Evaluating generative AI for dimensionally correct and manufacturable CAD models
Generative AI now writes Computer-Aided Design (CAD) programs and boundary representations; however, the output that executes or resembles a target need not be dimensionally correct, constrained, or manufacturable. This review asks what evidence supports the engineering validity of generated parametric CAD. Journal articles from 2022 to 2026 were retrieved from OpenAIRE and Crossref, and records outside design geometry were removed. A bibliometric map of 60 articles was combined with manual coding of 19 core studies against a six-rung validation ladder, technical appraisal of ten full texts, and a structured synthesis of 14 matched contrasts. Output rose from three articles in 2022 to 2023 to 50 in 2025 and the partial year 2026, and mentions of feedback loops more than tripled in 2026. Of the coded studies, 14 reported shape fidelity and 10 execution success, but only 3 checked dimensions or explicit constraints, none reported functional analysis, and one measured a manufactured part. Three of six appraised studies with shape metrics normalized scale away. Feedback from interpreters, kernels, or a rule monitor raised execution or validity by 3.9 to 19.4 percentage points, but none was compared with resampling under an equal budget, and shape fidelity fell in four of five contrasts. The highest correction rates, 32.6 to 36.9 points, came from runs two thirds of which carried author-injected defects, certified by the monitor supplying the feedback. No study held out part families. Generative CAD has learned to produce valid programs; evidence that it satisfies engineering requirements remains thin.
Authors
- Dimvy Rusefani Asetya (ORCID: https://orcid.org/0009-0004-5569-5863)
Publication Details
- Journal
- Journal of Engineering and High Technology
- Published
- 2026-11-15
- DOI
- https://doi.org/10.55942/jeht.v2i2.2375
- Primary Topic
- Manufacturing Process and Optimization
- Type
- article
- Field-Weighted Citation Impact
- 0.00