A dual-retrieval augmented Multimodal Large Language Model framework for automatic post-earthquake building safety evaluation
Year: 2026
Type: article
Abstract: Rapid post-earthquake building safety evaluation is critical for disaster response, infrastructure rehabilitation, and public safety. However, traditional manual inspections remain labor-intensive, time-consuming, and hazardous. While current task-specific inspection models excel at damage detection yet fall short in reasoning, Multimodal Large Language Models (MLLMs) offer semantic reasoning but lack the domain-specific seismic engineering knowledge required for standard-compliant safety evaluation. To address this, this study proposes SeisMLLM, a novel dual-path retrieval-augmented MLLM framework that bridges visual damage perception and expert-level safety evaluation. To enable this framework, this work constructs SeisMLLM-1k, the first case-level multi-view dataset comprising 3,856 images across 1,306 building cases from 17 global earthquakes, with annotations strictly grounded in standard protocols, e.g., ATC-20-1 Field Manual: Postearthquake Safety Evaluation of Buildings. In this study, SeisMLLM explicitly decouples visual perception from safety judgment via a domain-finetuned MLLM, and grounds the reasoning process through a dual-path retrieval strategy encompassing both historical damage cases and the regulatory logic of ATC-20-1. An ATC-20-1 compliant assessment report is ultimately generated via chain-of-thought reasoning. Experimental results demonstrate that SeisMLLM effectively mitigates critical failure modes of general-purpose MLLMs, outperforming representative open-source baselines by an average of 38.5 percentage points in safety evaluation accuracy, and achieving a leading accuracy of 66.13%. Overall, SeisMLLM transforms black-box predictions into an interpretable, standard-compliant solution for automated building safety evaluation.
Source: Advanced Engineering Informatics
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,560
OpenAlex ID: W7202183463