LLM-Driven CAD: From Script Generation to a Physical-World Entry Point

    Large language models are beginning to participate in CAD modeling and design, moving beyond text and image generation. This article examines their capabilities and limits in CAD, the path from script generation to validated design execution, multimodal and industrial data requirements, digital-twin use cases, a hypothetical valuation example, and practical integration options for developers.

    At the intersection of industrial software and artificial intelligence, a shift is underway: large language models are no longer limited to generating text or images. They are beginning to participate in modeling and design for the physical world. CAD connects digital models with physical manufacturing, and it is evolving from an interactive drafting tool into a platform that can understand intent, plan, and execute operations. Based on current technical progress, this article discusses the capability boundaries, implementation paths, and business implications of large language models in CAD.

    1. From Script Generation to Design Execution

    Early large language models played a limited role in industrial software. They were mainly used to generate parametric scripts or simple macro commands. With improvements in multimodal understanding and multi-step reasoning, a new generation of models can write CAD modeling scripts from natural-language descriptions and call modeling interfaces to perform the following work:

    • Asset construction and refinement: Generate basic geometry, surfaces, or feature trees, and set parameters such as dimensions, tolerances, and surface roughness according to constraints.
    • Part-level modification: Perform operations such as Boolean operations, chamfering, and shelling on individual parts in an existing model, while checking whether assembly relationships and topology remain consistent.
    • Layout adjustment: Adjust part positions, orientations, and interference relationships based on simulation results or assembly requirements, and generate multiple alternative schemes.

    These operations are not automatically safe. Before execution, the system needs API permissions for the CAD kernel. After execution, geometric validity checks are required, such as checking whether solids are closed and whether self-intersecting faces exist. If a step fails, the system should roll back to the last stable version and hand the issue to an engineer for review. The model can understand “what to do,” but whether it “does it correctly” still depends on deterministic validation rules and human judgment.

    For developers, secondary development workload may be significantly reduced. Tasks that traditionally required writing large numbers of API calls may in the future be completed with fewer descriptions. However, this effect needs to be measured in specific scenarios. For example, using the same batch of repetitive modeling tasks as samples, comparing average completion time and rework rate before and after the change can produce reliable conclusions.

    2. Multimodal Data and Industrial Knowledge

    The core of CAD software lies in complex data such as geometric topology, material properties, and manufacturing processes. A general-purpose large language model without industrial-domain training is prone to “hallucinations,” such as generating geometry that cannot be manufactured or parameters that violate basic physical laws. There are two common approaches.

    1. Multimodal alignment: Map text descriptions, engineering drawings, 3D point clouds, and simulation results into the same semantic space, enabling the model to infer 3D structures from 2D drawings or identify possible optimization directions from stress distribution diagrams.
    2. Domain data training: Use historical design files, standard part libraries, and process routes for domain adaptation. The model can learn implicit rules from a large number of successful and failed designs. This process is often compared to “digital apprentice” learning.

    It is important to note that historical design data often involves corporate secrets and customer information. Before use, the scope of data authorization should be confirmed, anonymization should be completed, and it should be clarified whether the data is used for training, inference, or both. Data without a compliance basis should not enter the training process even if it is technically available.

    In practice, many enterprises choose to cooperate with technology companies focused on industrial knowledge graphs to study issues such as neural network encoding of 3D CAD kernels and graph neural network reasoning for assembly constraints. This “general large model plus vertical fine-tuning” route is a common foundation for implementing “2D + 3D + cloud” solutions. “2D + 3D + cloud” means retaining the drafting efficiency of 2D engineering drawings while supporting 3D parametric modeling, and enabling multi-user collaboration, version management, and elastic computing resource allocation through the cloud. In this setup, a large language model can act as a dispatcher, deciding when to use 2D representation, when to switch to 3D verification, and when to synchronize the model to the cloud for rendering or simulation.

    3. From Design Tool to Digital Twin

    CAD with large language model participation is seen as an entry point to the physical world because it can handle information flows in two directions at the same time. The table below lists two typical paths.

    Workflow Input Model action Validation point Failure handling
    Forward design Functional requirements, such as withstanding a 5 kN static load Generate a CAD model and output manufacturing files Whether finite element simulation results meet the safety factor Revise constraints, regenerate, and record the reason
    Reverse reconstruction Physical point-cloud scan data Rebuild a parametric model Whether reconstruction error is below the tolerance threshold Switch to manual modeling or add scans

    In forward design, the CAD model is a digital description of a physical entity. In reverse reconstruction, the model converts scanned physical objects into editable parametric structures. Combined with sensor data and real-time simulation, CAD models can also develop into digital twins that reflect equipment status and support predictive maintenance. However, most digital twin projects are still in the pilot stage, and data access, model update frequency, and long-term maintenance costs all need to be evaluated separately.

    From an industry perspective, the global CAD market has long been dominated by a small number of closed-source software providers. Natural-language interaction may lower the barrier to professional operations and make it easier for small and medium-sized enterprises and individual users to complete basic design work. This could expand the user base, but whether the market structure truly changes still depends on data openness, interface standardization, and the response strategies of major vendors.

    4. Valuation Analysis: A Hypothetical Case

    The following hypothetical figures illustrate how a valuation calculation can be performed. These numbers are examples only and are not data from any real company.

    Assume a mid-sized industrial software company has a market capitalization of RMB 3.0 billion, cash on the books of RMB 1.3 billion, annual revenue of RMB 480 million, and net profit attributable to parent shareholders of RMB 110 million. The ex-cash market capitalization is RMB 1.7 billion (RMB 3.0 billion minus RMB 1.3 billion), corresponding to a price-to-earnings ratio of approximately 15.5x (RMB 1.7 billion divided by RMB 110 million). This calculation does not deduct non-recurring gains and losses. In actual analysis, the profit basis and reporting period need to be verified.

    Whether this multiple indicates an undervalued valuation depends on growth expectations. Traditional P/E valuation assumes that the business grows at its existing pace. If large language model features bring new charging models, the growth curve may change accordingly. Common analysis directions include the following three points.

    • Willingness to pay: If AI features are charged by number of calls or as value-added modules, revenue per user may increase. Paid conversion rate and renewal rate need to be observed.
    • Data accumulation: Design data generated by users can be used to improve models, but the premise is that the data is compliant and the improvement effect can be quantified.
    • Application expansion: Similar models may extend into architecture, machinery, electronics, and other fields, but each field requires separate data and validation.

    Risks are also obvious. Large language model training costs are high, access to industry data is constrained by compliance requirements, and integration with existing CAD kernels requires long-term investment. Investors can monitor the ratio of R&D investment to revenue, the scale of the partner ecosystem, disclosure of data assets, and changes in the share of AI feature revenue in total revenue. This content is for information exchange only and does not constitute investment advice.

    5. Developer Practice Guide

    Depending on their level of investment, developers can integrate at different levels of commitment. Two practical entry points are:

    1. Low-barrier layer: Use a cloud CAD platform that supports natural-language input to generate basic models through a conversational interface. This is suitable for prototype validation and teaching.
    2. Intermediate layer: Call the official SDK or scripting engine to embed models into existing workflows, for example, automatically generating sketches in parametric software.