Can AI Co-Design Structures? An AI Agent for Early-Stage Structural Design

1. Introduction

• Structural design is an iterative creative process that negotiates form, force, materials, and construction logic rather than merely satisfying performance constraints.Designers progressively develop the final result rather than automatically generating a final solution.Source: §1 · p. 1
• Recent text-to-3D systems generate visual 3D assets [17]; written-script work has begun exploring 3D CAD geometries and 2D FEA [18] [19] [22], while structural design requires joints, frames, areas, connectivity, and domain-specific metadata.Source: §1 · pp. 1–2
• The proposed workflow connects a language-model agent to a COMPAS-based modelling API through an MCP server to generate or edit graph-based models with explicit structural objects.The artifacts can be inspected, modified, serialized, and connected to downstream AEC workflows.Source: §1 · p. 2
• The proof-of-concept is demonstrated through prompt-generated examples, a language-based editing sequence, and progressive modelling, positioning editable structural artifacts as a basis for further research on natural-language design interfaces.The work frames these capabilities as a first step toward interactive natural-language interfaces for structural design.Source: §1 · p. 2

2. An Agentic Structural Modelling Workflow

• The agent interprets a prompt, uses MCP to communicate modelling actions, and supports follow-up instructions to review and adjust the resulting 3D structural model.The MCP server interfaces natural-language reasoning with deterministic modelling actions.Source: §2 · p. 2

2.1. A COMPAS-based graph data structure

• The COMPAS-based graph model represents structural elements, their connectivity, and associated metadata using semantically rich objects tied to structural engineering practice.The API includes joints, frames, areas, specialized beams, columns, braces, walls and slabs, along with material, section, grid, and level classes.Source: §2.1 · pp. 2–3

2.2. MCP as a bridge between AI agent and structural model

• An MCP server exposes the modelling API to the agent, whose calls execute deterministic operations and return the updated model state.The agent is pre-configured with MCP definitions and uses Google Gemini 3.1 Pro for reasoning on user input.Source: §2.2 · p. 3

2.3. Import and Export

• Generated structural objects can be serialized as structured JSON, and their geometric data can be imported or exported in common geometry formats.Connectors for common AEC software are planned as future work.Source: §2.3 · p. 3

3.1. Generating 3D Structural Models from Natural Language

• Prompts produced models ranging from a circular tower and L-shaped plan to terraced and rotating multi-story buildings, using structural vocabulary such as columns, beams, and slabs.The examples are stored as editable structural artifacts.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2. Editing object-based structural models

• A follow-up request extended the circular tower’s first three floors with beams and columns, then added slabs to those floors.The agent queried the model state before mapping the requested changes to modelling actions.Source: §3.2 · p. 4, Figure 3 · p. 5
• The workflow supports changes to an existing model, reflecting early-stage design in which new ideas are incorporated into an evolving structure.The passage connects this interaction to designers adapting models toward a final design solution.Source: §3.2 · p. 4
• The tower-editing example produced an overlapping slab at the circular footprint, so further instructions or manual edits were needed to correct that location.The passage attributes possible geometric errors to limited spatial reasoning or prompt ambiguity and calls for manual inspection and verification.Source: §3.2 · p. 5

3.3. Towards progressive structural co-design

• A sequence of prompts created two towers, connected them with a skybridge, and then added inclined cables to support the bridge.The designer developed the artifact over several prompts rather than providing the complete brief at once.Source: §3.3 · p. 5
• Progressive modelling resembles early structural design, where a simple organizational idea and preliminary model can become more complex over time.The example is described as a mode of use in which the designer develops the artifact incrementally.Source: §3.3 · p. 5
• Keeping the agent informed of the model state enables it to materialize intermediate artifacts for designers to inspect, revise, and redirect.The passage presents this capability as a prerequisite for future structural co-design workflows.Source: §3.3 · p. 5
• Reducing the effort of creating preliminary models could help teams pursue higher-performing design alternatives.The passage frames this as a potential benefit of reducing structural-modelling friction.Source: §3.3 · p. 5

• Prompt-based modelling can reduce repetitive setup effort, helping designers test structural ideas and explore alternatives more readily.The passage also states that explicit structural elements and access to the model state make the artifact inspectable, modifiable, and exportable.Source: §4 · p. 5
• A progressive example develops a pair of towers into a skybridge supported by inclined cables through successive prompts.This sequence illustrates modelling that becomes more complex over time.Source: Figure 4 · p. 6
• Future AI-assisted modelling could guide modifications, additions, or deletions toward lower material consumption and embodied carbon.The passage presents this as a potential sustainability application.Source: §4 · p. 6
• Future systems could combine text with sketches, annotations, reference images, or preliminary models to bring human and AI understanding closer.Source: §4 · p. 6
• The examples are proof-of-concept rather than a formal evaluation, and generated models do not guarantee structural correctness, code compliance, or analysis readiness.Prompt ambiguity, hallucination, misinterpretation, robustness across similar prompts, limited spatial reasoning, and the size and complexity of studied models remain concerns.Source: §4 · p. 6

• The workflow combines language-agent reasoning with a COMPAS-based structural modelling API exposed through an MCP server to generate editable early-stage building models.Source: §5 · p. 6
• Selected examples show that language agents can create structural artifacts that assist early structural design workflows.Source: §5 · p. 6
• The authors identify model size, analysis readiness, and intent misinterpretation as limitations of the preliminary work.Source: §5 · p. 6
• The proposed goal is to reduce friction between design intent and editable models toward more creative and efficient designs, rather than autonomous design.Source: §5 · p. 6