Does Nanoparticle Curvature Soften the Water Cage Around a Reaction?

Computer models of reactions at metal–water interfaces now simulate every water molecule. These explicit-solvent models show that rearranging the surrounding water can be a large share of a reaction's energy barrier. Yet every such simulation in this literature uses an infinite flat crystal slab, while real catalysts are particles a few nanometres across. A separate literature shows that water's interfacial free energy (its surface tension) changes sharply when the interface curves on the nanometre scale. This note asks whether that curvature effect carries over into reaction barriers. Concretely: on a small particle, does the water around an adsorbed molecule cost less to rearrange? And does that saving depend on temperature in a way a flat-slab model cannot capture?

The core idea fits in two sentences. We borrow a trick from nucleation theory: fix the flat-surface value from separate slab simulations, then fit only one curvature-correction length. We apply it to the part of a reaction barrier that comes from the solvent rather than the metal, measure that length at several temperatures for carbon dioxide activation on silver in explicit water, and test whether it follows a curvature law or is just an edge effect. If it works, catalyst modelers gain a one-parameter, temperature-aware correction from slabs to particles. If it fails cleanly, we learn that area-weighted facet models are enough, which is itself worth knowing.

Core Claim

Our hypothesis concerns the solvent part of the barrier, which we define as the explicit-water barrier minus the barrier computed without water on the same metal geometry. We claim this quantity, $\Delta G_{\rm solv}^{\ddagger}$, shrinks on small particles following a Tolman-like law in the mean curvature of the water–metal interface:

$$ \Delta G_{\rm solv}^{\ddagger}(R,T) \approx \Delta G_{\rm solv}^{\ddagger}(\infty,T)\left(1 - \frac{c\,\delta_r(T)}{R}\right). $$

In this equation, $R$ is the particle radius and $c$ is the number of curved directions (2 for a sphere, 1 for a cylinder). $\delta_r(T)$ is a single "reorganization Tolman length" that may depend on temperature $T$. The $1/R$ form is the leading-order curvature expansion used for surface tension. We adopt it here because the rearrangement penalty is, physically, the cost of reshaping an interfacial hydrogen-bond network.

The claim is new in two ways. First, it tests a geometric law with a falsifiable signature: a cylinder should shift the barrier by half as much as a sphere of the same radius. The competing explanation is edge proximity, where only the distance from the active site to the nearest facet edge matters. That explanation does not predict this factor of two. Second, temperature is used to split the curvature effect into an energy part and an entropy part, rather than serving only as a knob that keeps the simulated water liquid.

What the prior work does and does not settle

Three flat-slab, explicit-solvent studies motivate the claim that water rearrangement matters.

  • Qin, Vegge and Hansen ran constrained ab initio molecular dynamics of CO₂ activation on Au(110) with 44 water molecules and K⁺ ions. The barrier rises from 0.36 eV in static calculations to 0.61 eV with dynamic water, which they attribute to breaking and reforming hydrogen bonds around the bending, charging CO₂.
  • Tian et al. used a MACE machine-learned potential (a neural network fitted to quantum-mechanical energies, roughly 100,000 times faster). They combined it with grand-canonical DFT (simulations held at a fixed electrode potential) and enhanced sampling on Ag(111). They report a 0.59 eV barrier for the rate-limiting step *CO₂ → *COOH at −1.5 V versus the standard hydrogen electrode, driven by a potential-induced flip of interfacial water.
  • Mikkelsen et al. showed that 10 ns neural-network simulations on Pt(111) remove artifacts of frozen bilayer models and reveal a ~2 Å water depletion gap.

All three stay on flat slabs and at one temperature. We keep their reactions and methods and vary geometry and temperature.

The curvature literature does two jobs: one paper motivates the idea and the others supply the method.

  • Motivation. Kwon et al. measured an effective surface tension of a water nanomeniscus below 20% of bulk under 25 nm radius. The authors note this is an "effective" quantity that lumps in hydration and double-layer forces. So it motivates the idea but cannot be converted directly into a barrier shift.
  • Method. Cheng, Tribello and Ceriotti supply the key technique. Fitting curvature corrections to cluster free energies alone traps the bulk and surface parameters in a correlated fit. Pinning them to planar-slab values leaves one well-determined Tolman length. Their system was a Lennard-Jones fluid with no hydrogen bonds and no chemistry. We reuse the pinning with the flat-slab barrier as the pinned quantity.
  • Constraints. Prestipino et al. 2012 show that interface thickness matters when clusters are small. Prestipino et al. 2014 show that shape fluctuations matter. Together they warn that $1/R$ may fail below some radius, so we report the radius where the fit breaks rather than forcing it.
  • Why not simple averaging. Luo et al. showed that area-weighted wetting rules fail on nanoscale patchy surfaces. This is the reason not to assume facet additivity, but it is also a reason to test additivity directly.

For kinetics, Polino et al. supply the rate method: infrequent metadynamics at 390 and 490 K for urea decomposition in explicit water. They found that plain transition-state theory overestimates rates by nearly two orders of magnitude because of water proton wires. We adopt their approach to obtain rates, not just barriers, at each temperature.

The macroscopic temperature studies in the third direction fit bulk Arrhenius or pseudo-order laws and cannot isolate the solvent. Examples are nZVI dye reduction and Cr(VI) chemisorption. They frame the application question but play no quantitative role here.

Research Program

System. We use Ag and the *CO₂ → *COOH step because a MACE model and a reported flat-surface barrier already exist for it.

Potential. We extend the potential by active learning on Ag clusters and nanowires in water. The added DFT configurations sample facet centres, edges and corners with adsorbed CO₂ and COOH, using the same functional as the slab model. Error is monitored by ensemble disagreement, following the four-model ensemble practice of Mikkelsen et al.

Geometries. We build three families at matched radii of roughly 1 to 4 nm:

  • truncated-octahedral particles, curved in two directions;
  • (111)-faceted nanowires, curved in one direction;
  • elongated particles whose (111) facet width varies at fixed radius.

The third family is what lets us separate curvature from edge distance.

Measurements. For each geometry we compute the free-energy barrier at the facet centre in explicit water with a K⁺ counter-ion at fixed charge. We use the same collective variables and enhanced-sampling protocol as the slab reference. We also compute the vacuum barrier on the identical relaxed metal geometry. The vacuum barrier captures strain and electronic size effects, and subtracting it isolates the solvent part.

Temperature and the functional. Every calculation is repeated at four temperatures. PBE-family water over-structures, which is why Mikkelsen et al. ran at 400 K. We therefore state all temperatures relative to the model's own melting point and use the range where the model's water is liquid. Temperature derivatives of the barrier give the energy and entropy parts of $\Delta G_{\rm solv}^{\ddagger}$ at each radius. Rate constants at the same four temperatures come from infrequent metadynamics.

Independent predictor. Separately, and with no reaction present, we compute the curvature-dependent interfacial free energy of water against each metal shape using the tandem pinning method. This tests a second, cheaper hypothesis: that $\delta_r(T)$ tracks the Tolman length of the bare water–metal interface. If it does, one reaction-free simulation per temperature would predict barrier shifts for any adsorbate.

Baselines

Naive reference. The flat-slab barrier is applied unchanged to every particle. The file gives Tian et al.'s 0.59 eV for *CO₂ → *COOH on Ag(111) at −1.5 V. Our fixed-charge setup is not at that potential, so we recompute the flat Ag(111) value in our own protocol and use it as the pinned $\Delta G^{\ddagger}(\infty)$. Qin et al.'s 0.61 eV on Au(110) serves only as a sanity check on magnitude.

Strongest existing method. Explicit-solvent simulations of planar slabs and stepped slabs are combined by area weighting over the sites a particle exposes. This is how a careful practitioner would currently predict particle behaviour, and it already includes water rearrangement and undercoordinated sites. For fairness, the slabs use the same potential, the same sampling budget per window, the same temperatures and the same counter-ion.

Criterion for beating the baselines. The curvature law beats the naive reference if the barrier on the smallest particles differs from the flat value by more than three times the combined statistical error. Mikkelsen et al. report ~0.025 eV precision for comparable simulations, so we aim for ≤0.03 eV per barrier. The law beats the area-weighted slab combination if it predicts held-out geometries with lower error, especially the cylinder-versus-sphere ratio.

Ablations. Three ablations isolate the ingredient under test:

  • vacuum barriers on the same particles, removing the solvent;
  • implicit-solvent barriers on the same particles;
  • a frozen-water ablation, in which water is equilibrated around the particle and then held fixed during the reaction.

Why not just…?

Why not just compute slab barriers for terraces and steps and area-weight them? This is the most serious threat, and we treat it as a live hypothesis rather than a strawman. Luo et al. show that additivity fails for wetting on nanoscale patches, but wetting is not a reaction barrier. The elongated-facet family settles the question. If barriers depend only on distance to the nearest edge, and cylinders and spheres of equal radius give equal shifts, additivity plus an edge term wins and the curvature law is rejected.

Why not just use implicit solvent with a curvature-dependent cavity? Qin et al. show that non-dynamic treatments give 0.36 eV where dynamic water gives 0.61 eV, because the rearrangement work is missing. Tian et al. report that implicit models fail even to predict CO₂ binding at zero charge. An implicit model could still reproduce a curvature trend on top of a wrong absolute barrier. The implicit-solvent ablation tests exactly that, by comparing its predicted $\delta_r$ to the explicit one.

Why not just rescale the barrier with measured nanoscale surface tension or the textbook Tolman formula? Kwon et al.'s number is an effective tension that mixes several forces. It was measured on a water meniscus in air, not on water against a metal. Cheng et al. show that unpinned curvature fits are unreliable. The reaction-free tandem calculation is our principled version of this shortcut. If its Tolman length predicts $\delta_r$, the shortcut becomes the recommended method.

Why not just simulate at one temperature and assume Arrhenius behaviour? Polino et al. find that water dynamics can make transition-state rates wrong by nearly two orders of magnitude. The split between energy and entropy is precisely where curvature and the functional's temperature shift could interact. If $\delta_r$ turns out independent of temperature, we will say so, and future work can safely use a single temperature.

Decisive Checks

Five outcomes decide the proposal.

  1. Is there an effect at all? Does the solvent part of the barrier change with radius beyond 0.1 eV noise? If not, flat slabs suffice and the note reports a useful negative result.
  2. Curvature or edges? Is the cylinder-to-sphere shift ratio near one half, as the curvature law predicts, or near one, as edge proximity predicts?
  3. Transferability. Does a $\delta_r$ fitted on two radii with the slab pinned predict the barrier on a held-out third radius and on the nanowires?
  4. Temperature dependence. Does $\delta_r$ change with temperature, and does that change come from entropy? The hypothesis of a softened water cage predicts an entropic origin.
  5. Cheap predictor. Does the bare water–metal Tolman length predict $\delta_r$?

No reported result was found on Lacuna for any barrier on a curved metal particle in explicit solvent. Every particle number in this study will therefore be new, and the only quoted comparison points are the flat-slab values above.