Which Parts of a Porous Biochar Model Does the Scattering Data Actually Determine?

Materials made from crop waste, such as nitrogen-doped biochar from fruit pits or amorphous silica from corn stover ash, are disordered and full of pores. The standard way to get atomic structure for such materials is "big-box" refinement. One simulates a box of tens of thousands of atoms and moves them until the box's computed pair distribution function matches the measured one. The pair distribution function is a histogram of interatomic distances, obtained from X-ray or neutron scattering. Refinement returns one configuration, and papers then report ring statistics, pore topology, and dopant environments read off that single box. We propose to test whether these numbers come from the data or from the forcefield and optimizer that filled in whatever the data did not constrain. Our claim is that porous matter has a specific, predictable way of being underdetermined. Two dense regions separated by a void wider than the distance range the data covers contribute no cross-distances. So the data cannot tell where one region sits or how it is oriented relative to the other.

This is the structural counterpart of a known result in phase retrieval, the problem of recovering a signal from measurement magnitudes alone. There, a signal made of energy clusters separated by near-empty "lagoons" leaves the relative phase between clusters undetermined, and stability collapses (Stable Phase Retrieval in Infinite Dimensions). The fix in that literature is not to force the signal to be connected. Instead, one changes what counts as the answer: recover each connected "atoll" up to its own phase. We carry both halves over to refinement. First, a graph built from which atom pairs the data can see predicts which structural descriptors are stable. Second, refinement should report structure up to independent rigid motions of weakly coupled regions. It should state explicitly that descriptors spanning a void are prior-determined.

Core Claim

The objective being minimized has two parts: the weighted squared misfit between observed and model pair distribution functions, plus whatever forcefield term the method adds. We claim that the data-only part has near-flat directions, and that these directions are rigid translations and rotations of atom clusters joined by few data-visible pairs. The argument is short. Take any set $A$ of atoms and translate it rigidly by a unit vector. The model pair distribution function changes only through pairs with one atom in $A$ and one outside. Define $\kappa_{ij}$ as the sensitivity of the fitted pair distribution function to the length of pair $(i,j)$, computed from the weighting, the fitted distance window, and the instrumental damping. $\kappa_{ij}$ is zero for pairs longer than the fitted maximum distance $r_{\max}$. The curvature of the data misfit along this motion, per unit displacement norm, is then at most a cut ratio:

$$ h \;=\; \min_{|A| \le N/2} \frac{\Phi(A)}{|A|}, \qquad \Phi(A) = \sum_{i \in A,\, j \notin A} \kappa_{ij}, $$

where $N$ is the number of atoms. The data-only error amplification is therefore at least of order $1/\sqrt{h}$. This is a Cheeger-type quantity on what we call the pair-sensitivity graph: atoms are vertices and edges are weighted by $\kappa_{ij}$. It plays the role that the Cheeger constant and the algebraic connectivity $\lambda_2$ (the second-smallest eigenvalue of the graph Laplacian) play in the Locally Stable and Conditionally Connected (LSCC) theory of Stable Phase Retrieval from Locally Stable and Conditionally Connected Measurements.

The lower bound (a weak cut implies instability) is elementary. The substantive conjecture is the LSCC-style converse. Suppose each cluster is locally well determined, and suppose the coupling between clusters, restricted to their rigid motions, forms a well-connected "connection Laplacian" (a Laplacian whose edges carry small coupling matrices rather than scalars, as in phase synchronization). Then the data determine the whole structure, up to a global rigid motion, with a constant polynomial in the number of clusters rather than exponential. If this holds, it lets a practitioner certify, from a single refined model, which reported numbers are trustworthy.

Research Program

The first component is theory. We will state and prove the cut lower bound for the orientation-averaged pair distribution function with finite momentum range and noise. We will then try to prove the converse under an explicit local-stability assumption on each cluster, following the structure of the LSCC argument. Two features make this more than a transcription. The symmetry group per cluster is three-dimensional rigid motion rather than a unit phase. And the forward operator has none of the holomorphic structure that Stable Phase Retrieval in Infinite Dimensions exploits. That paper found that Gabor measurements degrade exponentially with lagoon width while multi-scale wavelets degrade only polynomially. We expect a comparable dependence here on how fast $\kappa_{ij}$ decays with distance, which is set by momentum-range damping and by $r_{\max}$. We will derive that dependence rather than assume it.

The second component is an algorithm. Given a refined box, we compute $\kappa_{ij}$ from the analytic gradient already available in differentiable refinement (Fitting Pair Distribution Function with Backpropagation). That paper treats the pair distribution function as a differentiable layer and fits tens of thousands of coordinates by gradient descent with a minimum-distance constraint. We reuse its machinery and do not change its objective. We then partition the pair-sensitivity graph into atolls by spectral clustering, and compute the smallest eigenvalues of the reduced rigid-motion coupling matrix. For every reported descriptor, we then ask whether it is invariant under independent rigid motions of the atolls. Descriptors that pass we call data-determined; those that fail we call prior-determined. Ring statistics and the $Q_n$ connectivity of silica tetrahedra (the number of bridging oxygens per silicon) are usually local and should pass. Inter-domain pore topology and long-range density correlations are likely to fail. Refinement can also gauge-fix each atoll, pinning its rigid motion, so that prior-determined degrees of freedom stop consuming optimizer effort.

The third component is design. The graph tells us which extra measurement would create edges across a weak cut and which would add only redundant edges within an atoll. Phase Retrieval In The General Setting Of Continuous Frames For Banach Spaces proves that redundancy alone cannot restore stability. Our framework makes a sharper, testable prediction. Extending $r_{\max}$ or adding small-angle scattering should help exactly when it creates cross-cut edges with non-negligible $\kappa_{ij}$. Raising the maximum momentum transfer should mainly help within atolls.

Agricultural waste is the application where this matters and where it is testable. Two of the file's materials serve here. The first is the jujube-pit biochar series of Two Birds with One Stone, where NBC-8 has 223 m²/g surface area and NBC-12 collapses to 31 m²/g. The second is the corn stover ash silica of Synthesis of silsesquioxanes... using corn stover ash, whose authors flag uncharacterized impurity and structural variation. The biochar series is a natural contrast built into the chemistry. Pore collapse should make the pair-sensitivity graph better connected, so NBC-12 should yield fewer prior-determined descriptors than NBC-8. For valorization, the practical question is whether the pyrrolic-nitrogen environments that the DFT calculations in the biochar paper tie to formaldehyde uptake are data-determined. They are local, so we expect they are. If so, papers can report dopant-site statistics with confidence while withholding claims about global pore architecture.

Baselines

The naive reference point is standard molecular dynamics with an empirical forcefield and no data fitting. On sodium silicate glass, New insights into the structure of sodium silicate glasses by force-enhanced atomic refinement reports a Wright's $R_\chi$ of about 16% for this approach. The strongest existing refinement method is Force-Enhanced Atomic Refinement (FEAR) from the same paper. FEAR alternates 5,000 Reverse Monte Carlo steps with Teter-forcefield energy minimization and reaches $R_\chi$ = 2.49% on $(Na_2O)_{30}(SiO_2)_{70}$ neutron data. It is the strongest competitor because its forcefield term already fills in unconstrained directions and yields low-energy structures. So any stability claim must beat it on descriptor accuracy, not on fit. The second strong competitor is differentiable pair distribution function fitting (Fitting Pair Distribution Function with Backpropagation). It reports convergence within 20 minutes on $10\times10\times10$ BaTiO₃ supercells, but the file lists no $R_\chi$ value; there is no reported fit-quality result found on Lacuna for it on silicate glasses. For biochar or biogenic silica, no reported refinement result was found on Lacuna for any method.

All methods will receive identical simulated or measured data and identical box sizes, momentum ranges, and $r_{\max}$. Each gets the same wall-clock budget per refinement and the same number of random restarts. Because our method is primarily a diagnostic rather than a new fitter, "beating" a baseline has two meanings. First, gauge-fixed refinement must match FEAR's $R_\chi$ within noise, and on synthetic ground truth it must reduce error in data-determined descriptors. Second, the diagnostic must predict which descriptors each baseline gets wrong.

Two ablations isolate the key ingredient. The first replaces the sensitivity-weighted graph with the ordinary chemical-bond coordination graph. That graph is connected through thin bridges even when data coupling is weak, so if it predicts instability equally well, our measurement-graph idea is unnecessary. The second replaces the graph partition with a raw Lanczos computation of the smallest Gauss–Newton eigenvalues. This checks whether the graph adds anything beyond interpretability and the ability to gauge-fix.

Why not just…?

Why not just run many random-seed refinements and read off the spread of each descriptor? This is the most serious alternative, and it may identify unstable descriptors about as well as our diagnostic does. Its weakness is that seed spread mixes two causes. One is the optimizer trapping that Fitting Pair Distribution Function with Backpropagation documents for sequential Monte Carlo. The other is genuine non-identifiability. With a strong forcefield, seed spread can also be small for descriptors the data never saw. We will run exactly this ensemble as the reference and measure whether the single-run graph prediction ranks descriptors the same way. Suppose it does, but adds nothing on the forcefield-masked cases or on predicting which extra measurement helps. Then we will concede that the ensemble suffices and reposition the contribution as theory and measurement design.

Why not just add a connectivity penalty on $\lambda_2$ to the refinement loss? This is the obvious combination of the four directions, and we think it is wrong for porous matter. The pores are real, and a penalty that rewards connectivity would push the model to close them. The lagoon analysis itself chose to redefine the answer up to local phase rather than regularize the signal toward connectedness (Stable Phase Retrieval in Infinite Dimensions). We will include the penalty as a baseline on synthetic porous boxes with known pore widths. We expect it to lower apparent variance while biasing pore-size and density descriptors away from the truth.

Why not just use a better forcefield, for example a machine-learned potential, inside FEAR? A better forcefield is complementary, not a competitor. It improves what fills the unconstrained directions but does not make them data-determined. The FEAR authors themselves note that the result is bound to the fidelity of the Teter potential. Our diagnostic is exactly what reveals which reported quantities rest on that potential. To show this, we will refine the same data under two different forcefields. The prediction is that data-determined descriptors agree between them, while prior-determined ones diverge.

Why not just rely on surface area and elemental analysis? The NBC-8 versus NBC-12 comparison shows that bulk metrics move in ways their chemistry alone does not explain: nitrogen content goes up while area and uptake go down. Bulk metrics cannot say which local sites exist. Our proposal does not replace BET measurements. It asks which atomic-scale claims layered on top of them are warranted.

Proposed Experiments

The first check uses synthetic ground truth. We start from sodium silicate glass boxes generated in the FEAR setting and carve spherical and slit pores of controlled width $w$ while keeping thin silica bridges. From these we simulate pair distribution functions with realistic noise, finite momentum range, and several $r_{\max}$ values, then refine with molecular dynamics, FEAR, differentiable fitting, the penalty variant, and our gauge-fixed variant. For each run we record the error of each descriptor against the true box, its seed spread, and our predicted stability.

The decisive test has two parts. First, the error of cross-pore descriptors should rise sharply once $w$ exceeds the effective range of $\kappa_{ij}$, while ring and $Q_n$ statistics remain accurate. Second, the Spearman correlation between predicted instability ($1/\sqrt{h}$ at the atoll level) and observed error should be substantially higher for the sensitivity-weighted graph than for the bond-graph ablation. We will fit how error scales with $w$ and compare the fitted scaling to the theoretically derived dependence on $\kappa$ decay.

The second check tests design. On the same boxes, we compare three additions against the prediction: a larger $r_{\max}$, simulated small-angle scattering, and a larger maximum momentum transfer. The theory predicts that the first two reduce cross-pore error and the third mostly does not. Seeing the opposite pattern would falsify the graph account.

The third check uses real materials. We will prepare NBC-8 and NBC-12 following the protocol of Two Birds with One Stone and silica from corn stover ash following Synthesis of silsesquioxanes... using corn stover ash, then collect X-ray total scattering. We refine under two forcefields and report which descriptors are data-determined. The pre-registered predictions are two. NBC-12 should have a larger cut ratio $h$ and fewer prior-determined descriptors than NBC-8. And nitrogen-site environment statistics should agree across forcefields in both materials, while inter-domain structure agrees only in NBC-12. Either outcome is informative: confirmation gives valorization studies a principled rule for what atomic-scale claims to publish, and failure would show that porous-carbon refinement is limited by something other than data connectivity.