Local Real Architectures in Urban Modeling

• Local Real Architectures are data-driven frameworks that capture authentic, site-specific urban features using multiscale, semantic, and adaptive computational methods.
• They integrate advanced techniques like MARL and ARCH-Elites to improve energy modeling accuracy and generate context-sensitive urban designs.
• These approaches extend to locally-oriented programming for distributed HPC, ensuring spatial data fidelity and scalable performance in structured grid environments.

Local Real Architectures are data-driven computational frameworks and methodologies designed to capture, synthesize, and operationalize authentic, site-specific architectural or urban morphological features in support of modeling, simulation, and generative design tasks. These approaches emphasize extracting and utilizing local geometric, semantic, and ecological characteristics and stand in contrast to traditional reliance on generic, nationwide, or idealized prototypes. Across domains such as urban energy modeling, wind comfort assessment, and distributed scientific computing on hybrid architectures, Local Real Architectures seek to align the representation and operation of models with the realities of granular morphological, programmatic, or computational context.

1. Principles of Local Real Architectures

Local Real Architectures systematically prioritize the preservation and utilization of genuine local features—spatial, geometric, semantic, or behavioral—when constructing models or performing computational tasks. Their distinguishing tenets include:

• Data-grounded Locality: Models are derived from spatially resolved datasets characterizing actual footprints, morphologies, or usage patterns.
• Multiscale and Semantic Fidelity: Encoding strategies capture both macro-level shapes and micro-level idiosyncrasies, ensuring representational fidelity to neighborhood, district, or context-specific variation.
• Automatic Adaptability: Algorithms autonomously adapt to heterogeneity in building scale, structure, or configuration, rather than imposing a priori typologies.
• Separation from Generic Prototypes: Local Real Architectures replace, augment, or critique traditional, generic templates (such as national building prototypes) with context-sensitive, empirically derived archetypes.
• Operational Transferability: Robustness to transfer across unseen districts or contexts is enabled via latent geometry, semantic constraints, or quality-diversity search.

These principles are exemplified across several contemporary research lines, notably in MARL for urban energy modeling (Zhuang et al., 2023) and ARCH-Elites for diverse urban design evaluation (Galanos et al., 2021).

2. MARL: Multi-scale Archetype Representation Learning

The MARL framework operationalizes Local Real Architectures through a deep representation learning approach for urban building energy modeling (Zhuang et al., 2023). Its workflow is as follows:

• Multi-scale Preprocessing: Each building footprint raster is processed across three scales (700×700, 224×224, 112×112 pixels), resized to 112×112, and stacked as channels to form a composite input.
• Encoding Architecture: A three-layer convolutional neural network (CNN) with residual blocks produces a 28×28×32 latent map per building; this map is subsequently vector-quantized via a learned codebook.
• Semantic Supervision via Downstream Task Pool (DTP): Auxiliary neural heads enforce semantic consistency in the latent space—program classification (residential category), vintage classification (binned construction periods), and height regression.
• Loss Function: The total loss is the sum of VQ-AE reconstruction and codebook terms, plus a weighted aggregation of DTP task losses:

$$L = -\log p\bigl(x\,|\,z_{q}(x)\bigr) + \bigl\|\operatorname{sg}\bigl(z_{e}(x)\bigr) - e\bigr\|_{2}^{2} + \alpha L_{DTP}$$

• Clustering and Archetype Discovery: The vector-quantized latent representations of all buildings are clustered (k-means, elbow criterion for k), yielding local archetypes as the exemplars nearest to each centroid.
• Evaluation: In LA County, MARL-derived archetypes enabled annual energy use intensity (EUI) simulations with significantly lower error than conventional DOE Prototype Building Models, improving accuracy from ~72% to ~93–98%, depending on configuration and test scenario.
• Locality Capture: MARL’s latent space reflects both district-specific morphology (e.g., the distinction between single-family and row-house typologies) and programmatic/temporal attributes enforced via DTP.

These elements constitute a pipeline that directly links local geometric structure to the performance of downstream energy modeling, replacing high-level generalizations with 'real' local archetypes (Zhuang et al., 2023).

3. Quality-Diversity Urban Design: ARCH-Elites

ARCH-Elites exemplifies Local Real Architecture at the urban-configuration scale through a MAP-Elites-based illumination algorithm, powered by surrogate models for rapid evaluation (Galanos et al., 2021). Key aspects include:

• Feature-Grounded Search Space: Each urban layout is modeled as a collection of building polygons with height and static/mutable status.
• Behavioral Descriptors: The feature space for illumination comprises wind comfort metrics: the percentage of open-space with 'comfortable' wind (<6 m/s) and total area with 'dangerous' wind (>15 m/s).
• Quality Function: Floor Space Index ($\mathrm{FSI}$) quantifies the intensity of use per plot.
• Surrogate Evaluation: A pretrained 'InFraRed' model rapidly approximates wind comfort outcomes, circumventing the computational cost of full CFD.
• Mutation Operators: Geometric blending and constrained height mutations facilitate exploration of layout alternatives within static and variable zoning envelopes.
• Quality–Diversity (QD) Archive: A 20×20 grid bins solutions by behavioral descriptor; for each bin, the highest-quality layout is retained.
• Local Applicability and Transfer: The pipeline handles arbitrary urban settings through GIS/OSM ingestion, scalable inference, and interpretable QD trade-off maps highlighting the interplay between density and comfort.
• Empirical Findings: For two Boston districts, ARCH-Elites raised maximum FSI by 8–19% and demonstrated coverage rates of 0.29–0.46 for behavioral-space exploration.

ARCH-Elites enables architects and planners to extract locally optimal, context-specific urban arrangements, guided by region-specific environmental and regulatory constraints (Galanos et al., 2021).

4. Locally-Oriented Programming for Distributed Computation

Local Real Architectures also encompass computational models that prioritize spatial and memory locality in high-performance computing (HPC) (Rasmussen et al., 2015). The locally-oriented programming (LOPe) paradigm features:

• Local Element and Halo Formalism: Stencil-based computations are programmed as per-element kernels, with write access limited to the central element and read-only access to a specified 'halo'—a hyperrectangular neighborhood.
• Language Extension: Fortran arrays are extended with HALO annotations (e.g., real, dimension(:,:), HALO(1:*:1,1:*:1) :: U) defining left/right halo sizes per dimension.
• Distributed Data Decomposition: Arrays are partitioned across a multi-level memory hierarchy—inter-node via MPI ranks/coarrays ('images') and intra-node via accelerator subimages (GPUs).
• Halo Exchange: Communication of overlapping halos is explicitly handled via intrinsic operations such as HALO_TRANSFER, abstracting over MPI and accelerator data movement.
• Compiler Transformations: Source-to-source translation maps LOPe constructs to OpenCL or MPI/OpenMP code, and guarantees algorithm–communication separation.
• Advantages and Limitations:

| Advantage | Limitation | |-------------------------------------|--------------------------------------------| | Separation of compute/comm | Only structured Cartesian grids | | Single-writer property (avoids races) | No direct AMR or unstructured support | | Compact, mathematical notation | No composition of multi-stage operators |

No quantitative benchmarks are reported, but the methodology offers a clear pathway for aligning computational kernels with locality and ecological realism in spatial data or simulations (Rasmussen et al., 2015).

5. Evaluation and Metrics

Quantitative and qualitative evaluation in Local Real Architectures centers on fidelity to real-world performance, robustness across contexts, and efficiency:

• MARL: Accuracy improvements of +20–24 percentage points over conventional models for energy consumption estimation, both in closed and open-set tests (Zhuang et al., 2023).
• ARCH-Elites: Gains of up to +19% in density (FSI) and QD coverage rates of up to 0.46, with computational feasibility for large geographic areas due to surrogate inference (Galanos et al., 2021).
• LOPe: Evaluation details are not supplied; primary claimed metrics are language simplicity, correctness guarantees, and architecture transparency (Rasmussen et al., 2015).

A plausible implication is that as fidelity to local instantiations improves, derived simulation or design outputs attain higher ecological validity and operational impact.

6. Implications, Scalability, and Prospects

The introduction of Local Real Architectures prompts several implications and possible developments:

• Improved Modeling Accuracy: By substituting context-agnostic prototypes with local archetypes, simulation outcomes become more predictive of real-world energy performance or environmental comfort.
• Automated Archetype Discovery: Data-driven pipelines such as MARL obviate manual archetype specification, enabling scalable deployment across jurisdictions and building stocks.
• Transferability and Generalization: Frameworks demonstrate the ability to generalize across unseen regions, suggesting robustness for metropolitan-scale applications.
• Future Extensions: Identified directions include support for adaptive mesh refinement, multi-scenario surrogate modeling (resilience to flood, heat, seismic), and richer regulatory/participatory constraints (Galanos et al., 2021, Rasmussen et al., 2015).
• Limitations: Some models support only regular grids or binary/static constraints; extensions to unstructured environments and multi-objective trade-offs remain active research areas.

In sum, Local Real Architectures represent an emerging, empirically grounded paradigm that fuses local data, semantic encoding, and scalable algorithms to bridge the gap between generic templates and real, context-sensitive architectural models in both simulation and design.