An LLM-Driven Agentic Framework for Automated Seismic Soil Liquefaction Hazard Assessment Using SPT- and CPT-Based Machine Learning Models

Soil liquefaction is a major earthquake-induced geohazard that can cause severe ground deformation and infrastructure damage, making rapid and reliable susceptibility assessment essential for geotechnical risk management. This study presents an automated decision-support framework that integrates Standard Penetration Test (SPT)- and Cone Penetration Test (CPT)-based machine learning models with a large language model (LLM) for natural-language-driven liquefaction assessment. Separate XGBoost classifiers were developed using 1,943 SPT and 479 CPT case histories incorporating key geotechnical and seismic parameters. The optimized models achieved testing accuracies of 0.96 and 0.91 for the SPT- and CPT-based datasets, respectively, with the SPT model showing particularly strong detection of liquefied cases. The predictive models were subsequently integrated with a Qwen2.5-7B-Instruct-based agent capable of extracting relevant site and seismic parameters from unstructured user inputs, identifying the appropriate SPT or CPT assessment pathway, executing the corresponding predictive model, and generating a structured engineering interpretation. Proof-of-concept evaluation demonstrated accurate parameter extraction and model routing across the tested scenarios. The proposed framework reduces manual data preparation and model-selection requirements while providing a practical conversational interface for preliminary liquefaction screening. The system is intended as an engineering decision-support tool rather than a substitute for conventional liquefaction analysis or professional judgement. The findings demonstrate the potential of combining field-test-based machine learning with LLM-assisted workflow automation to improve the accessibility and efficiency of seismic geotechnical hazard assessment.