Chemists often claim that one crystal face of a catalyst is more active than another. Cobalt oxide that exposes one face oxidizes carbon monoxide about 2.7 times faster per surface cobalt ion than cobalt oxide exposing a different face. Claims like this are usually tested by making two particle shapes, measuring each in bulk, and dividing the rates. A shape, however, never exposes only one face. Edges, corners, defects and size all change along with it, and particles of the same shape still differ from one another. This proposal asks a question that particle physics and cosmology ask routinely but catalysis has not asked formally: what is the smallest facet-activity ratio that a given experiment could ever resolve, and which choice of particle shapes and observation time makes it resolvable?
The central claim is that, once single particles can be counted, facet attribution is limited by two things. The first is particle-to-particle variability, which acts as a noise floor much like cosmic variance in cosmology. The second is how well the chosen set of shapes separates facet area from edge length and corner count. Counting noise, the familiar $\sqrt{N}$ Poisson scatter in the number of reactions observed, matters much less. A two-shape comparison can be structurally unable to separate a facet effect from an edge effect, however many reactions are counted. We will build an estimator that assigns activity to facets using known geometry, derive its sensitivity limit, and use that limit to choose shape libraries that make facet claims testable at stated significance.
Core Claim
Model particle $i$, observed for time $t_i$, as producing $y_i$ turnovers (single reaction events). The expected count is $t_i \lambda_i g_i$, where $\lambda_i$ is the particle's expected turnover rate and $g_i$ is a gamma-distributed multiplicative factor with mean 1 and variance $\phi$. The factor $g_i$ captures defects and other variability between particles that geometry does not explain. The expected rate is
$$\lambda_i = \sum_f k_f A_{if} + k_e L_i + k_c C_i ,$$
where $A_{if}$ is the area of facet $f$ on particle $i$, $L_i$ its total edge length, $C_i$ its corner count, and $k_f, k_e, k_c$ the per-area, per-length and per-corner rate constants to be inferred. A facet contrast such as $\log(k_{011}/k_{001})$ then has a variance with two parts. One part shrinks as more turnovers are counted per particle. The other is set by $\phi$ and does not shrink no matter how long each particle is watched; only more particles reduce it. Both parts are scaled by the inverse of the Fisher information of the geometry matrix, meaning how much independent spread in facet area, edge length and corner count the shape library provides.
Two consequences follow. In the regime where counting is cheap, the minimum detectable ratio approaches a floor of roughly $z\sqrt{\phi/(n\,\ell)}$. Here $n$ is the number of particles, $\ell$ is the effective spread ("leverage") the design provides on the contrast, and $z$ is the threshold for the chosen significance level. This form comes from the variance of a ratio estimator under gamma-Poisson mixing. The second consequence is that when the design matrix is nearly rank-deficient, $\ell \to 0$. That happens when facet fractions co-vary with edge density across the chosen shapes, as they do when micron-scale nanobelts are compared with 15–20 nm nanocubes. In that case the contrast cannot be identified at any level of counting precision.
Grounding in Prior Work
The motivating evidence comes from facet-engineering studies that report deterministic rate ratios without any uncertainty model. Surface active sites on Co3O4 nanobelt and nanocube model catalysts for CO oxidation compared {011} nanobelts (2–5 μm long) with {001} nanocubes (15–20 nm). It reported turnover frequencies (reactions per site per second) of $7.4\times10^{-3}$ and $2.7\times10^{-3}\ \mathrm{s^{-1}}$ per surface Co³⁺, and explained the difference structurally: on {001} the Co³⁺ ions sit 1.008 Å below an oxygen layer. We do not dispute that mechanism. Our point is that a two-shape design with a roughly hundredfold size difference cannot, on its own, separate the facet effect from edge and size effects. Similar attribution narratives appear in Single-faceted IrO2 monolayer enabling high-performing proton exchange membrane water electrolysis beyond 10,000 h stability at 1.5 A cm-2, which isolates IrO₂(101) and credits its performance to a downshift of the d-band center, a common electronic descriptor of how strongly a metal binds reaction intermediates. They also appear in An Improved Understanding of Chalcopyrite Leaching Mechanisms: The Influence of Anisotropic Crystal Planes. Both are ensemble measurements, where the particle-variance term is invisible.
The method draws on two papers. From Limits on non-Gaussianities from WMAP data we borrow two ideas. The first is fitting the amplitudes of known, separable templates, which tightened the local bound from $[-58,134]$ to $[-27,121]$. The second is adding an explicit correction for spurious signal caused by masks and uneven noise. In our setting, edge and corner terms play the role of the mask, and facet-area fractions are the templates. From Artificial Neural Network-Based Prediction and Morphological Evolution of Cu2O Crystal Surface Energy we take its library of 4,423 Wulff-constructed Cu₂O shapes. A Wulff construction is the standard geometric rule that predicts a crystal's equilibrium shape from the surface energies of its faces. The library spans {100}, {111}, {110} and {211} and supplies realistic geometry matrices $A_{if}, L_i, C_i$. That paper predicts surface energy ratios (MAE 0.02–0.03, $R^2$ 0.96–0.98) and does not address catalytic rates; we use only its geometry.
The sensitivity lens comes from rare-event physics. Measurement of the proton-air cross section with Telescope Array's Middle Drum detector and surface array in hybrid mode shows a regime where statistical uncertainty dominates systematic uncertainty. INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT shows how searching many windows (trial factors) erodes an apparent hotspot. Facet studies that screen many shapes and facets face the same multiplicity problem, and we will report post-trial significance. A constraining note comes from Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset: better event classification raised signal efficiency from 80.8% to 96.4%. Better event detection, in other words, increases the counts per particle, but under our model it cannot lower the particle-variance floor. Only design and the number of particles can.
Research Program
We will build the estimator as a negative-binomial template regression, which arises when Poisson counts are mixed with gamma-distributed variation between particles. Facet, edge and corner rate constants are constrained to be non-negative. We will derive its Fisher information in closed form and compute the minimum detectable facet ratio as a function of three quantities: the number of particles, the observation time, and the shape library. On top of this we will build a design optimizer. Given a candidate library, such as the Cu₂O Wulff set or a Co₃O₄ cube, belt and octahedron set, and a budget of particle-hours, it chooses the shapes and allocates time to maximize the determinant of the information matrix restricted to the facet contrasts of interest. We will also derive a diagnostic for structural unidentifiability: the condition number of the facet-versus-edge block of the design matrix. Analysis code will read traces stored in the format of Single-molecule dataset (SMD): a generalized storage format for raw and processed single-molecule data. That format was designed for biophysics, and we will extend it with Miller-index and geometry fields.
Baselines
The naive reference point is the two-shape ensemble ratio used in practice. One divides the per-site turnover frequency of shape A by that of shape B and attributes the whole ratio to the dominant facets. The reported number for this baseline is the 2.7-fold Co³⁺-normalized ratio ($7.4\times10^{-3}$ vs $2.7\times10^{-3}\ \mathrm{s^{-1}}$) from Surface active sites on Co3O4 nanobelt and nanocube model catalysts for CO oxidation. It will be run on exactly the same simulated particles as our method, by summing their counts.
The strongest competitor is the textbook exact test for comparing two Poisson rates: a test on the shape-A count conditional on the total count, or equivalently a Poisson regression with shape as the only predictor. This is the method a careful statistician would use without our geometric model. It gets the same counts, the same particles and no extra tuning. We beat it if, at matched particle-hours, our estimator recovers the true facet ratio with nominal 90% interval coverage in scenarios where edges carry 10–50% of the activity, while the exact test's coverage collapses. We also beat it if our designs reach a given minimum detectable ratio with fewer particle-hours than the best two-shape design.
Two ablations isolate the key ingredients. Removing the edge and corner terms from the template tests whether the "mask correction" matters. Fixing $\phi=0$, a pure Poisson model, tests whether the particle-variance floor is real. Each ablation should produce overconfident intervals exactly in the regimes our theory predicts. On the computational side, no reported result was found on Lacuna for facet-resolved turnover attribution from single particles. The relevant prediction-error reference is GemNet-OC's 0.374 eV in-domain total-energy MAE on The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts.
Why not just…?
Why not just use bigger ensemble reactors so counting noise vanishes? Collecting $10^{20}$ product molecules removes the Poisson term but leaves the confound untouched. The total rate is still a sum over facets, edges and corners, so if those co-vary across shapes the facet contrast remains unidentified. The Co₃O₄ comparison shows the pattern: nanobelts and nanocubes have similar BET areas (20.1 vs 22.6 m²/g) yet differ in size by about a hundredfold, so their edge densities differ too. Our simulation will show directly that the ensemble ratio is biased whenever edges carry part of the activity, at any number of counts.
Why not just fit a standard Poisson or negative-binomial regression on shape labels? This is a real threat. Our estimator is itself a negative-binomial regression, so the novelty lies elsewhere. It lies in the geometric template covariates, the closed-form sensitivity floor, and the design optimizer. If a shape-label model with good shapes chosen by intuition matched our optimized designs on detectable ratio per particle-hour, the design contribution would be void. The matched-budget comparison is built to settle exactly that.
Why not just screen facets with pretrained graph neural networks trained on density functional theory data? Density functional theory (DFT) is the standard quantum-mechanical method for computing surface energetics, and graph neural networks (GNNs) are now trained on large DFT datasets to approximate it quickly. GemNet-OC's errors on OC22 are 0.374 eV in-domain and 0.829 eV out-of-domain. Its MAE against literature perovskite adsorption energies is 0.57 eV. The Open Catalyst 2020 (OC20) Dataset and Community Challenges puts the target accuracy at 0.02 eV. These errors are comparable to or larger than the facet differences in question, so such models can propose hypotheses but cannot validate them. Our framework can use their predictions as priors on $k_f$ without needing them to be accurate.
Why not just measure single-crystal surfaces in ultra-high vacuum? This is a strong alternative for clean planes, as the step-edge studies in Hydroxyl on Stepped Copper and its Interaction with Water illustrate. It does not, however, answer the question for real nanoparticles under reaction conditions, where edges and particle variability are the point. We treat single-crystal rates as external calibration that the nanoparticle estimator should reproduce.
Proposed Experiments
The first experiment is a simulation benchmark. Using the 4,423 Cu₂O Wulff geometries, we will simulate per-particle turnover counts from known rate constants. The base rates will be set to the Co₃O₄ turnover frequencies, edges will carry 0–50% of the activity, $\phi$ will range from 0 to 1, and counts per particle will range from 5 to 5,000. Metrics are bias, 90% interval coverage, and minimum detectable ratio at 3σ and 5σ, all at matched particle-hours across the naive ratio, the exact Poisson test, our estimator and its ablations.
The decisive check is whether the measured minimum detectable ratio flattens at the predicted floor $z\sqrt{\phi/(n\ell)}$ as counts per particle grow. A second check is whether our identifiability diagnostic flags the belt-versus-cube style design as non-identifiable while optimized three- or four-shape designs pass.
The second experiment is a reanalysis. We will rebuild approximate geometry matrices for the published Co₃O₄ shapes from their stated dimensions and ask how large an edge contribution would be needed to explain the whole 2.7-fold ratio. If a physically plausible edge rate suffices, the published attribution is underdetermined. That is a falsifiable statement about existing data. It does not show that the mechanism is wrong.
The third experiment is an optional wet-lab validation, which we flag as beyond the file's data: no single-particle facet turnover dataset was found on Lacuna. Single-particle fluorescence turnover counting on Cu₂O cubes, cuboctahedra and octahedra, allocated according to the optimizer, would test whether the predicted design gains appear in real measurements.
References
- Surface active sites on Co3O4 nanobelt and nanocube model catalysts for CO oxidation
- Single-faceted IrO2 monolayer enabling high-performing proton exchange membrane water electrolysis beyond 10,000 h stability at 1.5 A cm-2
- An Improved Understanding of Chalcopyrite Leaching Mechanisms: The Influence of Anisotropic Crystal Planes
- Limits on non-Gaussianities from WMAP data
- Artificial Neural Network-Based Prediction and Morphological Evolution of Cu2O Crystal Surface Energy
- Measurement of the proton-air cross section with Telescope Array’s Middle Drum detector and surface array in hybrid mode
- INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset
- Single-molecule dataset (SMD): a generalized storage format for raw and processed single-molecule data
- The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts
- The Open Catalyst 2020 (OC20) Dataset and Community Challenges
- Hydroxyl on Stepped Copper and its Interaction with Water