Architecture modeling is an essential part of model-driven development of safety-critical software. AADL is a modeling language standard to design and analyze safety-critical software. However, there is usually a large gap between software requirements and architectural design. Effectively transforming requirements into formal software architecture models relies on a lot of manual experience and iterative exploration. In order to address these challenges, we conduct an exploratory study of using LLMs for fully automated AADL modeling. We assess three powerful LLMs, GPT 4o, DeepSeek-V3 and GLM-4-Plus. First, we decompose and refine high-level software requirements and design constraints, mapping them to the proposed prompt framework, RNL-Prompt, which significantly improved the accuracy of LLMs in generating different modeling elements. Our findings reveal that GPT-4o and DeepSeek-V3 perform on pair with each other, and both outperform GLM-4-Plus in complex modeling elements, such as modes, behavior annex, etc. To enhance the potential of GLM-4-Plus, we optimize its performance using N-shot prompting and retrieval-augmented generation (RAG). The results indicate that N-shot prompting performs more effectively. Finally, we demonstrate the effectiveness of our proposed approach in generating AADL architecture models in five examples of safety-critical domains. In addition, we implement a LLM-based modeling tool based on the AADL open source environment OSATE, which supports GPT-4o, DeepSeek-V3 and GLM-4-Plus. The tool is successfully applied to the modeling of an avionics control system in the industry.