Towards developing a language model based expert system for computational simulations

We present an expert system framework based on small language models (SLMs) for computational simulations (which for this work is a Søderberg electrode model). The framework is developed with the vision that the user can use prompts in natural language to perform actions (setup and run simulations) and seek information (from sources about the model/technology). Once the user provides the prompt, the back-end of the framework automatically performs the analysis and provides the required output for the user. This is enabled by using a combination of approaches like agentic workflows, retrieval augmented generation (RAG), intent based knowledge graphs and custom tools. Evaluation of the proposed framework has shown that it can provide reliable responses with minimal hallucination when the user is seeking information. The ability of the framework to automatically and accurately execute a range of actions based on the prompt (including prompts with typos) is also demonstrated. The proposed framework is deployed locally to protect the user's data privacy. We envision that such a framework, due to its ease of use, could enable or increase the use of computational models (especially fast models) by process operators as decision support tools during operations in industry.