deanlu.ai

Self-Improving CAD Generation Agents with FEA Feedback

15 min read

Source Snapshot

Garden Card

This note is a Quartz-ready system pattern for engineering agents. It shows how a model can generate CAD code while a deterministic controller validates the artifact with geometry checks, rich-view rendering, finite element analysis, typed feedback, and repair loops.

这篇笔记是面向工程智能体的 Quartz 系统模式。它展示了模型如何生成 CAD 代码，而确定性控制器负责几何检查、多视角渲染、有限元分析、类型化反馈和修复闭环。

  • Core question: How can an engineering agent move from plausible geometry toward validated engineering artifacts? 核心问题：工程智能体如何从“看起来合理的几何体”走向“经过验证的工程工件”？

  • Operational value: It turns validation evidence into targeted repair instructions and creates an auditable engineering record. 运营价值：它把验证证据转成定向修复指令，并形成可审计的工程记录。

  • Best connection: Agentic AI in Engineering and Manufacturing, Physical AI & Industrial Manufacturing, Core AI Platforms & Agents 最适合连接的内容：制造业智能体采用策略、物理 AI、企业智能体平台。

1. Executive Summary

The paper introduces an agent pipeline that converts a free-form engineering brief into an assembled STEP file and validates the artifact with finite element analysis. The agent writes CadQuery code, while a deterministic controller handles execution, rendering, meshing, simulation, requirement checks, and feedback routing.

这篇论文提出一个工程智能体流水线：将自由形式的工程需求转成装配后的 STEP 文件，并用有限元分析验证工件。智能体负责编写 CadQuery 代码；确定性控制器负责执行、渲染、网格划分、仿真、需求检查和反馈路由。

A structured blueprint and rich-view renderer help the agent inspect and revise the design. The benchmark remains difficult: frontier agents rarely produce fully valid artifacts on the first attempt, but repeated feedback-driven repair improves partial-credit performance and can eventually produce strict passes.

结构化蓝图和多视角渲染器帮助智能体检查并修复设计。这个基准仍然很难：前沿模型很少第一次就生成完全有效的工件，但多轮基于反馈的修复能提升部分得分，并最终产生严格通过的案例。

  • Main idea: Engineering-agent quality depends on the artifact-validation loop, not just the generated model or script. 主要观点：工程智能体质量取决于工件验证闭环，而不只是生成的模型或脚本。

  • Why now: CAD agents can create plausible geometry, but industrial use requires traceable checks for geometry, interfaces, clearances, load paths, stress, displacement, buckling, and metadata contracts. 为什么现在重要：CAD 智能体已能生成看似合理的几何体，但工业使用需要对几何、接口、间隙、载荷路径、应力、位移、屈曲和元数据契约进行可追溯检查。

  • Where it applies: Assisted CAD design, manufacturability checks, simulation-backed repair, engineering validation workflows, and controlled agent pilots. 可以应用的场景：辅助 CAD 设计、可制造性检查、仿真驱动修复、工程验证工作流和受控智能体试点。

Decision Signal

Put the agent inside a controlled engineering loop: explicit requirements in, auditable artifact out, deterministic validation, typed failure evidence, then targeted repair.

2. Key Technical Terms

Use stable Chinese translations for these engineering-agent concepts. Keep the English term first when it is the term people will search for later.

这些工程智能体概念建议使用稳定中文表达。如果未来检索时更常用英文术语，就把英文术语放在前面。

  • CAD generation agent / CAD 生成智能体: Agent that generates CAD programs or geometry artifacts from engineering requirements.

    根据工程需求生成 CAD 程序或几何工件的智能体。

  • Finite Element Analysis, FEA / 有限元分析: Numerical simulation method for stress, displacement, buckling, modal behavior, and related physical checks.

    用于应力、位移、屈曲、模态行为等物理检查的数值仿真方法。

  • Deterministic controller / 确定性控制器: Non-black-box control layer that executes code, calls tools, validates outputs, and returns evidence.

    执行代码、调用工具、验证输出并返回证据的非黑箱控制层。

  • STEP artifact / STEP 工程工件: Engineering artifact stored in a standard 3D product-data exchange format.

    以标准 3D 产品数据交换格式保存的工程工件。

  • CadQuery / 参数化 CAD 脚本工具: Python-based tool for creating parametric CAD geometry.

    用 Python 创建参数化 CAD 几何体的工具。

  • Typed feedback / 类型化反馈: Structured feedback by failure type, measured value, threshold, selector, load region, or repair scope.

    按失败类型、测量值、阈值、选择器、载荷区域或修复范围组织的结构化反馈。

  • Requirement checker / 需求检查器: Program that automatically checks whether an artifact satisfies geometric, physical, or metadata requirements.

    自动检查工件是否满足几何、物理或元数据需求的程序。

  • Repair loop / 修复闭环: Generate, validate, return evidence, modify, and validate again.

    生成、验证、返回证据、修改并再次验证的循环。

3. Core Notes

3.1 Problem

Plausible CAD geometry can still fail engineering constraints. A part can look correct in a rendered image while violating load paths, stress limits, displacement thresholds, clearances, interfaces, selectors, material assumptions, or metadata contracts.

看起来合理的 CAD 几何体仍可能违反工程约束。一个零件在渲染图中看似正确，却可能不满足载荷路径、应力限制、位移阈值、间隙、接口、选择器、材料假设或元数据契约。

  • Visual plausibility is not engineering validity. 视觉合理不等于工程有效。

  • First-shot generation is not enough for industrial use. 一次生成不足以支撑工业使用。

  • Engineering agents need validators that can produce repairable evidence. 工程智能体需要能产生可修复证据的验证器。

3.2 Mechanism

The agent writes CadQuery code and exports a STEP artifact. The controller creates isolated workspaces, executes code, runs geometry checks, renders rich views, launches finite element analysis, parses verdicts, and returns typed feedback to the agent for repair.

智能体编写 CadQuery 代码并导出 STEP 工件。控制器创建隔离工作区、执行代码、运行几何检查、渲染多视角图像、调用有限元分析、解析判定结果，并把类型化反馈返回给智能体进行修复。

  • The model proposes and repairs. 模型负责提出方案和修复。

  • The controller measures and governs. 控制器负责测量和治理。

  • The engineering record stores artifact versions, solver results, validator outputs, and repair decisions. 工程记录保存工件版本、求解器结果、验证器输出和修复决策。

3.3 Evidence

The paper introduces Hephaestus-CCX, a benchmark of 50 engineering briefs with executable requirement checkers. Requirements include stress, displacement, modal behavior, buckling, contact, clearance, selectors, and assembly constraints.

论文提出 Hephaestus-CCX，这是一个包含 50 个工程需求的基准，并配有可执行需求检查器。需求包括应力、位移、模态、屈曲、接触、间隙、选择器和装配约束。

  • In the main first-attempt sweep, 400 submissions produce no strict-passing artifacts. 主体首次尝试中，400 次提交没有产生严格通过工件。

  • After one FEA-feedback round, one strict pass appears across another 400 revised submissions. 一轮有限元反馈后，另外 400 次修订提交中出现 1 个严格通过。

  • One FEA-feedback round improves mean requirement pass by 13.4 percentage points on average across the reported model cells. 一轮有限元反馈使报告模型组的平均需求通过率平均提升 13.4 个百分点。

  • In the longest GPT-5.5/high run, mean requirement pass rises from 38.8% to 60.5%, with 9 strict-passing artifacts out of 50 cases. 在最长 GPT-5.5/high 运行中，平均需求通过率从 38.8% 提升到 60.5%，50 个案例中有 9 个严格通过工件。

Evidence Boundary

Treat the result as a promising engineering-assistance pattern, not proof of autonomous production readiness.

3.4 Boundary

The pattern is not production certification. Generated artifacts should not be used for safety-critical, regulated, or manufactured designs without independent review and validation.

这个模式不是生产认证。未经独立审查与验证，不应把生成工件用于安全关键、受监管或实际制造设计。

  • Validate solver configuration, meshing stability, selector binding, units, material properties, requirement provenance, and approval boundaries. 需要验证求解器配置、网格稳定性、选择器绑定、单位、材料属性、需求来源和审批边界。

  • Keep human engineering sign-off mandatory for high-consequence decisions. 高后果决策必须保留人工工程签核。

  • Use the pattern first where deterministic evaluators already exist. 优先在已经存在确定性评估器的流程中使用该模式。

4. Concept Map

Use wikilinks to connect this note into the broader Quartz graph.

使用双向链接把这篇笔记接入更大的 Quartz 知识网络。

flowchart LR
  A["Engineering Brief"] --> B["Typed Blueprint"]
  B --> C["CadQuery Program"]
  C --> D["STEP Artifact"]
  D --> E["Rich-view Inspection"]
  D --> F["Requirement Checks"]
  D --> G["FEA Simulation"]
  E --> H["Typed Feedback"]
  F --> H
  G --> H
  H --> I{"All Requirements Pass?"}
  I -- "No" --> B
  I -- "Yes" --> J["Human Engineering Review"]

Diagram labels stay in English for rendering consistency and easier reuse across published pages.

图中的标签保持英文，便于 Quartz 渲染后跨页面复用，也方便技术读者快速识别。

5. Source Visual

CAD-agent pipeline with blueprint, STEP assembly, rich-view inspection, and FEA feedback

The paper’s pipeline separates design decisions from execution control. The agent owns planning and CAD-code repair; the controller owns execution, measurement, composition, validation, and feedback routing.

论文中的流水线把设计决策和执行控制分开。智能体负责规划和 CAD 代码修复；控制器负责执行、测量、组合、验证和反馈路由。

Source credit: Figure 2 in the arXiv HTML version

来源说明：图片来自 arXiv HTML 版本中的 Figure 2。

6. Operating Pattern

The reusable system design is an evidence-producing controller around a model. This pattern is more operationally useful than a model-only benchmark because it defines where creativity, measurement, logging, and approval should live.

可复用的系统设计，是在模型外层放置一个能产生证据的控制器。这个模式比单纯模型基准更有运营价值，因为它定义了创造、测量、记录和审批分别应该放在哪里。

engineering_agent_loop:
  input:
    brief: free_form_engineering_requirements
    contract:
      - geometry
      - interfaces
      - selectors
      - physical_limits
  agent:
    owns:
      - blueprint
      - parametric_cad_code
      - repair_decisions
  controller:
    owns:
      - isolated_execution
      - artifact_export
      - deterministic_measurement
      - rich_view_rendering
      - meshing
      - fea
      - typed_requirement_verdicts
  retry:
    feedback:
      - failed_requirement
      - measured_margin
      - selector_or_load_case
      - recommended_repair_scope

Store every artifact version, validator result, solver version, requirement schema, and repair decision. This supports traceability, reproducibility, and controlled human approval.

保存每个工件版本、验证器结果、求解器版本、需求模式和修复决策。这样才能支持可追溯、可复现和受控人工审批。

Implementation Risk

Before using this pattern in production, validate solver configuration, meshing stability, selector binding, units, material properties, requirement provenance, and the approval boundary.

7. Quantitative View

xychart-beta
  title "GPT-5.5/high Mean Requirement Pass During Repeated Feedback"
  x-axis ["Early loop", "Longest reported loop"]
  y-axis "Mean requirement pass (%)" 0 --> 70
  bar [38.8, 60.5]

In the longest reported run, structured feedback and repeated repair increase mean requirement pass from 38.8% to 60.5%, with 9 of 50 artifacts achieving strict passes. The result is meaningful but still far from autonomous production readiness.

在最长报告运行中，结构化反馈和重复修复使平均需求通过率从 38.8% 提升到 60.5%，50 个工件中有 9 个严格通过。这个结果有意义，但距离自主生产可用仍有明显差距。

8. My Take

This paper gives a concrete blueprint for engineering agents: keep the model creative, but make the surrounding system deterministic, measurable, and auditable. The strategic lesson is that validation and repair loops may matter more than first-shot generation quality.

这篇论文给出了工程智能体的具体蓝图：让模型负责创造和修复，但让外层系统保持确定性、可测量、可审计。战略启示是：验证与修复闭环可能比一次生成质量更重要。

  • What changed my thinking: The controller is not plumbing; it is the governance layer that makes an engineering agent operational. 改变我理解的地方：控制器不是管道代码，而是让工程智能体可运营的治理层。

  • What I may do next: Identify one manufacturing workflow with an existing deterministic evaluator and design a small agent loop around it. 下一步可能行动：找一个已经有确定性评估器的制造流程，围绕它设计一个小型智能体闭环。

  • What still needs verification: Solver reliability, mesh stability, unit discipline, requirement schema quality, artifact storage, and human approval triggers. 仍需要验证的内容：求解器可靠性、网格稳定性、单位规范、需求模式质量、工件存储和人工审批触发条件。

Reuse Path

Convert this note into a controlled manufacturing-agent pilot where deterministic validators already exist.

References

Graph View

Backlinks