LLMs in CAD: From Design Assistance Toward Physical-World Execution

    This article examines how large language models are changing CAD workflows, from natural-language-driven script generation and validation to multimodal industrial data, cloud collaboration, valuation analysis, and reliability risks. It focuses on practical boundaries: what AI can draft, what engineers must verify, and what teams should clarify before adopting execution-capable tools.

    From Assistance to Execution: Where Design Work Costs Come From

    In a traditional CAD workflow, designers must handle geometric constraints, parametric modeling commands, and material adjustment panels at the same time. Much of the time cost is often not the design idea itself, but the process of translating that idea into concrete modeling steps. This translation cost is exactly where large language models have the greatest opportunity to reduce effort.

    Most earlier intelligent tools stayed at two tasks: explaining how a command works, or providing a code snippet for engineers to paste and run themselves. A new generation of models can understand natural-language requirements, generate a complete sequence of modeling operations, and pass it to the software as a script for execution. The model’s role therefore moves from a “copilot” that offers suggestions toward a party that directly executes tasks.

    The stronger the execution capability, the greater the impact of errors. Before introducing such tools, teams should clarify three things: which software and version the script targets; what permissions the model has, such as whether it can read and write files or call a rendering engine; and how to roll back when an error occurs.

    Automated Script Generation: What It Can Do and How to Check It

    Automated scripts are suitable for tasks that are highly repetitive and clearly rule-based, such as building basic parts from scratch, batch-adding fillets or shells, generating parametric arrays, and modifying the dimensions or materials of a component in a scene. The table below lists which parts of common tasks are suitable for an AI first draft and which parts require engineer review.

    Task type Suitable for AI-generated first draft Focus for engineer review
    Basic part creation Sketch profiles, extrude and revolve features Dimension units, datum plane selection
    Fillets and shells Feature parameters and target objects Whether wall thickness meets manufacturing requirements
    Parametric arrays Array count and spacing Feature parent-child relationships; whether the array updates with the parent feature
    Material adjustment Initial assignment of material properties Whether roughness, metalness, and other parameters meet rendering or physical standards

    A reliable validation process usually includes the following steps.

    1. Run the script on a test copy without directly modifying the production model.
    2. Check key dimensions, units, and tolerances to confirm they match design requirements.
    3. Inspect the feature tree or operation history to confirm that parent-child relationships and update order are intact.
    4. Have an experienced engineer review it, archive it after approval, and retain the script version number.

    Multimodal Fusion: Turning Industrial Data into Model Capability

    Logic-based reasoning alone is not enough to cover the complexity of the physical world. Industrial scenarios contain large amounts of unstructured data, including scanned historical drawings, 3D point clouds, and simulation result heatmaps. To let models use this data, the key step is to map it into the same representation space so that drawings, 3D models, text annotations, and performance data can be linked to one another.

    The table below describes common processing methods and application boundaries for several data types.

    Data type Common processing method Suitable application Main limitation
    2D engineering drawing scans OCR, primitive recognition, annotation structuring Search historical drawings; assist annotation Recognition rates are lower for hand-drawn and old drawings, requiring manual verification
    3D point clouds Registration, denoising, meshing Reference models for reverse modeling Accuracy depends on scanning equipment
    Simulation heatmaps Extract structured indicators such as stress concentration zones Flag potential weak locations Conclusions must be verified by simulation re-run
    Text annotations and design notes Vectorization and semantic search Find similar design schemes Missed detection when terminology is inconsistent

    A model can reference historical simulation data for similar structures and suggest improvement directions. Such suggestions are only clues and cannot replace simulation validation.

    From Desktop to Cloud: Migration Considerations

    Traditional desktop software is limited by hardware, and real-time rendering and collision detection often struggle with very large assemblies. After rendering and solving are deployed to cloud containers, compute can scale on demand, and lightweight pre-processing and post-processing can be completed in the browser. Multiple designers can work on modeling in the same digital twin scene and then merge the results.

    This collaboration model raises several issues that must be considered in advance.

    • Data security: Drawings and models should be stored encrypted, transmitted over encrypted channels, and access should be divided by project.
    • Permission isolation: Data from different customers and projects should be isolated from one another, and the AI assistant should only access files required for the current task.
    • Copyright and compliance: Ownership, scope of use, and data export requirements for customer drawings should be agreed in advance and comply with local regulations.
    • Concurrency conflicts: When multiple people modify the same assembly at the same time, clear locking mechanisms or merge rules are needed.
    • Network and offline access: When on-site networks are unstable, local caching or offline operation should be retained.

    When selecting a solution, teams can first list these requirements and then compare them item by item against the actual capabilities of candidate solutions, rather than looking only at advertised compute metrics.

    Investment Perspective: Reading Valuations

    This section introduces an analytical method and does not constitute investment advice.

    When evaluating industrial software companies, a common approach is to first subtract net cash from the market capitalization to obtain the implied value of the core business, and then compare it with the revenue and profit of the main business. Suppose a company has a market capitalization of RMB 10 billion and net cash of RMB 4 billion. The implied value after deducting cash is approximately RMB 6 billion. If the company’s annual net profit is RMB 300 million, the corresponding static price-to-earnings ratio is approximately 20x. These figures are used only to demonstrate the calculation method and do not represent any real company.

    The following questions are worth focusing on during analysis.

    • How much of revenue comes from software subscriptions versus one-time licenses, and is the growth of each sustainable?
    • Is the R&D expense ratio rising along with AI investment, and can profit margins support this investment?
    • Is the commercialization progress of new features consistent with the timeline in announcements?
    • Can design data form a closed loop that feeds model training, or is it still at the planning stage?

    Valuation analysis should be based on the company’s latest periodic reports and announcements, and financial definitions should be checked, for example, whether non-recurring gains and losses are included.

    Risk warning: Commercialization of AI features may be slower than expected, R&D investment may continue to compress profits, and the prices of related stocks may be highly correlated with sector sentiment. Readers should make independent judgments based on their own risk tolerance.

    Reliability Risks: Hallucinations and Human-in-the-Loop

    Current models can still produce hallucinations during long-chain logical reasoning, which is why human-in-the-loop review remains essential.