Feasibility Study on Reactor Preliminary Design via Fine-Tuning of Large Language Models

With the rapid advancement of large language models (LLMs) in knowledge integration, semantic reasoning, and information processing, their potential in chemical engineering and process systems design has become increasingly evident. However, systematic studies in chemical reactor design remain scarce, particularly frameworks integrating reaction conditions, reaction types, and material properties for preliminary material selection. This work investigates the feasibility of using LLMs for preliminary reactor design tasks, including residence time calculation, reactor volume estimation, and material selection, through model fine-tuning. We construct the first chain-of-thought fine-tuning data set tailored to preliminary reactor design, embedding multistep computational reasoning and material selection prompts. To address the gap between linguistic fluency and numerical correctness, a dual-evaluation framework combining semantic metrics with numerical accuracy is proposed. Five open-source LLMs (Qwen2.5-Omni, Qwen2.5-Math, DeepSeek-R1-Distill-Qwen, Gemma-3, and Llama-3.2) are fine-tuned and evaluated using both virtual and real reaction data sets. Results indicate that fine-tuned LLMs can reliably perform theoretical reaction volume calculations and preliminary material selection, achieving absolute relative error below 0.3 for over 50% of test cases across all models, while material selection accuracy exceeds 97%. Among them, DeepSeek-R1-Distill-Qwen and Qwen2.5-Omni exhibit superior performance, with a mean absolute relative error (MARE) of 0.206. Despite these improvements, LLMs remain less precise than traditional numerical methods in computation-intensive tasks, reflecting their text-oriented architecture. Analysis of reasoning behaviors and failure modes suggests that future work should adopt hybrid “LLM + tool” frameworks and expand data sets to include more complex reaction scenarios, enhancing the reliability of LLM-assisted reactor design.