Rafał Topolnicki, Michał Bogdan, Jakub Malinowski, Bartosz Naskręcki, Maciej Harańczyk, Paweł Dłotko
7 min
Abstract
Classical topological descriptors used in topological data analysis (TDA) are invariant under permutations of spatial axes and therefore cannot represent the loading direction, which is essential for modeling anisotropic mechanical response. Here, this limitation is addressed by introducing a \emph{direction-aware TDA framework} in which the compression axis is explicitly embedded into filtration functions used to compute both persistent homology and Euler characteristic profile descriptors. Across multiple porous-material datasets spanning a broad range of structural anisotropy, direction-aware descriptors yield higher predictive accuracy than their direction-agnostic counterparts, with performance gains that increase systematically with anisotropy. Notably, direction-aware descriptors remain competitive and often improve $R^2$ even for nominally isotropic ensembles, indicating sensitivity to mechanically relevant directional organization beyond bulk anisotropy measures. When used as inputs to gradient-boosted tree models, the proposed descriptors approach the accuracy of convolutional neural networks trained directly on voxelized structures while retaining a compact, transferable representation. The study considers multiple datasets spanning weak to strong anisotropy, enabling systematic validation of direction-aware topology across regimes. Overall, the results establish direction-aware TDA as a general route for linking porous structure to direction-dependent elastic properties and motivate its use in anisotropic materials modeling problems where a preferred direction naturally arises.
Alex: So these RTP ones ramp up the direction effect on purpose. What about the others?
Sam: A second dataset has topologically diverse structures—various pore shapes like Voronoi foams from random seeds or strut networks from lattices—all statistically isotropic, meaning no built-in direction bias. For a middle ground, they took those isotropic ones, downsampled, elongated the z-slice like pulling taffy to double its length, then tiled back to a cube; this adds controlled anisotropy without changing local connections much.
Alex: Right, so from balanced shapes to stretched ones. But how exactly do they bake the squeeze direction into the analysis?
Sam: The core is a cone-based filtration: at each tiny cube, or voxel, in the material, they draw two opposing cones along the compression axis—like narrow funnels pointing forward and back, wrapping around periodically so edges connect seamlessly. The filtration value there is simply the fraction of empty space inside those cones; as they grow the cones step-by-step, it reveals how pores connect or block specifically along that force path.
Alex: Huh, so the cone forces the math to 'look' straight along the load, spotting aligned voids that affect stiffness.
Sam: Yes—and for extra nuance, they add principal component multifiltration: around each voxel, they scan the local points, find the main stretch direction via principal component analysis—like picking the longest axis in a cloud of dots—then weight cones by how well it matches the load axis. They apply these filtrations at different scales across the grid, then summarize the changing shapes with tools like persistent homology—which tracks how connections, loops, and voids form and fade—and Euler characteristic profiles, which count them in a simpler curve. These become compact vectors fed into models like CatBoost, a type of boosted decision tree.
Alex: Okay, so interpretable summaries versus black-box image crunching. But with careful splits to avoid cheating on the tests?
Sam: Yes—eight-fold cross-validation at the structure level ensures no leakage, meaning all directions from one material stay together in train, validate, or test sets. Results show directional descriptors match CNN accuracy closely overall, but excel where anisotropy matters most—like in strongly anisotropic RTPxz sets, where non-directional predictions scatter widely because they miss how voids align sharply along the hard axis, but directional cones capture that channeling with much tighter errors.
Alex: So the gains get bigger as the material's direction-dependence ramps up?
Sam: Exactly. The pattern repeats across structure types: in balanced cases like the topologically diverse set, both perform similarly; stretch to moderate anisotropy, and directional pulls ahead. Benefits grow systematically with anisotropy strength. Combining persistent homology and Euler profiles produces the most accurate and stable results, smoothing out weaknesses in either one separately.
Alex: Huh, so it narrows where direction matters most. From a modeling standpoint, what does that mean for designing these materials?
Sam: It offers a general way to encode how voids connect and align directionally—crucial for stiffness—without custom geometric measures or heavy training on full 3D grids. These summaries work across structure types, no tweaks needed. Broader, it could extend to other direction-tied traits like flow through pores or full elasticity maps. The gains hold systematically, but the study notes they're clearest where anisotropy dominates; in near-isotropic spots, benefits are smaller but never a drawback. All code and data are open on GitHub.
Alex: So a solid, interpretable tool that scales with real-world direction effects in porous stuff. Makes sense for implants or storage where squeeze direction rules. Thanks for breaking it down, Sam.
Sam: My pleasure, Alex. This work shows how embedding direction sharpens topology for anisotropic challenges. Thanks for listening to ResearchPod.