Technical Summary: Knowledge-Constrained Neural Completion of Islamic Geometric Patterns with Symmetry Guarantees

Problem Definition

Islamic geometric ornament is defined not by stylistic resemblance but by exact geometric structure, including precise rotational symmetry ($N$-fold), preserved anchor points, and coherent vector geometry. Historically, these patterns are constructed via systematic compass and straightedge methods. The paper identifies a critical gap in modern neural generation: while neural networks can produce visually plausible imagery, they typically learn symmetry statistically and cannot certify that an output adheres to the strict structural rules of the domain. A minor symmetry violation renders an Islamic geometric pattern "structurally invalid."

The authors frame the task not as image synthesis, but as a structured geometric completion problem. Given sparse control geometry (minimal construction points and anchors derived from traditional layouts) and a target symmetry order $N$, the goal is to complete a full pattern as a vector graph. The output must consist of edges selected from a geometrically meaningful candidate lattice, optional bounded curve refinements, and must exhibit exact $N$-fold symmetry.

Methodology

1. Knowledge-Constrained Framework

The core of the methodology is embedding explicit geometric domain knowledge into the neural learning process.

• Candidate Lattice Construction: Instead of asking the network to invent arbitrary geometry, the system deterministically generates a candidate lattice from the sparse control geometry and the target symmetry order $N$. This lattice includes vertices on concentric rings and a set of admissible edges (ring edges, radial spokes, star-polygon chords, etc.).
• Orbit Structure: The candidate edges are partitioned into rotational orbits under the cyclic group $C_N$. An edge $e_i$ belongs to an orbit $O(e_i) = \{ g \cdot e_i : g \in C_N \}$. A completed pattern is valid if and only if its selected edge set is a union of complete orbits and refinement parameters are constant within each orbit.
• Neural Predictor: A graph neural network (GNN) operates on this candidate lattice. It takes vertex and edge features (position, radius, angle, anchor status, edge type) and predicts:
  • An edge score $s_i$ for selection.
  • A bounded refinement parameter $\alpha_i$ for ornamental curve adjustments.

2. Decoding Regimes

The paper compares four decoding strategies to isolate the impact of symmetry structure:

1. Orbit-Tied: The network averages score logits and refinement outputs over each orbit during the forward pass (both training and inference). This enforces $s_i = s_j$ and $\alpha_i = \alpha_j$ for edges in the same orbit.
2. Selection Projection: An unconstrained network is trained without tying. At inference, scores are aggregated to the orbit level before selection.
3. Full Projection: An unconstrained network is trained without tying. At inference, both scores and refinement parameters are averaged over orbits.
4. Unstructured (Free): An unconstrained network selects edges independently without using orbit information (negative baseline).

3. Bounded Refinement

Refinements are applied conservatively. Structural edges remain straight. Ornamental edges are refined by displacing the midpoint along the edge normal by $\tanh(\alpha_i) \alpha_{max} L_i$. The $\tanh$ saturation ensures the displacement never exceeds the bound $|\delta_i| \le \alpha_{max} L_i$, regardless of network output.

Key Contributions

1. Knowledge-Constrained Graph Neural Framework: A system where explicit construction rules define the candidate lattice, orbit structure, and validity-preserving decoding. The network predicts over this knowledge-defined space rather than a free combinatorial space.
2. Constructive Validity Guarantee: A formal proof (Proposition 1) demonstrating that the orbit-tied decoding scheme guarantees:
   • Exact $N$-fold rotational symmetry (zero selected-orbit violations).
   • Exact preservation of anchor vertices.
   • Bounded refinement displacements.
   • These invariants hold for any input and any orbit-level selection policy, independent of the network's learning state.
3. Empirical Characterization of Symmetry Structure: A controlled study across four decoding regimes showing that symmetry-structured inference (whether intrinsic or via projection) preserves validity with no measurable loss in completion fidelity compared to unstructured decoding.
4. Robustness and Failure Mode Analysis: An evaluation showing that while data augmentation (training with corrupted inputs) recovers fidelity in unstructured models, it does not restore exact validity. Conversely, symmetry-structured inference guarantees validity even when input geometry is missing, addressing complementary failure modes.

Experimental Results

Experiments were conducted on procedurally generated Islamic-geometry-inspired pattern graphs (symmetry orders 6, 8, 10, 12) across three random seeds.

• Fidelity on Clean Inputs: On clean test data, the orbit-tied model, selection projection, and unstructured free decoding achieved statistically indistinguishable edge-selection F1 scores (approx. 0.52). The procedural template baseline achieved a slightly higher F1 (0.534) but did so through severe edge over-completion (low precision, high recall), demonstrating that overlap metrics alone are insufficient for this domain.
• Structural Validity: The unstructured model violated orbit consistency on average 3.3 times per pattern, even on clean inputs. In contrast, the orbit-tied and projection methods achieved zero orbit violations, zero anchor drift, and zero excess refinement.
• Robustness to Missing Geometry: When 30% of the input control geometry was removed:
  • The unstructured model's F1 dropped significantly (0.524 to 0.415), and violations increased to 37.5 per pattern.
  • The structured methods maintained high fidelity (F1 ~0.50–0.51) and zero violations.
  • Training the unstructured model with missing-geometry augmentation restored its fidelity (F1 ~0.517) but did not eliminate orbit violations (9.0 per pattern).
  • Conclusion: Data augmentation improves fidelity under corruption, while symmetry structure guarantees validity. The two are complementary.

Significance and Claims

The paper claims that exact geometric validity can be imposed without a measurable loss in completion fidelity. The primary contribution is demonstrating how explicit geometric construction knowledge can constrain neural graph prediction to yield outputs that are not merely plausible but certifiably valid.

The authors emphasize that the "orbit structure" of the candidate geometry is the operative mechanism for validity, rather than a specific architectural device. Both the constructive (orbit-tied) and post-hoc (projection) approaches yield equivalent validity and fidelity, suggesting that enforcing the orbit structure at the decoding stage is sufficient.

The work positions itself as a step toward trustworthy AI in domains where structural rules are non-negotiable. It argues that for Islamic geometric patterns, validity is a structural requirement encoded as explicit knowledge, not a visual preference. The framework provides a scalable, vector-based approach to generating ornament that remains robust to imperfect input data, a common scenario in practical applications involving sketches or scans.

The authors remain modest, noting that the system does not claim superiority over procedural construction systems for faithful reproduction of historical families, nor does it claim to certify historical authenticity. The generated outputs are "completions inspired by Islamic geometric design" that are structurally valid according to the specified rules.