Transport-Compatible Interfaces in Graded Scaffolds for Competing Signals

Graded lattices are 3D-printed materials whose small repeating cells change in shape or density from place to place. They are designed today with one question in mind at the seams between neighbouring cells: do the solid struts connect, so that load passes across? For scaffolds that cells will live in, a second question matters just as much. Does the pore space connect, so that dissolved signaling molecules can diffuse across the seam? The two are different. A seam can be mechanically sound while its pore channels narrow or dead-end, and molecules then pile up on one side of it. When cells read two signals that suppress each other's pathways, a small local pile-up can decide which pathway wins in that region. A transport defect only a few cells wide then becomes a wrong biological outcome across a much larger area.

This note proposes to measure that effect and to design against it. We would build catalogs of 2D, and later 3D, microstructures whose faces share the same pore topology (where the openings are and how they connect) while the opening widths and the cell interiors stay free. We would then test whether this weaker compatibility rule keeps transport continuous across seams, keeps more of the range of achievable stiffness than fully identical boundaries, and lets the simplified homogenized model correctly predict the outcome of a two-signal competition. Homogenization is the standard shortcut that replaces each cell with its averaged material properties.

Core Claim

Our claim has two parts, each falsifiable. First, in graded scaffolds, the error of homogenized predictions for a mutually antagonistic two-signal readout is dominated by transport defects at cell interfaces. This error is larger, relative to the same defects, than the error for a single-signal readout, because antagonism acts like a threshold that amplifies small shifts in the ratio of the two concentrations. Second, requiring adjacent faces to share pore topology, rather than identical full geometry, removes most of that error while keeping a measurably larger property range than boundary-identical catalogs.

The biological motivation comes from two kinds of evidence in the file. Plant defense signaling resolves competing pathways through antagonism: silicon-primed potato raises ethylene and jasmonic acid while suppressing salicylic acid through NPR1, and silencing single components breaks resistance. Spatial confinement also decides signaling outcomes, as salivary gland morphogenesis shows. There, FGF8 rescues branching in culture by 24% but cannot replace FGF10 in vivo. We use these only to justify the type of readout, a bistable antagonistic switch. We do not fit to plant or gland data, and we make no in vivo claims.

Research Program

The design method builds on guided diffusion for boundary-identical microstructures. That work conditions a 2D diffusion generator, a model that produces shapes by progressively denoising random noise, on a fixed boundary mask so that every cell in a catalog shares an exact boundary. It reports an elastic-tensor fit above 0.95 R² and boundary match above 0.98. We change three things. We condition on a diffusivity tensor as well as the elasticity tensor. We replace the exact boundary mask with a face-topology template, which fixes the number, order and connectivity of the pore openings on each face but lets their widths vary within a tolerance. And we add guidance from a learned surrogate that predicts interface transport loss from a pair of face profiles.

The target quantity is an interface defect. For two adjacent cells A and B, we solve diffusion at full resolution through the two-cell pair to get a measured conductance $G_{AB}$. We compare it with the series conductance $G^{\mathrm{hom}}_{AB}$ that homogenization predicts from the averaged diffusivities $D_A$ and $D_B$ in the flux direction:

$$\delta_{AB} = 1 - G_{AB}/G^{\mathrm{hom}}_{AB}.$$

When $\delta_{AB}=0$, the seam is transport-transparent. Positive values mean pore channels pinch or dead-end at the seam. We compute averaged diffusivities with the same periodic cell-problem machinery already used for elasticity. This extends to transport the observation by Wang et al. that grading breaks the infinite-periodicity assumption.

The signaling readout is a standard mutual-inhibition switch driven by two ligands, $L_1$ and $L_2$, that enter from opposite sides of the scaffold and diffuse through the pore space. Local pathway activities $x$ and $y$ follow

$$\dot x = \frac{a\,L_1}{1+(y/K)^n} - x,\qquad \dot y = \frac{a\,L_2}{1+(x/K)^n} - y,$$

where $a$ is the activation gain, $K$ the inhibition threshold, and $n>1$ the steepness that makes the switch bistable. The readout is which pathway wins at each point. We report the fraction of the domain where the homogenized model and the full-resolution model agree on the winner, and the displacement of the boundary between the two winning regions. The single-pathway control removes the inhibition terms.

Baselines

The do-nothing reference tiles unconstrained cells from the initial 36,585-cell dataset of Feng et al. directly next to each other. This is the naive juxtaposition that Zhu et al. and Wang et al. associate with hanging nodes. It sets the ceiling on interface defects and on homogenization error.

The strongest competitor is the boundary-identical catalog of Feng et al. itself. Because faces are identical, pore openings line up by construction, so it should nearly eliminate $\delta$. Its weakness is a restricted property range. Its reported numbers are R² above 0.95 for elastic tensors and degree of match above 0.98. The file contains no reported result for transport or signaling on these catalogs, so we will compute both ourselves. To keep the comparison fair, we will regenerate its catalogs with diffusivity added to the conditioning, using the same network, sample count and guidance tuning budget as our method. Our method beats it only if two things hold together: the agreement on the winning pathway is not worse by more than a pre-registered margin of two percentage points, and the property range is measurably larger.

The blending family is a second strong competitor. It covers the latent interpolation with a compatibility gradient in MIND, which reports 0.29% property-matching error against 1.33% for an unconstrained diffusion baseline and 99.2% geometric validity. It also covers the implicit-surface transition layers in IH-GAN, which reports a 79.7% reduction in peak stress on a cantilever. These methods have the largest property range and were tuned for mechanics, so they test whether mechanically good seams are already transport-good. We will apply their blending rules to our 2D cells with transition layers of equal width.

Two ablations isolate the proposal's ingredients. The first keeps face-topology templates but drops the transport-surrogate guidance, which separates the effect of topology matching from the effect of width control. The second swaps the antagonistic readout for the single-pathway readout on every method, which tests whether antagonism amplifies interface error.

Why not just…?

Why not just use boundary-identical catalogs, since identical faces already align the pores? This is the most serious threat, and it may simply be enough. If the property range of boundary-identical catalogs turns out to be comparable to ours once diffusivity is part of the target, the proposal collapses into "use Feng et al. with a transport target." That would still be a useful finding, but it is not a new method. The property-range comparison at matched catalog size settles this directly. We commit to reporting it either way.

Why not just add diffusivity as a per-cell conditioning target and ignore the seams? Averaged diffusivity is computed for an infinitely repeating cell, so it says nothing about whether the openings of two different cells meet. The "blending without transport guidance" baseline and the topology-only ablation measure exactly how much $\delta$ remains when only cell-level targets are matched. If $\delta$ stays near zero in that setting, the interface surrogate is unnecessary and we will drop it.

Why not just use a uniform scaffold and soak it in the signals? For a uniform lattice, homogenization is close to exact, so there is no interface error at all. However, the scaffold can then neither be stiffness-graded nor place the competition boundary deliberately. Stollberg et al. report 35.2% higher stiffness for graded lattices than for uniform density on the 2D MBB beam at equal mass, a standard bending-beam benchmark. Jaskoll et al. show that bulk supply of a growth factor does not reproduce spatially localized outcomes. We include the uniform scaffold as a reference row to show the cost of giving up grading.

Proposed Experiments

The first experiment checks whether the problem exists. For each method, we generate 2D graded tilings that place stiffness and diffusivity gradients across the domain. We then compute three quantities: the distribution of $\delta$ over all interfaces, the agreement between homogenized and full-resolution winning-pathway maps, and the same agreement for the single-pathway readout. The decisive check is whether, across tilings, the antagonistic readout's disagreement rises with mean $\delta$ more steeply than the single-pathway readout's. If it does not, the claim that antagonism amplifies interface error is false, and the motivation for treating crosstalk specially goes away.

The second experiment is the trade-off. We plot the area of the achievable elasticity–diffusivity property region against mean $\delta$ and peak stress concentration at seams for every method. We also report compliance on the 2D MBB beam, where we include the uniform-density reference that Stollberg et al. use. Our method is useful only if it occupies a point that no baseline dominates.

The third experiment carries the comparison into 3D using the 180,000-cell MIND dataset. We add diffusivity tensors to its cells and compare face-topology templates against MIND's blending on the same assemblies. Feng et al. note that 3D boundary classification is an open problem, and topology templates may be a cheaper route through it. As a modest physical check, we would print a small set of 2D-extruded tilings and image dye fronts across the seams to confirm that the sign of the predicted $\delta$ holds. We will not run cell experiments until the in silico effect is established.