERATURE = 300.0
    11. &BOUNDARY
    12. LOWER = 0.0
    13. UPPER = 2.0 # covers the full range from outer- to inner-sphere Cl coordination
    14. &END BOUNDARY
    15. &PRINT
    16. &HILL_FILE OFFSET = 0
    17. &END
    18. &END PRINT
    19. &END METADYN
    20. &COLVAR
    21. KIND = coordination
    22. ATOMS_FROM = 1 # Cl
    23. ATOMS_TO = 41 42 # surface metal sites (indices illustrative)
    24. R0 = 2.8
    25. NN = 6
    26. NM = 12
    27. &END COLVAR

    Three easily missed points in this configuration:

    • When DO_HILLS is .TRUE.: a new Hill is added and the bias applied each time the deposition interval is reached; when .FALSE., the system always evolves on the unbiased potential surface and no Hills are written.
    • Setting .TRUE. alone is not WT-MTD: without the well-tempered flag, the bias keeps growing linearly. The “deposit Hills” flag and the “well-tempered” flag are independent; both must be set correctly to obtain the bias form needed for Stage 3.
    • Syntax varies across programs: some implementations specify Hill width as a variance rather than a standard deviation, or use deposition spacing instead of Gaussian width—check the definitions of your implementation before use.

    For a CV like CV1 that fluctuates in a narrow window, you can also first push Cl into the inner-coordination region using “bias + a lower boundary wall,” observe how the system responds under this artificial constraint, roughly determine the CV range of the target state, and then design the production sampling interval accordingly. Conclusions from such exploratory runs are valid only for parameter tuning, not as free energy results.

    6.3 Running and Monitoring

    In the production run, keep the Hill deposition flag .TRUE. and monitor four indicators:

    1. CV versus simulation time: does it shuttle between the initial and target states?
    2. Bias distribution over the CV: does the bias grow uniformly across intervals—i.e., are all free energy basins filled to the same level?
    3. Line count of the Hills file: does it grow linearly as expected, i.e., is deposition proceeding normally?
    4. Whether the system “runs away”: if the Cl–metal distance rapidly shrinks to an unphysical value (e.g., below 2 Å), the width or height is too large—pause and recalibrate.

    For multi-step processes, first coarse-scan with a continuous CV (such as a path CV or a sum of pairwise distances) to establish the overall route; then redefine CVs around key intermediates for refined one- or two-dimensional sampling. This is safer than stacking high-dimensional CVs from the start.

    7. Analysis: From Biased Trajectory to Free Energy Surface

    7.1 Convergence Assessment

    • Block averaging: Split the trajectory into equal time blocks, estimate F(s) for each, and overlay them. If the FES differences between adjacent blocks fluctuate toward zero in the target interval, the result has converged.
    • Repeat simulations: Change the random seed or the initial configuration and run 2–3 independent simulations; overlay their FES and check the spread in barrier heights and minima positions.
    • Hessian back-check: Estimate local curvature from the converged F(s) and compare with the actual sampled distribution. A large discrepancy indicates the bias has not filled the basins adequately, or sampling has not reached quasi-equilibrium.

    7.2 Slicing Along the Minimum Free Energy Path

    Extract representative configurations along the minimum free energy path and subject them to standard static optimization and more rigorous transition-state searches. This both verifies that the metadynamics pathway holds at the electronic-structure level and yields more accurate barriers.

    7.3 Consistency Check Against Unbiased References

    Reweight the metadynamics trajectory by exp[βV(s,t)] to recover the unbiased distribution:

    P(s) ∝ ⟨ δ(s − s(t)) · exp[β·V(s(t), t)] ⟩_t

    If the reweighted CV distribution agrees with the Stage 2 unbiased trajectory over overlapping ranges, the sampling and bias construction are self-consistent. The reweighted result can also be used to back out the unbiased probabilities of Cl–O contacts, HOCl configurations, etc., for comparison with experimental observations.

    7.4 From Free Energy to Mechanistic Judgment

    With the FES in hand, return to the original questions: in the CeO₂(111)–NaCl aqueous system,

    • If the Cl–O bound state is a pronounced minimum on the FES and its barrier is comparable to competing channels, spontaneous Cl–O bond formation is indeed observable under those conditions, and HOCl can follow;
    • If the bound state is only a shallow, narrow metastable region, it more likely appears as transient contacts, and product selectivity will be governed by the subsequent rate-determining step.

    On metal oxide anodes, chlorine evolution generally follows an adsorbate evolution mechanism, passing through Cl and oxygenated intermediates such as ClO and ClOO* before yielding Cl₂ or chlorine-oxygen products. This provides the rationale for the CV choices and pathway design in this example.

    8. Batch Execution and Engineering

    8.1 Multi-Walker Parallelism

    When using multi-walker parallelism (several independent trajectories sharing one bias potential), make sure the walkers’ initial configurations are spread across different sampling regions. Otherwise they redundantly fill the same basin and waste compute.

    8.2 Platform-Level Recommendations

    • Queued scheduling: hand batch jobs to a batch compute service or container platform, with retry-on-failure and priority policies configured, so no single job monopolizes the queue;
    • Data management: archive trajectories and CV outputs in object storage, bucketed per job, to ease cross-job comparison;
    • Observability: collect CV evolution and Hill counts from each job via a logging service, and use monitoring alerts to track abnormal exits and stalled output files.

    8.3 Illustrative Batch Script

    1. #!/usr/bin/env bash
    2. # Illustrative batch script: parallel scan over CV and bias parameter sets
    3. set -euo pipefail
    4. common=(--cv 1 --model ceo2_111_nacl)
    5. for sigma in 0.10 0.15 0.20; do
    6. for gamma in 8 12 16; do
    7. for seed in 1 2 3; do
    8. tag="s${sigma}_g${gamma}_r${seed}"
    9. job_id=$(batch-submit \
    10. --name "wtmtd_${tag}" \
    11. --params "${common[*]}" \
    12. --sigma "$sigma" \
    13. --bias-factor "$gamma" \
    14. --seed "$seed" \
    15. --hill-output "bos://wtmtd-results/${tag}/") # illustrative: object storage path
    16. log-query --job "$job_id" --keywords "HILL_ADDED,NORMAL Termination"
    17. done
    18. done
    19. done

    Note: the typical fluctuation amplitude of CV1 is not transferable across systems (Cl with Ru, Cl with Cu, Cl with Ce); the scan above is illustrative only. First fix the order of magnitude of σ in Stage 2, then narrow the scan range—otherwise most of the 27 jobs may land in ineffective parameter regions.

    9. Quick Troubleshooting

    • The CV “freezes” under bias: usually the width is too large, making the bias too flat in CV space so the system loses its “slope.” Reduce σ, or reduce the Hill height or increase the deposition interval.
    • Thousands of Hills deposited but the CV barely moves: usually the width is far smaller than the CV fluctuation, or deposition is too frequent, so each Hill affects only a tiny neighborhood.
    • A “slope” appears in the FES: usually insufficient basin filling or improper boundary handling. Check the buffer region and wall placement, and extend the simulation.
    • Large barrier differences between repeat simulations: first suspect an incomplete CV—hidden degrees of freedom correlated with the event but not captured by the CV. Redesign the CV rather than simply running longer.
    • The CV’s unbiased fluctuation is itself very large: the CV cannot distinguish initial from final states. Either switch variables, or use restrained simulations to “anchor” the initial and final states before calibration.

    10. Summary

    • To study Cl–O coupling and HOCl formation in the CeO₂(111)–NaCl aqueous system, the key is to follow the pipeline “base AIMD relaxation → unbiased CV probing → biased WT-MTD sampling,” rather than applying bias from the outset.
    • Gaussian widths come from the CV’s natural fluctuations: CV1 fluctuates only between 0.1 and 0.2 in the unbiased run, and widths on that order avoid both smearing out physical detail and stalling convergence. The bias factor, Hill height, deposition interval, wall buffering, and width form one integrated set that must be calibrated together.
    • The Hill deposition flag only determines whether bias is applied; the well-tempered flag only determines how the bias decays. Both must be configured correctly for genuine WT-MTD.
    • Convergence should be judged by block averaging and repeat simulations; agreement of reweighted results with the unbiased reference is an important self-consistency check.
    • For this example, under neutral open-circuit conditions chlorine exists as Cl⁻, so judging “whether hypochlorous acid forms spontaneously” requires making the oxidation conditions explicit at the model level and limiting conclusions accordingly.