Large Language Model Based Expert Assistant for Workflow Consistent NPR-DDA Analysis

NPR-DDA is a discontinuum method for analysing large deformation and failure in reinforced rock masses. Its practical application remains strongly experience-dependent due to the complexity of block modelling, parameter selection, contact treatment, and numerical stability control within a structured software environment. This study develops a large language model based expert assistant to support NPR-DDA analysis through knowledge-grounded guidance without modifying the numerical solver. The assistant converts natural language engineering descriptions into executable procedures consistent with PreDDA, DDAsolver, and PostDDA operations. Three implementations, DDA-GPT, NPR-DDA-DB, and NPR-DDA-YB, are deployed on different foundation model platforms while sharing a common domain knowledge repository. A Brazilian disc case and a circular slip slope case are used to examine procedural consistency and parameter guidance under distinct instability mechanisms. Results show that knowledge-supported interaction improves workflow reliability and maintains solver compatibility. Platform comparison indicates that multilingual coverage and response robustness vary with the underlying foundation model, whereas the core guidance logic remains stable. The study demonstrates the feasibility of integrating large language models into complex discontinuum analysis workflows.