建築智慧與 AI 的整合協作：建築師居中統籌雙端 AI 之 BIM 參數轉譯

Integrating Architectural Wisdom and AI: Architect-Mediated BIM Parameter Translation Coordinating Dual-End AIs

建築師累積的設計經驗，包括模式選擇邏輯、構造決策依據、以及法規與環境趨力間的權衡，多以隱性知識存在，難被資訊系統直接運用。有別於著重 AI 自動生成幾何或配置的「生成路線」，本研究探討一種「格式化路線」的協作模式：建築師扮演橋接角色，將自身設計判斷格式化，AI 分別在設計知識端與 BIM 模型端提供專業支援。以 Christopher Alexander《模式語言》為知識架構、Herbert Simon 滿足化理論為決策框架，開發一套將模式中的趨力（forces）轉譯為 IFC 參數約束的雛型系統，並以台灣社會住宅立面設計為驗證場域。

系統以「張力導航滿足化」（Tension-Navigated Satisficing, TNS）為理論架構，以 6 項定義與 3 條公理刻畫設計查詢作為異質知識空間中有限理性搜尋的特徵。建築師透過「橋接卡」（Bridge Card）格式化設計判斷，AI 知識顧問端從設計文獻檢索相關模式，模型解析端從 IFC 提取參數並執行 RASE 合規驗證。共建構 14 張橋接卡（含 3 張非住宅類），涵蓋 57 條趨力、36 組參數限制、35 條 RASE 規則與 17 對拮抗對。

實證結果顯示，系統在社會住宅 IFC 模型上成功完成從設計查詢到幾何參數提取的端對端流程。25 場景 × 8 條件消融實驗證明 AI 兩端能力互補且不可替代（向量檢索 ΔMRR = +0.169；橋接卡 ΔBHR = +0.787，均統計顯著）；微調嵌入呈現行為對齊，Recall@5 自 0.200 提升至 1.000，惟其反映模型對結構化推理的真正吸收抑或對訓練分布的記憶，尚待 held-out 任務驗證。跨棟別呈兩階段表現：幾何定義完整的 S1E（100%）、R5W（67%）可穩定運行，缺關鍵 IFC 構件的 C1 棟降至 0%；此斷崖落差源於前提崩潰而非推理鏈缺陷，量化了協作模式對建模品質的依賴邊界。橋接卡的條件演繹無法被向量空間近似（ΔBHR 正交於檢索指標），建築師的專業判斷是銜接設計知識與數位模型不可替代的環節。

 

Experiential design knowledge accumulated by architects—pattern selection logic, detailing rationale, and trade-offs among regulatory and environmental forces—largely remains tacit and inaccessible to information systems. Departing from the ""generative route"" in which AI produces geometry or layout directly, this study investigates a ""formalization route"": the architect serves as the bridge, formalizing design judgment while AI provides specialized support at both the design-knowledge end and the BIM-model end. Grounded in Alexander’s Pattern Language as knowledge framework and Simon’s satisficing theory as decision framework, a prototype system is developed that translates forces in pattern language into IFC parameter constraints, with Taiwanese social housing façade design as the empirical domain.

Underpinned by a Tension-Navigated Satisficing (TNS) framework of 6 definitions and 3 axioms, the system uses Bridge Cards through which architects formalize judgment: the AI knowledge-consultant retrieves patterns from design literature, while the AI modeling-agent extracts parameters from IFC models and performs RASE compliance validation. It comprises 14 Bridge Cards (3 non-residential), covering 57 forces, 36 parameter bindings, 35 RASE rules, and 17 antagonistic pairs.

Empirically, the system completes the end-to-end pipeline from design query to geometric parameter extraction on real IFC models. A 25-scenario × 8-condition ablation confirms the complementarity of AI’s dual ends (vector retrieval ΔMRR = +0.169; Bridge ΔBHR = +0.787, both significant). Fine-tuned embeddings exhibit behavioral alignment, raising Recall@5 from 0.200 to 1.000, though whether this reflects genuine absorption or memorization awaits held-out validation. Cross-building verification shows two-stage performance: stable on geometrically complete buildings (S1E 100%, R5W 67%) but collapsing to 0% on C1, which lacks key IFC components—a cliff-edge gap arising from prerequisite collapse rather than reasoning failure, quantifying the model’s dependence on modeling quality. Bridge’s conditioned deduction cannot be approximated in vector space, confirming the architect’s judgment as an irreplaceable link between design knowledge and digital models.