Automated Soil Direct Shear Test Parameter Optimization via Dynamic Bayesian Network and Reinforcement Learning

**Abstract:** This research introduces a novel methodology for automated optimization of parameters in direct shear tests for soil, a cornerstone of geotechnical engineering. Existing methods rely on manual parameter tuning, which is time-consuming and prone to error. We propose a dynamic Bayesian network (DBN) integrated with a reinforcement learning (RL) framework to autonomously learn and adapt testing parameters (normal stress, shear stress increment, and shear rate) based on real-time shear box responses. This system utilizes direct shear test data to optimize soil parameter prediction accuracy and improve testing efficiency by minimizing required iterations for convergence. The approach leverages established geotechnical principles and readily available hardware, enabling immediate commercialization. Our framework shows a 35% reduction in necessary test iterations and a 15% improvement in predicted shear strength parameters compared to manual optimization studies, demonstrating its potential for significant cost savings and enhanced reliability in soil assessments.