Large Models for CAD: From Script Generation to Physical AI
This article explains how large language models are changing computer-aided design (CAD). It covers natural-language-driven modeling scripts, multimodal validation, industrial data integration, unified 2D and 3D cloud architecture, the practical challenges that remain, and where AI-enabled industrial software fits within the broader physical AI trend.
1. Background: Large Models Are Opening the Door to the Physical World
For the past two years, large language models have mainly worked on digital content such as text, code, and images. Since 2024, a clear trend has emerged: AI capabilities are moving from the digital world into the physical world, and industrial design software, especially CAD, is the most typical entry point.
The reason is straightforward. CAD connects design intent with physical manufacturing. Every engineering idea ultimately has to be expressed through geometric models, parametric constraints, material definitions, and assembly relationships. This workflow has long faced three major pain points:
- High barriers to entry: Professional CAD software has a steep learning curve. Becoming proficient often takes months or even years.
- Low efficiency: Repetitive modeling, part modifications, and parameter adjustments consume a large share of engineering hours.
- Difficult collaboration: 2D drafting, 3D modeling, and cloud collaboration have long been fragmented, and data is repeatedly converted between multiple tools.
Large models offer new solutions to these three pain points. The core logic is to elevate natural language to a first-class interface for industrial design. The model translates plain-language requests into professional operating instructions and executes them.
2. Core Capability One: Driving Modeling Scripts with Natural Language
The most mature and practically deployable form of AI-enabled CAD today is a large model that generates and runs professional modeling scripts directly. The technical workflow breaks down into four steps:
- Intent parsing: The user describes the design requirement in natural language, for example, “Generate a shell with rounded edges and a 2 mm wall thickness.”
- Code generation: The model translates the requirement into a script language supported by the CAD kernel, such as a parametric modeling script or a geometry description language.
- Sandboxed execution: The script runs in a controlled environment and produces the geometric model.
- Iterative feedback: Based on rendered results or user feedback, the model modifies individual components and adjusts geometric dimensions, materials, and scene layout.
An illustrative pseudocode example is shown below:
# User input: "Generate a rounded container with a 40 mm diameter and 60 mm height, with a 2 mm wall thickness"model = create_part(sketch=circle(diameter=40), extrude=60)model = apply_fillet(model, radius=3)model = shell(model, thickness=2)set_material(model, name="ABS Plastic")layout.place(model, position=(0, 0, 0), rotation=(0, 0, 0))render(model) # Output a preview for multimodal validation by the large model
The value of this code lies not in the code itself but in the shift in working mode it represents. Engineers move from being operators to being reviewers. The model builds and refines the design elements, and people judge whether the design is engineeringly sound. This improves efficiency and naturally expands how often CAD software is used and who uses it. People who could not model before can now complete a preliminary design through conversation.
Several key details matter in engineering practice:
- Parametric models are the best anchor point for AI. Compared with direct manipulation of mesh geometry, the constraint system of a parametric model can be reliably understood and modified by a large model, and errors are easier to trace back.
- The execution environment must be isolated. AI-generated scripts may contain invalid geometric operations or infinite loops. Resources must be limited in a sandbox, and outputs must be validated.
- A multimodal feedback loop is essential. Text feedback alone rarely reveals geometric defects such as interference or failed fillets. Rendered results need to be sent back to a multimodal model for visual validation.
3. Core Capability Two: Integrating Multimodal AI with Large-Scale Industrial Data
Script generation addresses the operational layer, but accumulated data is equally important. Part of the competitive barrier in industrial software comes from its geometry kernel, and the other part comes from accumulated industry knowledge, including standard parts libraries, process specifications, and typical structural cases.
The combination of multimodal large models and large-scale industrial data is changing how knowledge is acquired:
- From drawings to models: Image recognition and 3D reconstruction can rebuild 3D models from engineering drawings, physical photos, or scanned point clouds, which significantly reduces the cost of reverse modeling.
- From cases to recommendations: Based on historical design data, models can recommend similar reference structures and typical parameter ranges when a new project starts.
- From experience to rules: Much implicit engineering experience, such as “the wall thickness of this type of shell is generally no less than a certain empirical value,” can be structured and converted into design validation rules.
The common industry approach to building this capability uses cloud-based model services and data processing platforms. Companies feed de-identified design data into training or fine-tuning workflows and use general cloud infrastructure for large-model training, inference, and vector retrieval to build industry-specific AI capabilities. Essentially, this forms a data flywheel: the more the system is used, the richer the data becomes, and the more precisely the model understands the industry.
4. Architectural Evolution: An Integrated 2D, 3D, and Cloud Design System
For AI capabilities to deliver real value, the software architecture itself must also be modernized. Traditional industrial software is mostly standalone. 2D drafting and 3D modeling belong to separate product lines, and their data formats are incompatible. The key technical elements of an integrated design system include:
- A unified data foundation: A common geometric representation and data model connects 2D drawings with 3D models, so drawing annotations can directly drive 3D parameters.
- A cloud-native architecture: Compute-intensive tasks such as rendering, simulation, and large-model inference move to containerized cloud environments. The local client keeps only lightweight interaction, which enables cross-device collaboration.
- AI embedded across the full workflow: From sketch ideation (text-to-image or text-to-geometry), to modeling (script generation), assembly (intelligent constraint solving), and drawing output (automatic annotation), AI serves as a cross-cutting capability throughout the design lifecycle.
This evolution means CAD software is shifting from a tool to a design collaboration platform, and large models are the engine behind the platform’s intelligence.
5. A Realistic View: Deployment Challenges and Evaluation Criteria
Beyond the technology narrative, engineering realism is also necessary. Deploying AI in CAD still faces several practical challenges:
- Geometric accuracy and engineering reliability: A model generated by a large model can look correct but still be impossible to manufacture. Traditional CAx capabilities, such as interference checks and manufacturing process validation, are required as a safeguard.
- Data security: Industrial design data is a core asset for companies. Cloud solutions must provide private deployment or strict isolation mechanisms.
- Hallucination risk: The model may generate constraint combinations that cannot be realized geometrically. A transparent parameter validation layer is needed.
- Responsibility boundaries between humans and AI: Once AI-generated designs enter production, the process must clearly define who is responsible.
The maturity of an AI and industrial software solution can be assessed across four dimensions: the executable rate of generated scripts, the accuracy of multimodal validation, the coverage of industry data, and the depth of integration with existing design workflows.
6. Industry Perspective: The Long-Term Value of Physical AI
From an industry standpoint, “AI plus industrial software” is a typical physical AI segment. Its value does not lie in generating a paragraph of text or an image. It acts directly on R&D efficiency, manufacturing costs, and product iteration speed. The industrial software market has long been dominated by international vendors. Domestic substitution combined with AI-native restructuring gives this segment room for long-term growth.
Here is a simple valuation framework. When an industrial software company holds a high proportion of its market value in cash, and its price-to-earnings ratio excluding cash falls within a reasonable range, the option value of its AI transition, such as user growth, subscription revenue growth, and cloud service premiums, is nearly “free.” For stocks in this category, the key indicators are not short-term earnings elasticity but whether AI capabilities turn into measurable usage growth. Examples include the share of monthly active users of AI features, the adoption rate of script generation, and the growth rate of cloud revenue.
Conclusion
Large models empowering CAD essentially translates the professional language of industrial design into everyday natural language that more people can use, freeing engineers from repetitive work. The path runs from script generation and multimodal validation, to an integrated 2D, 3D, and cloud architecture, and then to industry knowledge built up through a data flywheel. The industry is gradually validating this path. The door to the physical world has opened, and the real competition is just beginning.
