Abstract
Simulating the response of granular materials under realistic loading scenarios is essential for ensuring the reliability of geotechnical infrastructure. This task is particularly challenging because natural soils exhibit inherently non-uniform particle arrangements due to gravitational sedimentation and are subjected to complex, multidirectional loading conditions from environmental forces such as wind and seismic activity. Unlike crystalline solids, there is no closed-form mathematical framework that fully describes soil's collective response. In engineering practice, this complexity is typically addressed using nonlinear constitutive models calibrated against laboratory data. However, such data are often specific to the site and material, influenced by variations in soil type, particle morphology, experimental apparatus, and loading conditions, making them difficult to generalize. The discrete element method (DEM) offers a unique pathway to overcome these limitations by providing direct access to particle-scale kinematics, contact forces, and evolving microstructure. As assemblies of particles exhibit chaotic rearrangements under loading, predicting their collective behavior becomes highly nonlinear and computationally intensive. Here, deep-learning models offer a promising route to replicate these complex relationships. In this work, we develop a deep-learning model using DEM simulations to address fundamental challenges in predicting the response of granular media under multidirectional loading paths, with direct applications to pressing engineering problems such as optimizing wind turbine foundations.