Current tools for AI-enabled urban design, including Digital Twins, smart-city dashboards, and sensor networks, often optimize infrastructure and predict risk. What they often miss is harder to quantify but decisive: the drainage path a neighborhood has managed informally for decades, the market that anchors a block’s economy, the ritual site that has no address. Henri Lefebvre called this the gap between conceived space, the abstract city of experts and models, and lived space, the city as people actually inhabit, remember, and sustain it (Lefebvre 1991).

This paper introduces the Cultural Twin (CT) to address that gap. The CT is a model in which community knowledge, including vernacular practices, informal systems, and cultural memory, carries equal authority in scenario generation alongside engineering performance targets. We call this civic-encoded design: community claims are formalized as constraints the generative system must respect, not suggestions it can ignore.

Methodologically, the CT proceeds in three steps. First, civic and cultural data are gathered via fieldwork: story pins, walk-alongs, mapping exercises, and oral histories. Second, this material is translated into a Conflict Ontology, a set of Boolean, boundary, and relational constraints that govern what the model may or may not propose. Third, an Assumption Ledger records data sources, interpretive decisions, and trade-offs so outputs remain contestable and revisable.

We demonstrate the logic through a flood-prone district vignette. A conventional Digital Twin tends toward levees, floodplain compaction, and relocation, reducing risk on paper while erasing informal systems communities depend on. CT scenarios pursue comparable risk reduction while keeping markets, ritual spaces, and local housing typologies intact. The difference is not better data, but whose knowledge counts as binding constraint. The CT scales not by exporting forms but by sharing translation protocols, constraint-encoding methods, and governance structures. What works in one context becomes a template, not a blueprint.

The Architecture, Engineering, and Construction (AEC) industry is undergoing a transformative shift with the integration of AI-supported computational co-design tools. These platforms foster real-time, cross-disciplinary collaboration among architects, engineers, fabricators, construction professionals, and clients, enabling early-stage design processes that account for spatial, regulatory, material, and social factors. As a result, design outcomes are becoming increasingly innovative, context-sensitive, and construction-ready. This study investigates the practical implementation of co-design methodologies across disciplinary boundaries within the AEC sector. Using a combination of contemporary case studies and survey data from AEC professionals, the research examines how computational platforms and generative modeling tools facilitate more accurate, integrated, and efficient design workflows. Key findings indicate that computational co-design frameworks have potential to improve project timelines, cost-efficiency, stakeholder communication, and overall design quality. The research also identifies a broader industry trend: a departure from traditional, siloed design models toward iterative, collaborative processes. In addition, the study addresses the pedagogical and institutional challenges of embedding co-design into design and construction education. It suggests that overcoming these barriers is essential for the future adoption of more sustainable, efficient, and human-centered design practices in the AEC industry.

The global demand for rapid, sustainable post-disaster housing requires solutions that leverage local resources like bamboo. While deployable bamboo structures offer resilience and flexibility, their design is complex and computationally expensive, often requiring hours of parametric simulation. This research presents a data-driven artificial intelligence (AI) surrogate model that replaces slow optimization with instantaneous prediction, democratizing access to safe design. To identify the optimal tool, we conducted a benchmark of three machine learning algorithms—Enhanced Neural Network, Random Forest, and XGBoost—across two datasets: a general design space and a specific architectural application case. The benchmark identified XGBoost as the optimal predictive engine. The results revealed a critical "Complexity Gap": while the general unconstrained model successfully predicted geometric properties, it struggled to capture structural safety (R² ≈ 0.72). However, by constraining the problem to a specific topology, the application model achieved a R² of 0.982 and a reliability index (a20) of 0.936. Furthermore, analysis of the optimization frontier revealed a beneficial conservative bias, where the AI consistently underestimates the safety of hyper-optimized designs, acting as an implicit safety buffer. These results confirm that an empirically-tested, typology-based AI tool can provide local builders with structural assessment in real-time.