An LLM-based multi-agent framework for automated reliability analysis of spatially variable soils

Due to the inherent variability of soil properties, the Monte Carlo method is commonly used to conduct reliability analysis. To accelerate this process, time-consuming numerical simulations are often replaced with efficient surrogate models. However, current machine learning-based approaches for building these surrogates involve complex model construction, training, and hyperparameter tuning processes. Furthermore, accelerating reliability analysis via machine learning typically results in a fragmented workflow—requiring interaction with various software platforms, complex data transfers, and a heavy reliance on expert experience. Leveraging recent advancements in Large Language Models (LLMs), this paper proposes a hierarchical, LLM-based multi-agent framework for automated reliability analysis. We first demonstrate this approach using slope stability analysis as a primary example, and subsequently validate its generalization capabilities across other problems. The framework establishes a comprehensive workflow consisting of an Orchestrating Agent alongside three sub-agents responsible for dataset generation, surrogate model training, and large-scale prediction, respectively. The feasibility and accuracy of this framework have been validated across realistic scenarios of varying complexity. Experimental results demonstrate that the LLM-based multi-agent framework significantly reduces model construction time and minimizes the need for human intervention, while autonomously training accurate machine learning models. Additionally, this paper investigates the effects of self-correction, hyperparameter tuning, and prompt design on the performance of the proposed framework. The self-correction mechanism improves code-execution robustness, hyperparameter tuning enhances surrogate-model accuracy, and structured prompt design improves token and time efficiency during code generation. Ultimately, this study highlights the potential of LLM-based agents in advancing geotechnical engineering toward more intelligent and automated geotechnical reliability analysis.