ZnG Architecture Framework

• ZnG architecture is a layered, constraint-based framework that separates spatial topology search from geometric refinement in space layout planning.
• It employs a discrete topological model (Z) for adjacency and orientation alongside a geometric model (G) for integer dimension optimization.
• The approach enables exhaustive conceptual sketch enumeration and efficient cost optimization while ensuring clear separation of design concerns.

ZnG architecture refers to a layered, constraint-based computational framework that separates the search for feasible spatial topologies from subsequent geometric resolution in the context of space layout planning. Developed in the early 2010s for conceptual CAD by Medjdoub and Yannou, the ZnG model formalizes the workflow of architectural design by mapping ‘Z’ to discrete topological reasoning (adjacency, orientation, connectivity) and ‘G’ to metric geometric realization (assigning concrete integer dimensions to spaces). This clean partition enables exhaustive enumeration of conceptual sketches followed by tractable geometric refinement, contrasting sharply with earlier approaches that attempted to solve topology and geometry in a single monolithic phase (Medjdoub et al., 2013).

1. Topological Model Z: Discrete Adjacency and Orientation

At the Z level, the layout is abstracted as a system of isothetic (axis-aligned) rectangles representing individual architectural spaces (rooms, corridors, etc.), with explicit disregard for their exact size and position. Variables associated to the Z model include:

• Symbolic domains for length $L_e$ and width $W_e$, corner coordinates $(x_{1,e},y_{1,e}),(x_{2,e},y_{2,e})$, which remain uninstantiated.
• Discrete variables for spatial relationships between pairs of spaces $e_i, e_j$:
  • Non-overlap variable $$\delta_{ij} \in \{\mathrm{N},\mathrm{S},\mathrm{E},\mathrm{W}\}$$, enforcing the relative positional exclusivity.
  • Adjacency variable $$\alpha_{ij} \in \{\mathrm{N},\mathrm{S},\mathrm{E},\mathrm{W}\}$$, representing required side-by-side configuration.
  • Contact and separation lengths $d^1_{ij} \in [0, \infty)$, $d^2_{ij} \in \{0\}$, controlling minimal shared edges or mandatory gaps.
• Contour proximity variables $$\gamma_{i}\in\{\mathrm{N},\mathrm{S},\mathrm{E},\mathrm{W}\}$$ for bounding-wall adjacencies.

Core constraints are specified as follows:

• Geometric arc-consistency: For each space $e$, $x_{2,e} - x_{1,e} = L_e$, $y_{2,e} - y_{1,e} = W_e$, $S_e = L_e \cdot W_e$.
• Generalized adjacency: For $\alpha_{ij}=N$, the inequalities $y_{1,j} - y_{2,i} = d^2$, $x_{2,i} - x_{1,j} \leq -d^1$, $x_{1,i} - x_{2,j} \leq -d^1$ fix $e_j$ north of $e_i$.
• Non-overlap enforcement: e.g., for $\delta_{ij}=N$, $y_{1,j} - y_{2,i}\geq 0$.
• Symmetry-breaking: Total ordering on lower-left corners for interchangeable spaces avoids permutational redundancy.

Enumeration over these variables uses a backtracking constraint satisfaction procedure, selecting the most constrained variable at each search node for assignment, and propagating arc-consistency plus implicit pruning. When all orientation variables are fully instantiated, a valid topological solution is recorded.

2. Geometric Model G: Integer Dimension Assignment and Optimization

Given a fixed topological configuration $Z_k$ (a unique set of discrete $\{\delta_{ij},\alpha_{ij},\gamma_i\}$ assignments), the G phase instantiates integer-valued coordinates and sizes for each rectangle:

• Variable domains: $x_{1,e}, x_{2,e}, y_{1,e}, y_{2,e} \in \mathbb{Z}$, $L_e, W_e \in \mathbb{Z}$, grounded relational constraints.
• All previously abstract topological relationships now become hard geometric inequalities.
• Additional geometric constraints (corridor widths, aspect ratios) are applied, e.g., $$\frac{L_e}{W_e}\in[a_1/a_2, b_1/b_2] \Leftrightarrow a_1W_e < a_2L_e, L_eb_1 < W_eb_2$$.

The geometric problem is typically addressed as a branch-and-bound optimization – the search seeks either feasible placements or cost-minimizing layouts for criteria such as corridor area or total wall length. Each $Z_k$ gives rise to one or more $G$-solutions, admitting rapid pruning due to the strong domain restriction imposed by topology.

3. Z→G Interface and Feedback Mechanisms

The interface between Z and G is characterized by a freeze-and-check workflow:

• Selection of a topological (Z-level) solution for attempted realization at the G level.
• If the geometric constraint programming problem is infeasible, the topology is rejected; otherwise, it is refined or optimized.
• No full mixed-mode backtracking is performed between G and Z in the original implementation; infeasibilities detected at G are met by discarding Z$_k$ and continuing enumeration.

A plausible implication is that interactive or automated feedback from geometric failures to topological choices could further enhance design efficiency, though this is not implemented in the base system (Medjdoub et al., 2013).

4. Computational Workflow and Enumeration Heuristics

The detailed workflow is as follows:

• Apply all architectural specification constraints to Z variables.
• Propagate geometric arc-consistency and implicit pruning.
• On each non-leaf search node, choose the variable with the smallest non-unit domain (most constrained) for assignment.
• Upon recording a topological solution, generate the corresponding ground constraint set for G; solve using propagation, domain splitting, and cost-bound updating.
• Efficient variable and value ordering, bolstered by arc-consistency at both Z and G stages, enables scaling from small (10–20 space) to moderate (50+ space) layouts.

In practice, case studies report enumeration of 49 topologies for a two-floor house with 11 spaces, and 72 topologies for a house with 11 rooms in a prescribed area, with rapid downstream optimization at the geometric level (Medjdoub et al., 2013).

5. Benefits, Limitations, and Potential Extensions

Benefits:

• Exhaustive yet sharply reduced enumeration: orders of magnitude fewer conceptual sketches than possible full geometry instantiations.
• Human-aligned sketch output: Z-level outputs match early design sketches with imprecise sizing, supporting conceptual decision-making.
• Rigorous separation of concerns: Z level handles global adjacency/orientation, G enforces all metric constraints and supports architectural cost optimization.

Limitations:

• Strictly isothetic (axis-aligned) rectangles and grid-aligned coordinates; nonrectilinear or nonorthogonal spaces are not supported.
• Management of implicit constraints (symmetry-breaking, pruning of unreachable domains) is critical; over-pruning can exclude valid topologies.
• Full bi-directional feedback (from geometric infeasibility back to topological relaxation) is not realized in the original system.

Potential extensions:

• Generalization to rectilinear polygons, support for more complex space shapes (L, circular), or continuous (not discrete) orientations.
• Uncertainty weighting on design specifications, effectively integrating soft constraints and probabilistic cost into the G phase.
• Expanded cost criteria to include acoustics, lighting, or circulation.
• Application to industrial layouts or high-rise multi-storey planning.

6. Distinction from Related Spatial Reasoning Architectures

ZnG’s layered approach contrasts with constraint programming models that treat orientation and geometry in a monolithic or unlayered fashion. By mimicking the architect’s real workflow—sketch, enumerate, then refine—the architecture achieves near-complete coverage of feasible design options without combinatorial explosion in the number of fully-dimensioned alternatives. The discrete orientation-based enumeration, grounded in symbolic variables and explicit constraint propagation, sets ZnG apart from purely continuous or numeric optimization models (Medjdoub et al., 2013).

7. Impact and Applications

ZnG has been implemented and evaluated in the ARCHiPLAN conceptual CAD system, yielding exhaustive sets of topological sketches on architectural briefs of moderate size. For each distinct Z-level solution, the system delivers one (or an optimized few) geometric layouts, enabling side-by-side comparison for design refinement and client communication. The approach’s computational efficiency and clarity of design separation have influenced subsequent research in automatic space planning, computational design exploration, and constraint-based architectural software (Medjdoub et al., 2013).