The integration of computational fluid dynamics (CFD) with deep learning in tunnel fire research is currently constrained by excessive reliance on manual operations and low overall efficiency. To address these limitations, this study presents a multi-agent collaborative framework driven by large language models, which enables full automation of the fire source characteristic inversion process. This framework reorganizes the conventional research pipeline into four dedicated, specialized agents: physical modeling, data governance, model training, and evaluation analysis. As a typical automated verification task, five deep learning models are systematically benchmarked under 45 experimental configurations to implement multi-task continuous regression inversion, which fully demonstrates the framework’s capability of automated, reproducible and large-scale comparative experiments. The experimental results demonstrated that the CNN-LSTM model outperforms other models in extracting spatiotemporal correlation features from temperature time-series data, enabling high-precision prediction of multiple fire parameters. With a 6 s observation window and 10 m sensor spacing, the average R2 attains 0.942, an improvement of 2% over the baseline LSTM model, and the RMSE decreases by 28.8%. For sparse sensor deployment at 30 m spacing, the average R2 remains at 0.917, confirming the effectiveness of integrating spatial feature extraction with temporal modeling. This study provides an efficient technical pathway for intelligent tunnel fire identification and advances the research paradigm by shifting the traditional manual optimization process to a multi-agent system-based optimization workflow.
Zeng et al. (Mon,) studied this question.