Bridging Grain-Scale Fidelity and Continuum Efficiency on Ballasted Tracks using Diffusion Accelerated Workflows

Simulating granular media, such as railway ballast with Discrete Element Method (DEM) remains a prohibitively expensive task, especially during initialization phases, which dominate total runtime through iterative force-displacement solutions at sub-milli second time steps. Ballasted tracks are hence represented with overly simplified continuum models in traditional simulations to be compatible with conventional time domain solver pipelines, often losing important grain level characteristics. We introduce the first end-to-end generative pipeline based on diffusion models, that sidesteps the lengthy initialization phase by learning the distributions of fully settled assemblies and directly sampling physically consistent final states. The three-stage pipeline is: i) Synthesis of 3D Representative Volume Elements (RVEs) using unconditional diffusion models ii) Stitching of RVEs using diffusion inpainting, iii) Segmentation and exporting as DEM compatible libraries. A speed up of 200x is observed on 1.2 meter of statistically validated ballasted tracks, compared to conventional DEM initialization. Leveraging this generation backbone, we run a sleeper-track campaign comprising of ~5000 DEM samples with varying loading conditions and homogenizing them using a custom toolkit to get a dataset of continuum fields like macro and micropolar stresses/strains, that capture heterogenous details from DEM simulations. These fields become supervised targets for a neural operator that learns to output a rollout of these continuum fields given an initialized DEM sample, without ever doing a complete DEM simulation. Paired with fast granular sample generation, this physically consistent rollout gives near realtime load-response estimates for granular media applications, bridging grain-scale fidelity with continuum-scale computational efficiency.