Local Real Architectures

• Local Real Architectures are computational frameworks that generate empirically grounded models capturing the morphological, cultural, and functional diversity of urban environments.
• They employ automated pipelines using 2D heat-map representations, VQ-VAE neural encoding, and latent clustering to produce simulation-ready 3D building prototypes.
• Quantitative evaluations demonstrate significant improvements in energy simulation accuracy and stylistic classification compared to conventional, one-size-fits-all models.

Local Real Architectures refer to computational and analytic paradigms for constructing, identifying, and utilizing representations of the built environment that are simultaneously faithful to local context ("local"), empirically grounded in real-world data ("real"), and structured for systematic analysis and simulation ("architectures"). These models aim to capture the morphological, cultural, and functional diversity of buildings and urban forms at a regional or even neighborhood scale. Local Real Architectures stand in contrast to generic, globally parameterized models by prioritizing empirical calibration, locale-specific typologies, and direct integration into quantitative workflows such as energy modeling, microclimate simulation, or stylistic analysis (Zhuang et al., 2024, Zhong et al., 9 Jun 2025, Galanos et al., 2021).

1. Foundations and Motivation

The impetus for Local Real Architectures arises from the inadequacy of prototypical or one-size-fits-all models in urban energy simulation, climatic performance, and architectural style studies. Traditional archetype libraries—such as the U.S. DOE Prototype Building Models (PBM)—are unable to reproduce the fine-grained morphological, envelope, and usage distinctions that drive energy demand and environmental behavior at the district or city scale (Zhuang et al., 2024). Similarly, region-agnostic generative design methods fail to account for the cultural and stylistic particularity embedded in actual urban fabric (Zhong et al., 9 Jun 2025).

A fundamental tenet is that high-fidelity architectural analytics demand empirically anchored, contextually differentiated prototypes. These must be extracted or synthesized directly from local datasets: polygonal footprints, heights, envelope characteristics, or image corpora annotated for cultural provenance.

2. Automated Extraction of Local Real Archetypes

The generation of Local Real Architectures for simulation-ready modeling is exemplified by the pipeline introduced by Zhuang et al. (Zhuang et al., 2024), consisting of:

• Geometric Representation: Buildings are encoded as single-channel 2D heat-maps (128 × 128) with pixel intensities representing height (0–100 m) and binary supports encoding plan footprints. Data are obtained from municipal sources and normalized per building.
• Neural Encoding: A Vector-Quantized Variational Autoencoder (VQ-VAE) ingests these images. The convolutional encoder transforms each map into a spatial tensor (e.g., 16 × 16 × 64), quantized against a learned codebook (K ≈ 512).
• Latent Clustering: After quantization, per-building latent vectors are averaged spatially to yield a D-dimensional embedding $\bar{z}_n$. Clustering (K-means) in latent space creates typological clusters. Cluster count K is chosen by the WCSS elbow criterion.
• Prototype Generation: Cluster representatives are decoded to yield canonical 2D plans and extruded with mean heights to construct 3D mass models.
• Local Parameter Integration: Envelope properties (insulation, WWR), climatic data (local .epw files), and cultural metadata (e.g., year built, usage type) are matched to each archetype for simulation.
• Assembly and Simulation: Prototypes are exported as geometry-compatible files (e.g., Rhinoceros/Grasshopper) and subjected to energy simulation with real district parameters.

This pipeline ensures empirically grounded, locale-calibrated archetypes suitable for district-scale modeling.

3. Quantitative Evaluation and Fidelity

The primary evaluation of Local Real Architectures involves both reconstructive and predictive metrics:

• Reconstruction Error: For the Russian Hill dataset, the VQ-VAE achieves a per-pixel MSE of ≈0.015 after ~2,000 epochs, demonstrating geometric fidelity (Zhuang et al., 2024).
• Simulation Accuracy: Aggregated annual energy estimates based on sampled archetypes approach measured consumption with high precision: for Russian Hill, archetype-based simulation yields 89.6% accuracy relative to metered data, reducing error by over 20 percentage points compared to conventional PBMs. Sampling the closest cluster member further outperforms decoded averages (10.4% vs. 20.4% energy error).
• Transferability: Similar accuracy improvements (up to +34% over PBM) are obtained across multiple districts, indicating robustness.

A plausible implication is that Local Real Architectures substantially narrow the discrepancy between simulated and real-world aggregate demand, thus enabling higher-confidence policy analysis and retrofit targeting.

Method	Russian Hill Annual Energy (kWh)	Accuracy vs. Real (%)
DOE PBM	1.48×10¹¹	66.0
Real (metered)	2.24×10¹¹	100
Local Real (average)	2.70×10¹¹	79.6
Local Real (sampled)	2.01×10¹¹	89.6

4. Regionally Differentiated Style and Typology Analytics

Local Real Architectures are operationalized in architectural stylistic analysis through frameworks such as ArchiLense (Zhong et al., 9 Jun 2025). The process includes:

• Dataset Construction: A comprehensive, architect-anchored image corpus (ArchDiffBench, 1,765 images, 10 regional groups) encoding local styles.
• Vision-Language Embedding: Images and textual descriptions are embedded into a unified 1,024-dimensional latent space (CLIP ViT-G/14).
• Signature Extraction: Descriptive linguistic traits are automatically distilled via LLMs (GPT-4V, BLIP-2) and filtered for discriminative power using cosine similarity metrics and statistical tests.
• Statistical Evaluation: High expert-consistency (92.4%) and classification accuracy (84.5%) confirm that learned signatures correspond to genuine local stylistic distinctions.

The resulting analytic pipeline allows direct, quantitative comparison of regional or historical building styles, supplanting subjective prose with repeatable, objective metrics and extracting measurable patterns unique to each locale or architectural lineage.

5. Extensions to Arbitrary Graph-Structured Architectures

The theoretical foundation for Local Real Architectures extends to quantum information and computational design: in "Local random quantum circuits form approximate designs on arbitrary architectures," the architecture is generalized to any connected graph $G=(V,E)$ (Mittal et al., 2023). Here, "architecture" refers to the allowed structure of local operations—e.g., the topology of two-qudit gates.

Key results demonstrate that:

• For graphs with bounded-degree spanning trees and logarithmic height, approximate k-designs can be generated with circuit size $O(|E|n\,\mathrm{poly}(k))$.
• For all connected architectures, even without degree or height constraints, quasi-polynomial circuit size suffices.
• Explicit spectral gap bounds (via Knabe's criteria or the Detectability Lemma) determine mixing times; for $k=2$, $Δ_0\leq38$, and $τ\leq90|E|(4n+θ)$ is sufficient for ε-approximate 2-designs.

This unifies the notion of "architecture" across both physical (built) and algorithmic (graph-topological) domains, with "local real" connoting empirically or physically realizable pathways rather than purely formal constructs.

6. Implications and Applications

Local Real Architectures empower high-fidelity, context-aware simulation and analysis across built environment and computational domains:

• Urban Energy Modeling: Local prototypes calibrated by real geometry/metadata, reducing energy modeling errors and enabling more granular retrofit targeting (Zhuang et al., 2024).
• Stylistic Classification: Embedding-based, region-aware pipelines for style recognition, comparison, and classification; supporting both automated and expert-guided studies (Zhong et al., 9 Jun 2025).
• Urban Design Optimization: Algorithms such as ARCH-Elites utilize actual site data, local constraints, and surrogate models to generate diverse, feasible urban forms under practical constraints, illuminating trade-offs between density, comfort, and regulatory regimes (Galanos et al., 2021).
• Quantum Circuit Complexity: Rigorous bounds on design generation for arbitrary architectures, with parallelizable, local constructions scaling to large system sizes and yielding nearly optimal results on many graph families (Mittal et al., 2023).

A plausible implication is that continued advances in locality-sensitive modeling and analysis will further entwine architectural, urban, and computational research communities, enabling new paradigms for energy, comfort, and cultural analytics that are simultaneously scalable, precise, and representative of actual built form.

7. Limitations and Future Directions

While Local Real Architectures offer substantial gains in fidelity and applicability, several caveats attend their use:

• Data Quality and Coverage: Model accuracy is controlled by the quality, granularity, and completeness of local building datasets or style image corpora.
• Algorithmic Generality: VQ-VAE encoding and clustering pipelines may fail to capture rare morphological outliers or emergent urban forms absent in historical data.
• Physical Realism: Surrogate models (e.g., for microclimate/wind) may not fully encompass dynamic or transient phenomena unless retrained or supplemented.
• Transferability: While demonstrated across multiple districts/regions, the method requires careful recalibration when extrapolated beyond the domain of the training set.

Future work may incorporate adaptive model selection, outlier detection, and multi-modal integration (e.g., integrating materials, usage patterns, and high-frequency sensor data), as well as further theoretical unification of locality principles across architectural, urban, and computational design frameworks.