Self-Improving CAD Generation Agents with Finite Element Analysis as Feedback

Computer-aided design (CAD) is the backbone of modern industrial design, yet learned CAD generators still fall short of real engineering pipelines: they neither iterate like engineers nor evaluate what engineering requires. Prior work has treated CAD generation as two disjoint steps, part synthesis and assembly, where the former is graded by proximity to a gold reference and the latter, when handled at all, is reduced to a separate constraint solving step. In this work, we introduce a more industry-native task formulation that requires a model to produce a fully assembled multi-part STEP file from a free-form engineering brief, which is then validated via finite element analysis (FEA). FEA validation reveals that Codex (GPT-5.5) and Claude Code (Opus-4.7) agents do not produce a single strict-passing artifact in the main first-attempt sweep, with the best configuration meeting only about 20% of typed requirements on average. Moreover, we introduce two additional supervision signals, a novel text-only blueprint schema and a 21-view image renderer that aids the agent's visual inspection, that better align the generation loop with how engineers iterate in practice. On S2O and Fusion360, the same feedback tools improve geometric reconstruction, with GPT-5.5/xhigh rising from 0.444 to 0.592 Box-IoU on S2O and from 0.397 to 0.505 on Fusion360. Together these signals move CAD programs toward artifacts that are not only visually plausible but also checked against physical and structural requirements.

📋 한 줄 요약

**[CAD 생성 / FEA 피드백]** Free-form brief에서 fully assembled multi-part STEP을 생성하고 FEA로 validate — Codex·Claude Code의 strict-pass 0%, blueprint·21-view renderer 피드백으로 S2O Box-IoU 0.444→0.592, Fusion360 0.397→0.505.

🎯 핵심 기여도

CAD가 modern 산업 design의 backbone이지만, 학습된 CAD generator가 real engineering pipeline에 미치지 못함 — engineer처럼 iterate하지도, engineering이 요구하는 평가를 수행하지도 못함을 진단.
선행 연구가 CAD 생성을 part synthesis(gold reference 근접도로 채점)와 assembly(별도 constraint solving)의 disjoint 두 단계로 다뤘다는 한계 지적.
More industry-native task formulation 도입 — free-form engineering brief에서 fully assembled multi-part STEP file을 만들고 FEA로 validate.
FEA validation 결과 Codex(GPT-5.5)·Claude Code(Opus-4.7) agent가 main first-attempt sweep에서 strict-passing artifact를 단 하나도 산출하지 못함, 최고 구성도 average 20% 정도의 typed requirement만 충족. 두 추가 supervision signal — novel text-only blueprint schema와 21-view image renderer — 도입.

💡 핵심 아이디어

CAD 생성을 진정한 engineering 작업으로 만들려면 single-shot part synthesis가 아니라 FEA validation을 포함한 iterative pipeline이 필요하며, blueprint schema(text)와 21-view rendering(visual)의 두 피드백 도구를 추가하면 frontier agent도 geometric reconstruction에서 유의미 향상을 보인다.

🔬 기술적 접근법

**방법론**: Industry-native task formulation + FEA validation + dual supervision signal.
**핵심 기법**: (1) Free-form engineering brief → fully assembled multi-part STEP file 생성, (2) FEA validation으로 physical·structural requirement 검증, (3) Novel text-only blueprint schema (engineer-style 명세 표현), (4) 21-view image renderer로 agent의 visual inspection 보조, (5) Engineer-like iterative loop와 정렬된 생성 루프.

📊 주요 결과

Codex(GPT-5.5)·Claude Code(Opus-4.7) main first-attempt sweep에서 strict-passing artifact 0개.
최고 구성도 typed requirement의 평균 20%만 충족.
Blueprint + 21-view 피드백으로 GPT-5.5/xhigh의 S2O Box-IoU 0.444→0.592, Fusion360 0.397→0.505.
Visually plausible 수준을 넘어 physical·structural requirement 검증된 CAD artifact로 이동.

💭 의의 및 한계

**의의**: CAD 생성 평가의 engineering-grade 표준 정립, FEA를 supervision으로 통합해 hallucination 방지, blueprint·multi-view 피드백의 효과 정량 입증. **한계**: Frontier agent도 strict-pass 0%로 절대 성능 미흡, FEA validation의 도메인·시간 비용, blueprint schema 설계 품질에 의존.

🚀 실용적 활용

산업 CAD agent의 평가·학습 표준.
Engineering-grade 생성 시스템의 supervision 설계.
FEA-in-the-loop AI 보조 design 도구.