A GraphRAG-Enhanced Multi-Agent Framework for Blast Design in Drill-And-Blast Tunnels

Blast design in drill-and-blast tunnelling requires geological interpretation and parameter selection for each excavation cycle, yet these tasks are often conducted separately, making it difficult to adapt plans to changing ground conditions. This study develops a multi-agent framework that integrates multimodal large language models with graph-based retrieval-augmented generation (GraphRAG) to connect geological assessment with preliminary blast-plan generation. A geological analysis agent, adapted through low-rank adaptation (LoRA)-based supervised fine-tuning on 1222 training cycles, combines tunnel-face images, ground-penetrating radar (GPR) profiles, and geological logs to classify the rock mass. A blast design agent uses the predicted class, cross-sectional geometry, planned advance, and site constraints to retrieve code provisions, design rules, and historical cases from a domain knowledge graph. It generates a structured preliminary plan with associated source and applicability records. On the held-out tunnel test set, domain adaptation increased classification macro-F1 from 78.6% to 89.7%. A seven-condition ablation supported the complementary contributions of the three modalities, with geological logs providing the strongest single-source performance. Across 60 design tasks under a common evidence-budget limit, GraphRAG improved blast-plan macro-F1 over conventional RAG by 4.1 percentage points with Qwen2.5-VL-7B-Instruct and 7.7 percentage points with MiMo-VL-7B. On an independent project, classification and blast-plan macro-F1 reached 82.4% and 84.0%, respectively, exceeding the unadapted classification model and conventional-RAG comparator by 7.5 and 4.4 percentage points. An illustrative Jinan Tunnel plan received an aggregate expert score of approximately 91. These findings support the framework’s use in integrating geological information and design evidence to prepare preliminary blast plans for engineering review.