Cluster-Based Isotropy Enhancement (CBIE)

• CBIE is a dual-domain framework that enhances isotropy by clustering high-index particles in metasurfaces and correcting anisotropy in contextual embeddings.
• In photonics, it employs symmetry-breaking perturbations to create nearly azimuthally uniform electric near-fields with isotropy (I ≥ 0.9) and high-Q resonances.
• In NLP, CBIE applies local principal component removal within embedding clusters, significantly improving semantic expressiveness and downstream task accuracy.

Cluster-Based Isotropy Enhancement (CBIE) designates a class of methods that employ clustering and symmetry principles to enhance the isotropy—or statistical and spatial uniformity—of a target property across a complex system. In the physical sciences, CBIE characterizes the paradigm of constructing metasurface unit cells from structured clusters of high-index dielectric particles, leveraging controlled perturbations to achieve highly uniform, enhanced electric near-fields. In representation learning, CBIE denotes a local, cluster-wise post-processing technique to correct anisotropy in contextual embedding spaces, improving semantic expressiveness and downstream task performance. The two principal threads—the electromagnetic and the embedding-space frameworks—are connected by the central notion of mitigating undesirable anisotropy through cluster-specific analyses and interventions.

1. Definition and Scope of CBIE

In all-dielectric metasurfaces, Cluster-Based Isotropy Enhancement is realized by assembling the unit cell from multiple subwavelength, high-refractive-index elements, and introducing tailored symmetry-breaking perturbations. This configuration fosters the excitation of collective trapped electromagnetic modes with strong, nearly isotropic electric near-field enhancement across the in-plane region of the cluster, quantified by an isotropy parameter

$$I = 1 - \frac{\max_\phi E(\phi) - \min_\phi E(\phi)}{\max_\phi E(\phi) + \min_\phi E(\phi)}\,,$$

which approaches unity for optimal isotropy (Kupriianov et al., 2019).

In contextual word representations (CWRs), CBIE addresses representation degeneration, whereby embedding spaces become highly anisotropic under language-model pre-training objectives. Here, CBIE refers to explicitly clustering embeddings and performing within-cluster isotropy correction, generally by removing dominant principal components per cluster, thereby aligning each local covariance toward isotropy and restoring semantic expressiveness (Rajaee et al., 2021).

2. Theoretical Underpinnings in Metasurfaces

CBIE in metasurface engineering is grounded in the electromagnetic eigenproblem for periodic arrays of high-permittivity resonators. The underlying structure is described by time-harmonic Maxwell’s equations, with the local permittivity profile $\varepsilon(\mathbf r)$ encoding both host and resonator materials:

$$\nabla \times \nabla \times \mathbf{E}(\mathbf{r}) - \omega^2 \mu_0 \varepsilon(\mathbf{r}) \mathbf{E}(\mathbf{r}) = 0\,,$$

subject to interface and radiation boundary conditions (Kupriianov et al., 2019). The eigenvalue problem for the metasurface lattice includes the interaction between individual clusters via the lattice Green’s function. A trapped (quasi-BIC) mode appears when

$$\det \left[ \alpha^{-1}(\omega)\, \mathbf{I} - \mathbf{G}(\omega, \mathbf{k}_\parallel)\right]=0\,,$$

where $\alpha(\omega)$ is the polarizability tensor of an isolated particle and $\mathbf{G}$ is the lattice Green’s function. The resulting high-$Q$ trapped mode produces a near-field enhancement quantified by $\eta = |E_{\text{loc}}|/|E_0| \sim Q$ at resonance. Isotropy of the in-plane field is analytically controlled by cluster symmetry and perturbative parameters, ensuring that $I \gtrsim 0.9$ across relevant designs.

3. Algorithmic Framework in Embedding Spaces

In contextual embedding models, CBIE comprises a four-stage pipeline:

1. Clustering: Embeddings $X = \{x_i \in \mathbb{R}^D: i=1..N\}$ are grouped using $k$-means; the number of clusters $C$ is a tunable hyperparameter (e.g., $C = 10$–$30$).
2. Local Isotropy Assessment: Within each cluster $c$, centered embeddings $\bar{X}_c$ are analyzed by forming the empirical covariance $\Sigma_c$ and computing its eigendecomposition.
3. Dominant Direction Removal: The $k$ leading principal components are selected per cluster. For an embedding $x_i$ in cluster $c$, its projection onto the dominant subspace is subtracted:

$$x'_i = x_i - \sum_{j=1}^k (u_{c, j}^\top x_i) u_{c, j}\,,$$

where $\{u_{c,j}\}$ are the top $k$ eigenvectors.

1. Output Formation: The corrected ensemble $\{x'_i\}_{i=1}^N$ is used for downstream tasks (Rajaee et al., 2021).

Cluster-specific isotropy metrics provide a finer description than global ones, revealing that the embedding geometry is more locally isotropic after this cluster-based principal component removal.

4. Symmetry, Perturbation, and Control of Isotropy

In metasurfaces, the isotropy of the near-field is governed by the point group symmetry ($C_{4v}$ for square clusters), and three physical perturbation strategies are detailed:

• Disk displacement: Shifts each disk toward the cluster center by $s$, breaking inversion symmetry (asymmetry quantified by $\theta = 2s/(p-2r_d)$).
• Off-center circular holes: Introduces an inner hole of radius $r_h$ offset by $s$ ($\theta = 2r_h / r_d$).
• Coaxial-sector notches: Carves a sector-shaped notch of angle $\alpha$ ($\theta = \sin(\alpha/2)$).

These perturbations minimally couple the otherwise dark, symmetry-protected mode to free space, raising $Q$ by orders of magnitude while maintaining a negligible anisotropic angular component $a_4(\theta)$. The in-plane field structure is nearly azimuthally uniform, and even for moderate asymmetry required to achieve $Q \sim 10^2$, the isotropy parameter $I$ remains above 0.9.

In embedding spaces, cluster structure reflects discrete linguistic phenomena, such as punctuation or verb tense. The removal of cluster-local principal axes unmasks semantic structure otherwise dominated by non-semantic variance. The empirical outcome is that after CBIE, sense similarities in verbs supersede tense-based clustering, and structural tokens (punctuation, stopwords) show reduced style bias in nearest neighbor organization (Rajaee et al., 2021).

5. Empirical Results and Design Guidelines

In metasurface experiments and finite-element simulations, the CBIE paradigm was demonstrated at microwave frequencies using clusters of high-permittivity disks ($\varepsilon_d = 23$) arranged on a $p = 32$ mm lattice:

Parameter	Value / Range	Description
Lattice period $p$	32 mm	Unit cell size
Disk radius $r_d$	4.0–4.5 mm	Physical disk size
Disk height $h_d$	2.5 mm
Substrate $h_s$	25 mm
Permittivity	$\varepsilon_d=23$, $\varepsilon_s=1.1$	Disk/substrate

For a coaxial-sector notch design with $\theta = 0.3$, the trapped mode centered at $\approx 7.1$ GHz shows $Q \approx 150$, $\eta_{\max} \approx 2 \times 10^4$, and $I \gtrsim 0.92$. The in-plane field is spatially homogeneous, validated by experimental near-field scans and simulated $E_x$ distributions (Kupriianov et al., 2019).

In contextual embeddings, CBIE delivers substantial improvements on semantic benchmarks. For instance, on semantic textual similarity tasks (STS), the average Spearman $\rho$ rises (e.g., RoBERTa-base: baseline 51.2→global 65.7→CBIE 66.4). Classification accuracy also increases across multiple datasets, with BERT models improving from 61.2% (baseline) to 63.2% (CBIE). The parameter $k$ (PCs per cluster) is empirically tuned for each model, with optimal settings ranging from $k=12$ (BERT, RoBERTa) to $k=30$ (GPT-2) (Rajaee et al., 2021).

6. Practical Considerations and Applications

In metasurface fabrication, the CBIE approach is robust to moderate lithographic or milling errors, as field isotropy ($I \gtrsim 0.88$) and quality factor ($Q$) are only weakly sensitive to small deviations ($\Delta \theta \lesssim 0.02$), well within typical process tolerances. The full-width at half-maximum bandwidth of the trapped-mode resonance is $\Delta f \approx f_\mathrm{tr}/Q$; for $Q=100$ and $f_\mathrm{tr}=200$ THz, $\Delta f\approx 2$ THz, affording a practical frequency range for isotropic near-field enhancement. The spatial distribution of the enhanced field, largely external to the high-index resonators, is conducive to integration with gain or nonlinear media—e.g., quantum dots or nonlinear polymers embedded in the cluster gaps—for flat spaser or nanolaser implementations, ultrafast optical modulation, or highly uniform surface-enhanced Raman spectroscopy (SERS) (Kupriianov et al., 2019).

In the field of contextual embeddings, CBIE’s local corrections alleviate down-stream representation bottlenecks without necessitating retraining. The approach is especially effective in re-balancing the relative prominence of semantic versus syntactic/structural information, yielding measurable gains in both similarity and classification tasks (Rajaee et al., 2021).

7. Summary and Outlook

Cluster-Based Isotropy Enhancement furnishes a design and analysis framework grounded in cluster symmetry and local covariance control, applicable both to engineered materials and high-dimensional representation spaces. In metasurfaces, it enables the coexistence of high-$Q$ resonances and uniform near-field enhancement, facilitating advanced photonic and nonlinear device platforms. In natural language processing, CBIE restores isotropy in degenerate embedding spaces, improving their suitability for semantic tasks. The distinct domains are unified by a common mathematical and algorithmic motif: locally adaptive interventions within clustered structures achieve global isotropy and enhanced system performance (Kupriianov et al., 2019, Rajaee et al., 2021).