LLM-Powered Multi-Agent Framework for Automated PPV Prediction in Tunnel Blasting

Accurate prediction of blasting-induced peak particle velocity (PPV) is critical for assessing structural damage risk and ensuring safe tunnel construction. This study proposes an AI agent-based Evaluator-Optimizer workflow that automates the model-development pipeline from prepared dataset input through model training, performance evaluation, hyperparameter optimization, and ensemble construction, with limited manual intervention after dataset definition. The framework employs a multi-agent architecture comprising three collaborative agents—an Orchestrator, an Evaluator, and an Optimizer—supported by a large language model (LLM) reasoning layer. The Evaluator agent analyzes model performance across multiple metrics and generates diagnostic insights; the Optimizer agent translates these insights into structured optimization plans; and the Orchestrator coordinates the evaluate-optimize loop and stopping logic. The workflow was applied to a dataset of 102 tunnel blasting events. Nine candidate regression models spanning tree-based, kernel-based, neural network, and regularized linear families were trained and evaluated. The results show that the workflow enables three substantive observations: (i) across five tree-based models the powder factor is the dominant predictor (28.7–50.5% relative importance); (ii) under 50 Monte-Carlo repeated 80/20 splits, KNN and the Voting ensemble are statistically indistinguishable and form the most stable performance cluster, while Gradient Boosting lies within the same cluster with larger variance; and (iii) under nested 5 × 5 cross-validation, the un-leaked R2 for the top models is about 0.84–0.86, which quantifies the small-sample over-optimism that any future PPV study on single 80/20 splits should expect. The study therefore contributes both a portable agent architecture for tabular geotechnical regression and a concrete cautionary result about single-split benchmarking.