Architectural Design with Large Language Models

Scientific goals

Identify high-impact intervention points within architectural design workflows that are amenable to augmentation by LLM-based agentic systems.Develop and prototype computational frameworks that enable LLM agents to operate within geometric design environments, equipping them with tools for the creation, manipulation, and analysis of architectural form. Establish novel methods for qualitative design evaluation by leveraging the descriptive and analytical capabilities of Large Language and Vision Models (LLMs/VLMs) as automated design critics.

Experiments

Experiment 1: General Generative Agent

Motivation

To accelerate early-phase design exploration by enabling the rapid translation of abstract, metaphor-driven architectural ideas into tangible 3D parametric models without requiring specialized coding skills.

Method

Implementation of a multi-agent framework where specialized agents sequentially translate a conceptual metaphor into a structured brief, generate a corresponding Python script for parametric modeling in Rhino/Grasshopper, and evaluate the resulting geometry.

Finding

The agent-based system successfully translated ambiguous, high-level design intent into actionable parametric geometry, demonstrating a viable workflow for rapidly expanding conceptual exploration in architecture.

Experiment 2: Layout design for multistory buildings

Motivation

Multi-story building layout design is a high-dimensional optimization challenge requiring simultaneous satisfaction of geometric, programmatic, and circulation constraints that can be formulated as Mixed-Integer Quadratic Programming (MIQP) problems.

Method

We integrate Multimodal Large Language Models (MLLMs) with MIQP solvers to automate the translation between natural-language briefs and mathematical constraints, orchestrate optimization runs, and perform qualitative evaluation through vision-language analysis of generated floor plans, rendered views, and adjacency graphs.

Finding

Our framework demonstrates MLLMs' effectiveness in bridging design intent with mathematical optimization while preserving designer agency, making sophisticated layout generation more accessible to architectural practice.

Experiment 3: Facade Design

Motivation

To develop an LLM agentic framework that generates contextually-aware and structurally-coherent facade concepts, moving beyond abstract geometries to designs with greater architectural integrity.

Method

The method utilizes a material taxonomy derived from a reference image to guide the generation of a parametric script, which is then validated by a vision transformer and refined through an autonomous feedback loop.

Finding

The experiment revealed that design improvement hinges more on the material coherence of the initial concept brief than on code-based correction, as the specificity of this input dictates the architectural quality throughout the entire generative pipeline.