Spatial semantic information model (SSIM): A review for a multi-dimensional data framework of AI-driven architectural design

With the rapid advancement of digitalization and intelligence, architectural design remains constrained by the dominance of geometry and the weakness of semantics and associations, limiting the interpretability and transferability of AI in design. This study addresses how to construct a building data architecture suitable for AI-based modeling and reasoning by focusing on semantic enhancement of architectural geometry models. The research employs bibliometric analysis and a comprehensive review of semantic enhancement techniques in architectural geometry to critically examine the limitations of geometric representations across the design–performance–construction chain. The study has three objectives: to systematically categorize semantic enhancement methods; to synthesize the semantic requirements across the domains of generative design, building performance, and intelligent construction, addressing challenges in their integration with existing geometric models; and to identify conditions necessary for AI-oriented geometry model data structure. The findings reveal three primary research trajectories: data-driven geometric representation, ontology and rule-based structured semantics, and cross-modal geometry-semantic integration. Within these trajectories, nine distinct methods have been identified. The synthesis of domain-specific requirements indicates that generative design requires semanticization of multimodal geometry; performance optimization relies on bidirectional coupling and real-time feedback between performance and geometry; intelligent construction depends on rule-based, executable semantics derived from models. Building on this foundation, the paper proposes the formal definition of the Spatial Semantic Information Model (SSIM) as a multilayer graph-structured semantic web. Centered on the minimal semantic unit, it achieves a unified geometric representation of the multidimensional domains of Form, Performance, Stream, Intent, and Rule. Through a graph–tensor bidirectional mapping, the framework further specifies key constraints, including semantic closure, cross-domain alignment, and data governance. This architecture establishes a computational and extensible semantic basis for natural language driven geometric generation, performance construction integration, and traceable intelligent construction, thereby providing theoretical support for research agendas and future developments across the entire building lifecycle.