Rotor-Based Layers: Computation and Fluid Dynamics

• Rotor-based layers are defined by rotational mechanisms, organizing structures across domains such as neural networks, fluid flows, and sensing films.
• In machine learning, these layers employ Clifford rotors and rotary positional encodings to achieve manifold-preserving state updates with significantly reduced parameter counts.
• In physical systems, rotor-induced stratification and sensor coatings demonstrate precise control over flow dynamics and measurement sensitivity, offering high performance in diverse applications.

In the literature, “rotor-based layers” denotes several domain-specific constructions in which rotation, rotors, or rotary structure is the organizing mechanism. In machine learning, the term most directly refers to layers built from Clifford rotors or to structured linear maps composed from bivectors and rotors. Closely related work studies rotary structure inside transformer attention and rotation-equivariant convolutional or contour-processing layers. In fluid mechanics, the phrase can describe layered rotating flows generated by a rotor, and in sensing it can refer to surface-bound molecular-rotor films used as active measurement layers (Huy et al., 10 Feb 2026, Pence et al., 15 Jul 2025, Jonasson, 3 Mar 2025, Viazzo et al., 2013, Nalatamby et al., 18 Feb 2025).

1. Terminological scope and disambiguation

The literature uses the expression in several distinct senses. Some usages describe learnable neural modules; others describe physical stratification induced by rotation; still others describe surface coatings or coupled mathematical layers in abstract routing models (Huy et al., 10 Feb 2026, Pence et al., 15 Jul 2025, Li et al., 2017, Viazzo et al., 2013, Nalatamby et al., 18 Feb 2025, Auger et al., 2024).

Setting	Layer notion	Defining mechanism
Geometric sequence models	Hidden-state layer	Rotor multiplication on a Spin manifold
Structured linear modules	Linear layer	Composition of bivectors and rotors
Transformer attention analysis	Rotary feature structure	RoPE rotations of query/key pairs
Rotation-aware vision and contour models	Equivariant layer	Filter rotation, radial tying, ring rotation, complex circular convolution
Rotating-flow physics	Boundary-layer system	Ekman, Bödewadt, Stewartson, and geostrophic organization
Microfluidic sensing and routing theory	Surface or coupled layer	Molecular-rotor polymer films; coupled vertex/arc graphs

A terminological false friend is layer rotation in optimization. There, rotation means the cosine distance between a layer’s current weights and its initialization,

$\mathrm{rot}_l(t)=d_{\cos}(w_l^{t_0},w_l^t),$

and the reported network-independent optimum is that all layers reach cosine distance $1$. This is a training diagnostic and control variable, not a rotor-based layer construction (Carbonnelle et al., 2018).

2. Clifford rotors as explicit learnable modules

The most literal neural usage appears in architectures that replace ordinary vector-space updates with manifold-preserving geometric evolution. Versor represents hidden states as multivectors in the conformal Clifford algebra $Cl_{4,1}$, the $32$-dimensional space generated by $\{e_1,e_2,e_3,e_+,e_-\}$, with $e_o=\tfrac12(e_- - e_+)$ and $e_\infty=e_-+e_+$. A $3$D point is lifted by

$X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$

which satisfies

$X_i\cdot X_j=-\frac12 \|x_i-x_j\|^2.$

Rotors $1$0 act by the sandwich product

$1$1

and the recurrent state update in the Recursive Rotor Accumulator is

$1$2

with

$1$3

The paper presents the Cayley transform as a practical substitute for $1$4, attributes $1$5 time and $1$6 memory to the recurrence, and states that removing manifold normalization causes immediate divergence/NaNs. Its Geometric Product Attention decomposes

$1$7

into a scalar proximity term and a bivector orientational term, with attention score

$1$8

Reported results include orders of magnitude fewer parameters ($1$9 vs. Transformers), $Cl_{4,1}$0 MCC on the topology task versus $Cl_{4,1}$1 for ViT, and custom Clifford kernels with up to $Cl_{4,1}$2 speedup (Huy et al., 10 Feb 2026).

A complementary formulation treats a rotor-based layer as a structured linear map

$Cl_{4,1}$3

with $Cl_{4,1}$4. This construction uses the identification $Cl_{4,1}$5, exponentiates bivectors into rotors, and gives the closed form

$Cl_{4,1}$6

for simple bivectors. A general bivector is decomposed into at most $Cl_{4,1}$7 mutually commuting, orthogonal, simple bivectors, and the resulting parameter count is $Cl_{4,1}$8 rather than $Cl_{4,1}$9. Applied to attention projections in LLaMA-3.2 $32$0B, the reported dense parameter counts for key, query, and value are $32$1, $32$2, and $32$3, whereas the rotor counts are at most $32$4, $32$5, and $32$6; the paper states that rotor layers are typically best or second-best against LR1, LR4, and BH1 baselines (Pence et al., 15 Jul 2025).

3. Rotary structure inside transformer layers

A distinct, but closely related, line of work studies rotary positional encodings rather than explicit rotor modules. RoPE pairs dimensions in the query and key vectors and rotates each pair by a position-dependent angle,

$32$7

The key identity is

$32$8

so the query-key dot product depends only on relative position. Using mean rotary vectors $32$9 and $\{e_1,e_2,e_3,e_+,e_-\}$0, the contribution of pair $\{e_1,e_2,e_3,e_+,e_-\}$1 at relative distance $\{e_1,e_2,e_3,e_+,e_-\}$2 is

$\{e_1,e_2,e_3,e_+,e_-\}$3

The analysis identifies a structured division of labor. Higher-frequency cyclic rotary features can build sharp local patterns such as sub-diagonal or previous-token attention. Low-frequency rotary offset features are defined by

$\{e_1,e_2,e_3,e_+,e_-\}$4

and can keep the dot product suppressed for most positions. The paper derives two bounds: $\{e_1,e_2,e_3,e_+,e_-\}$5 for the frequency, and

$\{e_1,e_2,e_3,e_+,e_-\}$6

for the initial query-key angle. Empirically, outliers concentrate in low-frequency partial-cycle rotary features, large-radius features cluster near the theoretical angle boundary, and frequencies above the upper bound rarely become offset outliers. The same pattern is reported for Phi-1, Llama-2-7B, and DeepSeek-V2-Lite, and the paper argues that rotary offset features help create attention sinks by making only a few keys align positively with the query while most other positions remain negatively aligned (Jonasson, 3 Mar 2025).

4. Rotation-equivariant layers without Clifford rotors

A large adjacent literature implements rotation-aware layers by constraining filters, channels, or local operators rather than by evolving states on a Spin manifold. Deep Rotation Equivariant Network uses cycle layers, isotonic layers, and decycle layers. The cycle layer rotates filters into four orientations and converts image rotation into a cyclic permutation among four channels. The isotonic layer preserves that permutation structure through constrained $\{e_1,e_2,e_3,e_+,e_-\}$7 filter hyper-matrices, and the decycle layer removes the permutation and outputs ordinary rotation-equivariant feature maps. The paper reports more than $\{e_1,e_2,e_3,e_+,e_-\}$8 speedup over feature-map-rotation methods with less memory overhead, $\{e_1,e_2,e_3,e_+,e_-\}$9 error on Rotated MNIST for DREN with max pooling, and $e_o=\tfrac12(e_- - e_+)$0 test error for $e_o=\tfrac12(e_- - e_+)$1-ResNet-20(conv1-13)$e_o=\tfrac12(e_- - e_+)$2 on CIFAR-10 (Li et al., 2017).

Other architectures impose rotational structure more locally. RAD constrains a kernel weight to depend only on its distance from the center; RSDW inserts a depth-wise separable radial middle layer; and RING rotates each ring of the kernel independently. The same paper emphasizes that fully connected layers are rotation invariant only when the representation before them is reduced to a one-dimensional vector, which is why its architectures end with pooling. On CIFAR-10, model b) with RING $e_o=\tfrac12(e_- - e_+)$3 reaches $e_o=\tfrac12(e_- - e_+)$4 across all four tested rotations, whereas an ordinary CNN $e_o=\tfrac12(e_- - e_+)$5 varies strongly with rotation. RRL, by contrast, is a parameter-free regional canonicalization module inserted before convolutional layers and after the last convolutional layer. For $e_o=\tfrac12(e_- - e_+)$6, it claims that feature maps can become exactly identical after the RRL processing; reported gains include ResNet-18 from $e_o=\tfrac12(e_- - e_+)$7 to $e_o=\tfrac12(e_- - e_+)$8 on CIFAR-10-rot and MobileNet-tiny-yolov3 from $e_o=\tfrac12(e_- - e_+)$9 to $e_\infty=e_-+e_+$0 mAP (Fuhl et al., 2020, Hao et al., 2022).

For contour data, RotaTouille formulates equivariance under

$e_\infty=e_-+e_+$1

with action

$e_\infty=e_-+e_+$2

on complex-valued signals $e_\infty=e_-+e_+$3. Its linear operator is complex circular convolution,

$e_\infty=e_-+e_+$4

and pointwise nonlinearities must satisfy

$e_\infty=e_-+e_+$5

The paper also defines equivariant coarsening and a global invariant pooling

$e_\infty=e_-+e_+$6

thereby combining cyclic-shift equivariance and planar-rotation equivariance in a fully complex-valued architecture (Gardaa et al., 22 Aug 2025).

5. Rotor-induced layering in rotating flows

In fluid mechanics, “rotor-based layers” often refers to the literal boundary-layer organization created by rotating machinery. In a confined rotor-stator cavity at

$e_\infty=e_-+e_+$7

the flow is organized into an Ekman layer on the rotating disk, a Bödewadt layer on the stationary disk, and a geostrophic core between them. The Bödewadt layer is described as almost twice as thick as the rotor layer. High-order LES with dynamic Smagorinsky closure (LES-FD) and spectral vanishing viscosity (LES-SVV) both reproduce the secondary meridional circulation and the layered structure, but LES-SVV is reported as slightly better overall. At $e_\infty=e_-+e_+$8, the core swirl ratio is approximately $e_\infty=e_-+e_+$9 for LES-SVV and experiments, $3$0 for LES-FD, and $3$1 for the RSM model; LES-SVV also predicts rotor transition at the theoretically expected local Reynolds number $3$2 (Viazzo et al., 2013).

Magnetized rotating boundary layers add Lorentz forcing to the classical Bödewadt problem. The Bödewadt-Hartmann analysis uses the Elsasser number

$3$3

and gives the effective thickness

$3$4

The paper emphasizes that there is one composite thickness, not nested Ekman and Hartmann sublayers. In rotating Rayleigh-Bénard convection, Ekman layers at the plates scale as $3$5, while Stewartson sidewall layers scale as $3$6 and $3$7. These boundary layers drive a secondary circulation that carries hot fluid upward along the lower sidewall and cold fluid downward along the upper sidewall, thereby preserving a vertical wall-temperature gradient even after the large-scale circulation disappears (Davidson et al., 2020, Oliveira, 2011).

A related rotor literature studies ensembles of driven microscopic rotors in a viscous membrane and ideal point vortices. That work reports a disordered hyperuniform state with approximately

$3$8

and, at $3$9, a transition to a hexagonal lattice. It concerns self-assembled rotor phases rather than boundary layers, but it extends the physical meaning of rotor-organized structure beyond wall-bounded flow (Oppenheimer et al., 2021).

6. Surface-bound and abstract coupled-layer constructions

In microfluidic sensing, a rotor-based layer can be an immobilized fluorescent film whose active unit is a molecular rotor. The sensing monomer MECVJ is a methacrylate-functionalized julolidone-based molecular rotor incorporated by RAFT polymerization into poly(DMA-\textit{s}-MECVJ). After grafting onto gold-coated glass, the resulting film acts as a viscosity-sensitive surface. The polymer preserves the viscosity-sensitive fluorescence response, follows Förster–Hoffmann behavior, and remains usable after repeated washing; the paper reports that after $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$0 hours of continuous washing the fluorescence lifetime response was still satisfactory. In the demonstration chip, a PDMS block with a straight channel of $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$1 mm width and $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$2m height is sealed against the grafted slide and read out by FLIM/LIFA at $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$3 nm. The authors report average $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$4 for grafted surfaces and chips, and dynamic switching between about $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$5 mPa·s and about $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$6 mPa·s flowing solutions (Nalatamby et al., 18 Feb 2025).

In combinatorics and distributed dynamics, rotor-routing introduces a different layered meaning. The generalized rotor mechanism couples a vertex layer $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$7 with an arc layer $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$8, where the vertices of $X=\mathcal{K}(x)=x+\frac12 x^2 e_\infty + e_o,$9 are the arcs of $X_i\cdot X_j=-\frac12 \|x_i-x_j\|^2.$0. A routing vector $X_i\cdot X_j=-\frac12 \|x_i-x_j\|^2.$1 acts through the linear map

$X_i\cdot X_j=-\frac12 \|x_i-x_j\|^2.$2

so routing simultaneously updates arc configurations and vertex configurations. Standard cyclic rotor-routing appears as the special case in which each $X_i\cdot X_j=-\frac12 \|x_i-x_j\|^2.$3 is a directed cycle on $X_i\cdot X_j=-\frac12 \|x_i-x_j\|^2.$4. The paper proves that deciding the existence of a routing vector is polynomial-time, that legal reachability in general GRM multigraphs is NP-complete, and that legal reachability in cyclic GRM multigraphs is in $X_i\cdot X_j=-\frac12 \|x_i-x_j\|^2.$5 (Auger et al., 2024).

These usages show that “rotor-based layers” is not a single canonical object. In current research it names a family of constructions in which rotational action, rotary feature pairs, rotor-induced stratification, or rotor-coupled state updates define the operative layer structure.