Generation of Granular Deposition Interfaces using conditional Generative Adversarial Network (cGAN)

TL;DR

The paper addresses the challenge of efficiently predicting granular deposition interface growth under varying media by training a conditional GAN with a 1D U-Net generator and a 1D ResNet discriminator to map random deposition interfaces to dynamically simulated interfaces from LAMMPS DEM data. It demonstrates that the model learns the complex mapping across water, acetone, and hexane, producing interfaces whose large-scale shapes and roughness evolution closely resemble those from MD simulations, while enabling rapid generation of hundreds of samples to stabilize statistics. The approach yields substantial computational savings (e.g., generating 500 profiles in under an hour on a multi-core CPU versus months for MD) and provides a scalable surrogate for studying KPZ-like surface growth in granular systems. This work highlights the potential of deep generative models to capture stochastic, physics-informed interface growth, offering practical benefits for statistical analysis and exploration of deposition dynamics.

Abstract

This work aims at generating 1D interface profiles of granular deposition by a conditional generative adversarial network (cGAN). Our cGAN model employs a U-Net generator and a ResNet discriminator that, in competition with each other, produce granular interfaces. The network is trained on dynamic simulation data from the LAMMPS granular package. Different fluids (water, acetone, and hexane) were used for the medium of the deposition cell to check the model performance in different growing conditions. The same model with the same hyperparameters was trained on data from different media separately. The ML-generated interfaces are compared with those of dynamic simulations, and a large number of interfaces are then produced to obtain more stable statistical properties of granular deposition. This way, the computationally extensive molecular dynamics simulation is substituted by the AI model. The statistical trend of interface growth is diagrammed, and the generated interfaces are also analyzed in terms of statistical features. Keywords: Conditional Generative Adversarial Networks, ResNet, U-Net, Granular Deposition, Interface Growth.

Figures (8)

• Figure 1: Architecture of the proposed cGAN. Left: U-Net-based generator. Right: ResNet-based discriminator.
• Figure 2: One simulation snapshot of deposition in the water medium. The noisy curve in the left image represents the random deposition interface at one time frame. At right, the interface generated by dynamic simulation for the same time frame is shown, and the corresponding model prediction of the interface is shown above it in red color.
• Figure 3: The upper image shows the growth of 240 interfaces during time for one dynamic simulation in water medium, and the lower image shows 240 of the corresponding model predictions in the test data. Different colors only represent different times of deposition.
• Figure 4: Roughness diagram of water medium for 50 dynamic simulations, 50 predictions of the model, and 500 predictions of the model.
• Figure 5: The upper image shows the growth of 240 interfaces during time for one dynamic simulation in acetone medium, and the lower image shows 240 of the corresponding model predictions in the test data. Different colors only represent different times of deposition.