Double-Fractal Architecture

• Double-fractal architecture is defined by two coupled, self-similar processes that create distinct micro and macro organizational scales.
• It is applied to model phenomena such as superconducting networks in FM/SC bilayers, rheological behavior in colloidal suspensions, and recursive computational architectures.
• Key insights include quantitative relationships between fractal dimensions, percolation thresholds, and dual-scale dynamics that enhance materials and computational performance.

Double-fractal architecture denotes an organization in which two coupled fractal hierarchies, or two independent fractal-generating processes, jointly determine structure, dynamics, or function. In the literature represented here, the term is explicit in some domains and interpretive in others. It is used directly for superconducting pathways in ferromagnetic/superconducting bilayers and for rheological modeling of attractive colloids, while several computational, neural, and architectural studies describe closely related two-level self-similar constructions without always adopting the same label (Ruiz-Valdepeñas et al., 2013, Bouthier et al., 2022, Fu et al., 10 Nov 2025, Moharil et al., 18 Mar 2025). The common feature is not merely self-similarity, but a two-tier organization in which a lower-scale fractal process and a higher-scale fractal process remain structurally distinct and jointly control observable behavior.

1. Conceptual definition and terminological scope

A single fractal description assigns one dominant self-similar rule or one effective fractal dimension to an object. A double-fractal architecture instead separates two levels of organization. In the strongest form, the two levels are explicit and quantitatively distinct, as in particle-within-cluster versus cluster-within-network descriptions, or nucleus placement versus cluster growth in percolative superconductivity (Bouthier et al., 2022, Ruiz-Valdepeñas et al., 2013). In weaker or extrapolated forms, the phrase denotes two nested recursions, such as macro- and micro-level recursive segmentation in neural or hardware systems (Moharil et al., 18 Mar 2025, Fu et al., 10 Nov 2025).

Domain	Lower-scale fractality	Higher-scale fractality
FM/SC bilayers	Fractal expansion of reversed-domain clusters	Random arrangement of initial reversed domains
Attractive colloids	Particles within clusters	Clusters within a percolated network
FractalCloud	Shape-aware binary-tree decomposition	Recursive compute/memory orchestration

This usage should be distinguished from generic multifractality. In fractal scale-free networks, the relevant notion is often bifractality: only two local fractal dimensions, $d_f^{\min}$ and $d_f^{\max}$, are needed to characterize structural inhomogeneity (Yamamoto et al., 2023). It should also be distinguished from anisotropic self-similarity. The Borobudur study reports a single box-counting dimension for the whole 3D form together with different horizontal and vertical scaling exponents, and explicitly states that this is better described as anisotropic self-similarity than as a demonstrated “double-fractal” architecture (Situngkir, 2015).

2. Double percolation and superconducting pathways

The clearest physical realization is the NdCo$_5$/Nb PMA ferromagnetic/superconducting bilayer, where superconductivity nucleates on a tunable fractal network shaped by magnetic history (Ruiz-Valdepeñas et al., 2013). The two controlling processes are independent: first, a random arrangement of initial reversed magnetic domains establishes the coarse map of nascent superconducting islands; second, each reversed-domain cluster expands fractally, with mass dimension $D_m$, determining how the effective superconducting area grows. This is why the paper characterizes the system as a double percolation and a dual-scale fractal architecture.

The central relation is that the superconducting area fraction is not identical to the reversed magnetic area fraction. With

$p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$

and with fractal cluster growth,

$$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$

Using this form, the resistance near percolation follows

$R \sim (p_c-p_{\mathrm{sup}})^s,$

with $s \approx 1.32$ and $D_m \approx 1.8$. The percolation threshold is history dependent: $p_c \approx 0.16$ for minor loop mL1, $d_f^{\max}$0 for mL2, and $d_f^{\max}$1 in the major loop. These values support the interpretation that two random processes, rather than one, govern the onset of global superconductivity.

The second aspect of the architecture is stray-field confinement. The NdCo layer supports labyrinth domains, and the Nb layer experiences a spatially modulated out-of-plane field. At remanence, the average negative field over the “down” domains is $d_f^{\max}$2 Oe, strengthening to about $d_f^{\max}$3 kOe by $d_f^{\max}$4 kOe. In ordered stripes, the nucleation condition is expressed through the localization length

$d_f^{\max}$5

with a crossover at $d_f^{\max}$6 kOe, where $d_f^{\max}$7 nm and $d_f^{\max}$8 nm become comparable to $d_f^{\max}$9 nm. In labyrinth configurations, once percolation is complete, the upper critical field follows

$_5$0

a non-integer exponent consistent with superconductivity confined on a fractal network rather than a homogeneous 2D film or a purely 1D wire network.

A common misconception is that the double-fractal character here is simply a decorative description of domain morphology. The measurements show otherwise: the transport plateaus, the history dependence of $_5$1, the inferred $_5$2, and the anomalous $_5$3 exponent are all tied to the two-level control of superconducting connectivity.

3. Colloidal suspensions and two-level rheological scaling

In attractive colloidal suspensions, double-fractal architecture is introduced to relate rheological observables explicitly to microstructure across two scales (Bouthier et al., 2022). The first level is intra-cluster: particles of radius $_5$4 assemble into clusters of size $_5$5 with fractal dimension $_5$6 and chemical dimension $_5$7. The second is inter-cluster: clusters of size $_5$8 assemble into a sample-spanning network of size $_5$9 with fractal dimension $D_m$0 and chemical dimension $D_m$1. The model’s defining claim is that rheology depends not only on particle volume fraction $D_m$2 but also on cluster size $D_m$3 and on possible mismatch between $D_m$4 and $D_m$5.

The two-level packing relation is

$D_m$6

which already shows that the network-scale and cluster-scale fractal dimensions enter differently. The elastic model further introduces $D_m$7 as attraction energy scale, $D_m$8 as interaction range, $D_m$9 as bending-stretching mixing parameter, and $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$0 as strong-link versus weak-link parameter, with

$p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$1

In the exact multi-step extension, $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$2, $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$3, and $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$4 are written as products over hierarchical levels; the two-scale specialization makes both $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$5-exponents and $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$6-exponents depend separately on $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$7 and $p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$8.

For percolated systems, the closed-form storage modulus is

$p_{\mathrm{mag}}(H_z)=0.5[1-M(H_z)/M_s],$9

with analogous expressions for $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$0 and $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$1. A decisive structural consequence is that when $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$2, the yield stress loses any dependence on $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$3, and more generally cluster-size influence is suppressed. The paper therefore treats heterogeneity between cluster and network fractal dimensions as necessary for explaining experiments in which $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$4 changes strongly at fixed $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$5 when cluster size changes.

The model also predicts a two-slope small-angle scattering signature through an intensity $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$6 containing both $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$7 and $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$8 sectors, allowing $$p_{\mathrm{sup}} \propto r_0^2 \propto (p_{\mathrm{mag}})^{2/D_m}.$$9, $R \sim (p_c-p_{\mathrm{sup}})^s,$0, $R \sim (p_c-p_{\mathrm{sup}})^s,$1, and $R \sim (p_c-p_{\mathrm{sup}})^s,$2 to be inferred from Guinier knees and fractal slopes. In a comparison to carbon black gels under ultrasound, representative parameters $R \sim (p_c-p_{\mathrm{sup}})^s,$3, $R \sim (p_c-p_{\mathrm{sup}})^s,$4 nm, $R \sim (p_c-p_{\mathrm{sup}})^s,$5, $R \sim (p_c-p_{\mathrm{sup}})^s,$6 nm, $R \sim (p_c-p_{\mathrm{sup}})^s,$7 \AA, $R \sim (p_c-p_{\mathrm{sup}})^s,$8, $R \sim (p_c-p_{\mathrm{sup}})^s,$9, $s \approx 1.32$0, $s \approx 1.32$1, $s \approx 1.32$2, $s \approx 1.32$3, and $s \approx 1.32$4 produced $s \approx 1.32$5 kPa, $s \approx 1.32$6, and $s \approx 1.32$7 Pa against experimental targets of about $s \approx 1.32$8 kPa, $s \approx 1.32$9, and $D_m \approx 1.8$0 Pa. The paper characterizes this as close order-of-magnitude agreement rather than exact matching.

4. Mathematical abstractions: bifractality, micro/macro duality, and coarse geometry

In network science, a double-fractal architecture appears as bifractality in fractal scale-free networks (Yamamoto et al., 2023). The relevant object is the partition-function exponent $D_m \approx 1.8$1, which becomes piecewise linear when the number of nodes in a minimal box is proportional to the degree of the corresponding super-node in the renormalized network:

$D_m \approx 1.8$2

Under this condition,

$D_m \approx 1.8$3

so the local fractal dimension takes only two values,

$D_m \approx 1.8$4

The lower branch describes hub-centric regions; the higher branch describes periphery-centric regions. The paper conjectures that any fractal scale-free network is bifractal.

A different abstract formulation comes from multi-valued dynamical systems, where the duality is between micro-fractals and macro-fractals rather than between two simultaneous physical scales (Banakh et al., 2013). For a multi-function $D_m \approx 1.8$5, the fixed fractal is

$D_m \approx 1.8$6

with

$D_m \approx 1.8$7

If $D_m \approx 1.8$8 is contracting and compact-valued on a complete metric space, $D_m \approx 1.8$9 is a compact micro-fractal. If instead $p_c \approx 0.16$0 is contracting, then $p_c \approx 0.16$1 is the micro-fractal and $p_c \approx 0.16$2 becomes the dual macro-fractal, typically discrete and visible only after suitable zooming or spherical/projective embedding. This formulation does not use “double-fractal architecture” for a simultaneous two-scale material, but it does establish a precise micro/macro duality between two fixed-fractal constructions generated by the same underlying system.

A third abstract route is the coarse-geometric analysis of deep weight spaces (Moharil et al., 18 Mar 2025). The paper does not explicitly introduce a “double-fractal” architecture. However, it formalizes recursive segmentation with a coarse group action $p_c \approx 0.16$3 on the integer grid

$p_c \approx 0.16$4

at scales

$p_c \approx 0.16$5

and then extends this to two nested ladders, macro and micro. The proposed macro- and micro-level recursions, along with regularizers $p_c \approx 0.16$6 and $p_c \approx 0.16$7, supply a formal blueprint for enforcing discrete scale invariance at two levels in a neural architecture.

5. Recursive computational and neural architectures

FractalCloud provides a hardware-centered instance in which the phrase “double-fractal” is not explicitly coined, but the architecture is naturally described as having two fractal levels (Fu et al., 10 Nov 2025). The first is a data fractal: a binary-tree, self-similar decomposition of a point cloud down to a threshold block size using the rule

$p_c \approx 0.16$8

followed by recursive partition until $p_c \approx 0.16$9. The second is a compute fractal: recursive execution of point operations at block granularity with inter-block and intra-block parallelism, where the search space for grouping or interpolation is restricted to a leaf block and, at most, its parent. The resulting DFT block layout produces spatially coherent streamed memory access, and the total partition work is $d_f^{\max}$00 with $d_f^{\max}$01. Implemented in 28 nm as a chip layout with a core area of $d_f^{\max}$02, FractalCloud achieves $d_f^{\max}$03 speedup and $d_f^{\max}$04 energy reduction over state-of-the-art accelerators while maintaining network accuracy.

Fractal generative models provide an explicit two-level recursive generative formulation (Li et al., 24 Feb 2025). With atomic generative modules $d_f^{\max}$05, a double-fractal is the $d_f^{\max}$06 case: a parent module at level $d_f^{\max}$07 generates child contexts, and shared level-$d_f^{\max}$08 modules produce the leaf distributions. In the image instantiation, the two-level model has a top-level sequence length of $d_f^{\max}$09 pixels and a leaf sequence length of $d_f^{\max}$10 channels on ImageNet $d_f^{\max}$11, requiring about $d_f^{\max}$12 GFLOPs and reaching NLL $d_f^{\max}$13 bpd. The paper contrasts this with a three-level hierarchy at about $d_f^{\max}$14 GFLOPs and NLL $d_f^{\max}$15–$d_f^{\max}$16 bpd, so the double-fractal is presented as a strong but suboptimal trade-off relative to deeper recursion.

A related template-driven formulation appears in the FractalNet-style generator for large-scale model exploration (Mittal et al., 10 Nov 2025). There, a single fractal is a recursively defined DAG with join operator

$d_f^{\max}$17

and a double-fractal replaces the atomic base block by an inner fractal:

$d_f^{\max}$18

$d_f^{\max}$19

Without weight sharing across copies, parameters and FLOPs scale as $d_f^{\max}$20. The study reports more than $d_f^{\max}$21 variants, training with PyTorch, AMP, and gradient checkpointing on CIFAR-10 for five epochs, with mean training time per epoch of about five minutes, mean GPU memory of $d_f^{\max}$22–$d_f^{\max}$23 GB, and success rate of about $d_f^{\max}$24. The paper does not isolate a clean single-versus-double ablation, but it identifies moderate macro depth $d_f^{\max}$25–$d_f^{\max}$26 and small micro depth $d_f^{\max}$27–$d_f^{\max}$28 as the practically stable regime.

Across these computational examples, the defining feature is not simply recursion. It is the coexistence of two recursive ladders that remain distinguishable: data layout versus execution layout in hardware, parent versus child generative levels in probabilistic models, and outer versus inner fractal graphs in neural templates.

6. Architectural analogues, topological lattices, and limits of the term

Architectural and materials studies show both the reach and the limits of the concept. In the Borobudur analysis, the temple is described as “stupas within stupas,” and the study reports scale-invariant counts of stupa-like elements together with a box-counting dimension

$d_f^{\max}$29

and directional power laws

$d_f^{\max}$30

The proposed generative rules

$d_f^{\max}$31

recover the macro Head:Body:Foot ratio $d_f^{\max}$32. However, the paper explicitly states that it does not demonstrate two distinct scaling regimes across different size intervals and therefore stops short of claiming a double-fractal architecture; the observed structure is better described as anisotropic self-similarity, with a possible double-fractal extension only suggested (Situngkir, 2015).

In topological photonic lattices, the 2026 study establishes exponential scaling of boundary states for single-scale fractal-inspired unit-cell hierarchies and then identifies a double-fractal architecture as a natural next step (Song et al., 1 Apr 2026). For the quasi-1D Koch chain,

$d_f^{\max}$33

while for the 2D Sierpiński tiling,

$d_f^{\max}$34

The integer $d_f^{\max}$35 in $d_f^{\max}$36 is symmetry determined: $d_f^{\max}$37 by inversion symmetry in the Koch chain and $d_f^{\max}$38 by $d_f^{\max}$39 symmetry in the higher-order topological Sierpiński tiling. The paper then proposes a two-scale construction in which a fractal unit cell of generation $d_f^{\max}$40 is periodically arranged into a supercell that is itself fractal-like at generation $d_f^{\max}$41, leading, under multi-topological-phase constraints, to factorized scaling such as

$d_f^{\max}$42

This suggests a materials interpretation of double-fractal architecture as concomitant fractality at the unit-cell and superlattice scales rather than merely two exponents in one spectrum.

Taken together, these studies show that “double-fractal architecture” is not a single standardized doctrine. In explicit uses, it denotes two distinct fractal processes that are both necessary for prediction, as in superconducting percolation and colloidal rheology. In mathematical and computational settings, it often denotes a two-level self-similar construction or a bifractal decomposition. In architectural and topological analogues, it can function as a proposed extension rather than an experimentally established designation. The strongest common denominator is therefore structural: two coupled fractal organizations must remain distinguishable, and observable behavior must depend on their interaction rather than on a single effective fractal description.