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

When the ground shakes during an earthquake, the soil beneath our feet can sometimes behave like a liquid. This phenomenon, known as soil liquefaction, occurs when loose, wet sand loses its strength and stiffness, causing buildings to sink, roads to buckle, and infrastructure to fail. It is one of the most destructive forces unleashed by seismic events, responsible for catastrophic damage in history's major quakes. To protect communities, engineers must determine before construction begins whether the ground at a specific site is likely to turn to liquid under stress. This assessment relies on two primary methods of probing the earth: the Standard Penetration Test, which measures how hard it is to drive a sampler into the ground, and the Cone Penetration Test, which pushes a cone through the soil to measure resistance continuously. Traditionally, interpreting these tests requires a geotechnical engineer to manually gather data, perform complex calculations, and decide which method applies, a process that is time-consuming and prone to human error.

A team of researchers from India, Australia, and the United States has developed a new system designed to make this critical safety check faster and more accessible. They created an automated framework that combines powerful computer learning models with a sophisticated artificial intelligence capable of understanding human language. Instead of forcing engineers to format data into rigid spreadsheets, this new system allows them to describe a site's conditions in plain English. The computer then reads the description, identifies the necessary soil and earthquake data, chooses the correct calculation method, and delivers a clear engineering report. The goal was not to replace human judgment, but to act as a smart assistant that handles the heavy lifting of data processing and model selection, allowing experts to focus on the final safety decisions.

The researchers began by teaching two separate computer programs how to predict liquefaction. One program was trained on nearly 2,000 real-world cases where the Standard Penetration Test was used, while the other learned from almost 500 cases involving the Cone Penetration Test. These programs, built on a technique called extreme gradient boosting, learned to recognize the complex, non-linear patterns between soil density, water levels, and earthquake intensity that lead to failure. When tested on new, unseen data, the program trained on the Standard Penetration Test correctly identified whether soil would liquefy or remain stable in 96 percent of cases. The program trained on the Cone Penetration Test achieved a success rate of 91 percent. Both systems proved highly reliable, with the Standard Penetration Test model showing a particular strength in spotting dangerous, liquefiable soil, a crucial trait for safety-critical engineering.

The true innovation of this work lies in how these prediction tools were connected to a large language model, a type of artificial intelligence designed to understand and generate human language. In previous attempts to use machine learning for geotechnical engineering, the computer models were isolated; they could only work if a human first extracted every single number from a site report and fed it into the system in a specific format. This new framework removes that barrier. The researchers built an "agent," a digital assistant that acts as the brain of the operation. When a user types a natural description of a site—mentioning details like soil depth, the force of the ground shaking, or the resistance measured by a probe—the agent reads the text, pulls out the relevant numbers, and figures out which of the two prediction models to use. If the description contains cone resistance data, it routes the request to the cone model; if it contains blow counts, it sends it to the penetration test model.

To ensure this system works in the real world, the team tested it with ten different scenarios, each describing a unique site condition. In every single case, the agent correctly extracted the required parameters from the text and chose the right prediction model to run. It then took the raw output from the computer model and translated it into a structured, professional engineering report that explained the risk and offered recommendations. This end-to-end workflow demonstrates that it is possible to bridge the gap between raw, unstructured field data and high-precision machine learning predictions without requiring the user to be a data scientist. The system effectively turns a complex, multi-step engineering analysis into a simple conversation.

However, the researchers are careful to define the limits of their creation. They emphasize that this tool is intended for preliminary screening and decision support, not as a final replacement for traditional analysis or professional judgment. The system is designed to catch potential hazards early and streamline the initial review process, but it cannot account for every unique geological nuance or the full scope of a site's history. For marginal cases where the risk is unclear, or when the input data is incomplete, the system is programmed to recommend further investigation by a qualified engineer. The study concludes that while the automated framework significantly improves the accessibility and efficiency of seismic hazard assessment, it works best as a partner to human expertise, handling the routine data work so that engineers can focus on the safety of the structures they build.