WITHDRAWN: Integration of Natural Language Processing and Large Language Models for Automated SysML Generation with Topological Robustness Benchmarking in MBSE

The current generation of aerospace systems is getting much more complex and that pace continues to accelerate. This demands Model Based Systems Engineering (MBSE) workflows that turn text based documents into traceable SysML components while preserving design structure and minimizing manual effort. This work generated an AI automated framework that converts text based system descriptions into Papyrus compatible SysML components, eliminating a time consuming portion of early phase modeling tasks. This paper presents two extraction pipelines: (1) a rule-based NLP path built on spaCy and (2) a Large Language Model (LLM) path that uses prompt engineered queries to GPT-4o. Both pipelines output Papyrus standard .uml, .di, and .notation files. A gold standard manual model was introduced and verified by the authors to be reliable for current standard set from reputed sources. Then these pipelines were tested on a GPT-4o generated synthetic reference set of SysML models which was curated from the gold standard upon which further analyses are conducted. The study outlines the system architecture, requirement decomposition, and evaluation protocols. It shows that the LLM pipeline delivers higher recall on complex relationship structures, while the NLP pipeline offers deterministic execution time and fine grained controllability over the reference models. Building on this foundational analysis, the study conducted additional experiments which include a 100 system synthetic aerospace corpus with reference SysML models, a chaos monkey perturbation engine that injects lexical/semantic/structural noise, and a Graph Edit Distance (GED) metric to capture topological fidelity beyond token level F1. Present SOTA LLMs were benchmarked under clean and noisy conditions and an agentic Reflexion loop was implemented for self correction. Topological fidelity remained stable under perturbation (NLP: mean GED 23.1 ± 7.8 clean vs 20.6 ± 7.3 noisy, p = 4.3 × 10−4; LLM: 16.0 ± 6.8 clean vs 15.7 ± 5.4 noisy, p = 0.51), while Reflexion improved 19% of samples but degraded 54% (mean ΔGED = +1.13, p = 0.007). In the nine model benchmark, Grok-4 (0709) achieved the lowest GED (28.2), and Gemini 3 Pro delivered the best cost intelligence trade off (GED 33.2 at $2); Claude Sonnet 4.5 was mid at $3. These results establish GED as a structural complement to F1. Overall, the framework reduces modeling effort from hours to minutes, a significant advantage in AI assisted MBSE.