Synthetic Filtrations

• Synthetic filtrations are designed systems that integrate controlled architectures, including graded pore networks and measure-based indices, to modulate transport and retention.
• They combine deterministic and stochastic models, such as Hagen–Poiseuille-based flow and DTM-based topological filtrations, to address fouling and blockage challenges.
• These approaches enable tailored applications across porous media, topological data analysis, and categorical probability, offering actionable insights for optimizing filter performance.

Searching arXiv for the cited works to ground the article. arXiv search query: (Gu et al., 15 Aug 2025) Synthetic filtrations designate deliberately constructed filtration structures rather than filtrations taken as raw givens. Across the cited literature, the expression covers engineered filters whose pore geometry, transport laws, and fouling mechanisms are explicitly synthesized; measure-based filtrations in Topological Data Analysis that replace raw point-cloud distances by distance-to-measure weights; and geometric or categorical filtrations in which time itself is context-dependent rather than linearly indexed (Gu et al., 15 Aug 2025, Anai et al., 2018, Adachi, 16 Sep 2025). A plausible unifying description is that the filtration is made synthetic by inserting an additional designed structure—geometric, stochastic, weighted, or categorical—between the underlying system and the filtration it induces.

1. Terminological scope and common structure

In porous-media engineering, synthetic filtration refers to engineered filters whose microstructure and internal transport mechanisms are prescribed rather than merely phenomenologically averaged. The cited work includes mathematically engineered graded pore networks, porosity-graded homogenized media, PEG–PEGDA hydrogel membranes, corrugated nanochannels, branched-channel ricochet devices, and random or polydisperse fibrous media (Gu et al., 15 Aug 2025). In these settings, the synthetic element lies in the controlled architecture and in the explicit coupling of flow, transport, and retention.

In Topological Data Analysis, a DTM-based filtration is synthetic because it is built from a distance-to-measure function instead of raw pairwise distances. This replaces a filtration indexed directly by the geometry of a finite point cloud with one indexed by a smoothed, measure-based proxy that is robust to noise and outliers (Anai et al., 2018). In this sense, the filtration is synthetic because the indexing function is itself a constructed statistical object.

In categorical probability, synthetic filtrations are contravariant functors $X:\Sigma^{\mathrm{op}}\to \mathbf{Prob}$, where $\Sigma$ extends the simplex category $\Delta$ by objects encoding context-dependent times. Here the synthetic element lies in the time domain: the present can be formed by synthesizing multiple possible pasts, so information is no longer organized by a single linear chain of $\sigma$-algebras (Adachi, 16 Sep 2025). Singular Lie filtrations and their associated weightings provide a related geometric generalization: filtrations are imposed on the sheaf of vector fields by locally finitely generated submodules satisfying Lie-bracket compatibility, and these induce weighted normal bundles and osculating structures (Loizides et al., 2022).

2. Engineered pore networks and graded membrane filters

A central physical realization of synthetic filtration is the graded pore network model for membrane filters. In the stochastic sieving–adsorption model, the membrane is represented as a metric graph $G=(V,E,D)$ with cylindrical pore throats, and its thickness is partitioned into $m$ horizontal bands whose initial radii satisfy

$R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$

Within each band, porosity is constrained to remain close to a prescribed value $\phi_0$, typically $\phi_0\approx 0.6$, so that pore size and junction density are co-designed rather than independently assigned. Flow in each edge obeys the Hagen–Poiseuille law,

$$Q_{ij}=K_{ij}\big(P(v_i)-P(v_j)\big),\qquad K_{ij}=\frac{\pi R_{ij}^4}{8\mu L_{ij}},$$

and small foulants evolve by continuum advection–reaction with adsorption-driven radius shrinkage, while large particles arrive by a Poisson process and move through the network by a flux-weighted random walk until they either exit or block a pore (Gu et al., 15 Aug 2025).

This synthetic coupling of deterministic and stochastic mechanisms changes the qualitative behavior of filtration. Adsorption alone produces smooth, convex flux–throughput decay, but the addition of sieving generates kinks and sudden drops in the outlet flux, because complete blocking removes edges from the hydraulic graph in discrete events. The model also shows a phase transition in filter lifetime as the dimensionless sieving arrival rate $\Sigma$0 increases: below a critical threshold, lifetime is close to the adsorption-only value, while above it sieving sharply reduces lifetime. In the ungraded case, the critical $\Sigma$1 is reported as roughly $\Sigma$2 in the dimensionless units used (Gu et al., 15 Aug 2025).

An earlier adsorption-only graded-network model isolates the effect of pore-radius grading. There the membrane again consists of bands with

$\Sigma$3

equalized initial band porosity, and flow governed by the same Hagen–Poiseuille framework. For operation until flux extinction, total throughput $\Sigma$4 is maximized at $\Sigma$5, essentially independent of the maximum pore length $\Sigma$6; under a flux threshold $\Sigma$7, the maximizing gradient shifts to $\Sigma$8. By contrast, initial flux decreases monotonically with $\Sigma$9, while accumulated outlet contaminant concentration decreases monotonically with $\Delta$0 (Gu et al., 2022). The design tension is therefore explicit: stronger grading improves polishing but penalizes initial permeability.

3. Soft, deformable, and nanoscale realizations

Hydrogel membranes provide a distinct synthetic filtration platform in which selectivity is governed by a soft polymer network rather than rigid pores. In PEG–PEGDA membranes, free-standing hydrogel sheets are formed by UV-initiated photopolymerization of PEGDA, high-molecular-weight PEG, and Irgacure 2959 at a fixed PEGDA/water ratio of $\Delta$1. Pure PEGDA hydrogels exhibit a primary network with mesh size $\Delta$2 and a secondary population of water-filled cavities of diameter $\Delta$3–$\Delta$4, while PEG-containing samples display micron-scale cavities and, in the $\Delta$5 PEG sample, nanometric heterogeneities of order $\Delta$6 in the PEGDA-rich walls (Eddine et al., 2023). Functionally, $\Delta$7 particles permeate only in the PEG-containing samples and only above a pressure threshold of about $\Delta$8, whereas $\Delta$9 and $\sigma$0 particles do not permeate despite the presence of $\sigma$1 cavities. The work therefore argues that the effective transport pathways are local $\sigma$2–$\sigma$3 defects in the cavity walls, not the cavities themselves (Eddine et al., 2023).

At the nanoscale, corrugated nanochannels act as synthetic nanofilters through an entropic rather than steric mechanism. The studied device alternates narrow slits and deeper pits under pressure-driven flow with full hydrodynamic interactions and thermal fluctuations. When pit and polymer sizes are compatible and the pressure difference is sufficiently small, polymers exhibit contour-length-dependent entropic trapping in the pit, due to competition between advection and chain relaxation. This produces a non-monotonic dependence of mobility on size. Circular polymers move faster than linear polymers of the same $\sigma$4 because hairpin entry places two strands in the slit and effectively doubles the driving force, while star polymers often collapse onto the same mobility curve as linear polymers when plotted against $\sigma$5, indicating that effective size dominates in much of the parameter range (Ollila et al., 2014).

Cross-flow ultrafiltration of non-ionic microgels offers a third soft-matter realization. Here the macroscopic model combines Darcy permeation,

$\sigma$6

with concentration-dependent collective diffusion and viscosity for solvent-permeable spheres. Microgel permeability enters through the hydrodynamic radius $\sigma$7, increases $\sigma$8, lowers $\sigma$9, reduces membrane-surface concentration, and enlarges the permeate velocity relative to impermeable hard spheres (Roa et al., 2015). The synthetic character lies in the fact that particle permeability, membrane resistance, osmotic pressure, and concentration polarization are co-modeled as design variables rather than absorbed into a single empirical resistance.

4. Multiscale transport, heterogeneity, and anomalous retention

Synthetic filtrations in porous media are often designed through multiscale upscaling. In the homogenization theory for porosity-graded filters, the filter is modeled as a near-periodic array of spherical obstacles whose radius varies slowly in space, so that porosity becomes a macroscale field $G=(V,E,D)$0. Homogenization yields a Darcy law for the flow and an effective advection–diffusion–reaction equation for the volume-averaged solute concentration,

$G=(V,E,D)$1

A decreasing porosity with depth has a much larger effect on the uniformity of adsorption than on total adsorption, and this is the mechanism by which it lowers the risk of localized blocking while maintaining the rate of total contaminant removal (Dalwadi et al., 2015).

A related multiscale framework for fibrous filters begins from fully resolved Stokes flow and contaminant transport in random or regular two-dimensional fibre cross-sections, then computes effective permeability $G=(V,E,D)$2, diffusivity $G=(V,E,D)$3, and surface area $G=(V,E,D)$4. Because fibres grow by adsorption, the model includes an agglomeration algorithm that merges neighbouring fibres upon contact and then recomputes the effective coefficients. The influence of microstructure is pronounced for permeability and modest for diffusivity: at $G=(V,E,D)$5, the ratio $G=(V,E,D)$6 is about $G=(V,E,D)$7, whereas at $G=(V,E,D)$8 it is about $G=(V,E,D)$9; by contrast, the different random microstructures have almost identical mean $m$0 (Printsypar et al., 2018). This separates two roles often conflated in empirical models: microstructure primarily reshapes advective access and specific surface area, not merely diffusion.

The intermittent-transport study of colloid filtration makes the same point stochastically. In a one-meter microfluidic porous medium with controlled heterogeneity, colloids alternate between long-range “flights” through pore channels and localized “dives” near grain surfaces. Attachment occurs at constant rate during dives, but not during flights, so a distributed flight-length law creates anomalous filtration. The measured breakthrough curve exhibits a power-law tail $m$1 after one pore volume, while the deposition profile is power-law near the inlet and exponentially truncated downstream. An alternated CTRW constrained by the measured pore and grain statistics captures both signatures and links the cutoff length to geometry, critical distance, and attachment rate (Miele et al., 5 May 2025).

A high-Reynolds-number branched-channel filter furnishes a different multiscale limit. There, multiple-scales analysis replaces a discrete family of thin branches by an effective leakage boundary condition on the lower wall of the main channel,

$m$2

so that the total leakage satisfies $m$3. Coupled with a particle model using a wall-bounce condition, this yields a design-level relation between leakage, Stokes number, and the fraction of particles diverted into branches in ricochet separation (Fastnedge et al., 7 Nov 2025).

5. Measure-based, weighted, and categorical filtrations

In Topological Data Analysis, DTM-based filtrations are synthetic because they replace raw point-cloud distance with the distance-to-measure function

$m$4

For an empirical measure $m$5 with $m$6, this reduces to the average squared distance to the $m$7-nearest neighbors. The resulting weighted Čech filtration $m$8 and its weighted Rips approximation are robust to outliers because DTM is $m$9-Lipschitz and satisfies

$R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$0

The circle-plus-outliers example shows that the DTM-filtration of noisy data recovers the persistent $R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$1-dimensional class of the clean circle, whereas the ordinary Čech filtration on the contaminated sample does not (Anai et al., 2018).

Singular Lie filtrations generalize regular filtered manifolds by replacing filtrations of the tangent bundle with filtrations of the sheaf of smooth vector fields,

$R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$2

where each $R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$3 is locally finitely generated and $R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$4. For an $R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$5-clean submanifold $R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$6, the induced filtration on functions defines a weighting of order $R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$7, and under constancy assumptions the weighted normal bundle is identified with a homogeneous space of osculating groupoids,

$R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$8

This makes the filtration synthetic in the sense that quasi-homogeneous structure is reconstructed from filtered vector-field data rather than from ordinary tangent directions (Loizides et al., 2022).

The categorical theory pushes this construction to the level of time itself. Contexts form

$R_n = R_m + (m-n)sW,\qquad 1\le n\le m.$9

and an object of the context-dependent time category $\phi_0$0 is a time $\phi_0$1 together with an equivalence class of contexts. A synthetic filtration is then a contravariant functor $\phi_0$2. Dirichlet filtrations are concrete examples in which each object carries a probability measure on a simplex induced by a Dirichlet distribution, and Bayesian updating is interpreted as a categorical update of the Dirichlet functor. The framework is designed to model both parameter uncertainty and contextual uncertainty in situations where the present is synthesized from multiple possible pasts (Adachi, 16 Sep 2025).

6. Design principles, misconceptions, and limitations

Several design principles recur across the physical literature. Significant pore-radius gradients are advantageous when the feed contains a mixture of small adsorptive foulants and sporadic large particles; placing smaller pores deeper localizes sieving and preserves upstream adsorption, while very moderate gradients can perform poorly when large particles arrive frequently (Gu et al., 15 Aug 2025). Decreasing porosity with depth improves adsorption uniformity more than total adsorption and therefore lowers the risk of localized blocking without sacrificing overall removal (Dalwadi et al., 2015). For soft colloids, permeability-aware transport coefficients are essential: permeable microgels weaken concentration polarization and increase flux relative to impermeable particles under the same operating conditions (Roa et al., 2015).

A persistent misconception in soft filtration is that functional pore size is directly readable from static imaging. The PEG–PEGDA hydrogel study explicitly argues the opposite: micron-scale cavities coexist with a functional pore size controlled by $\phi_0$3 wall defects, and the pressure threshold for $\phi_0$4 particle permeation reflects elastic pore deformation rather than a rigid cavity aperture (Eddine et al., 2023). A parallel misconception in porous-media theory is that constant-rate, homogeneous filtration laws are structurally adequate. The graded-network and intermittent-transport studies show that discrete blockage, cooperative interior sieving, broad flight statistics, and power-law deposition profiles are inaccessible to such laws, even when mean flux decline can still be represented empirically (Gu et al., 15 Aug 2025, Miele et al., 5 May 2025).

The mathematical variants carry their own limitations. DTM-filtrations require tuning of the mass parameter $\phi_0$5 and the exponent $\phi_0$6, and the strongest robustness results occur for $\phi_0$7; the paper also constructs a counterexample showing that no uniform global bound of the form $\phi_0$8 can hold for all measures (Anai et al., 2018). The categorical model of context-dependent time is currently discrete, and its authors identify continuous-time extension as an open direction (Adachi, 16 Sep 2025). Singular Lie filtrations require $\phi_0$9-clean submanifolds and additional constancy hypotheses to assemble the fiberwise graded Lie algebras into global algebroids and groupoids (Loizides et al., 2022).

Across these literatures, synthetic filtrations remain a plural concept rather than a single theory. In engineered media, the phrase denotes co-designed architectures and transport mechanisms; in topology, it denotes robustified weighted filtrations; and in geometry and probability, it denotes filtrations indexed by weighted or context-dependent structures. What unifies these uses is the replacement of an unstructured filtration by an explicitly constructed one whose architecture is intended to control robustness, selectivity, or inferential content.