Synthesizing realistic sand assemblies with denoising diffusion in latent space

Nikolaos N. Vlassis, WaiChing Sun, Khalid A. Alshibli, Richard A. Regueiro

TL;DR

Researchers have developed a generative AI pipeline that uses denoising diffusion models to synthesize thousands of hyper-realistic 3D sand grains. By training on a small dataset of real micro-CT scans, the model can "hallucinate" massive, statistically accurate sand assemblies for use in high-fidelity engineering and physics simulations.

What Problem Does This Paper Address?

In fields like geotechnical engineering, petroleum extraction, and even computer animation, the specific shape of a sand grain matters. The way grains interlock dictates how a soil slope fails, how oil flows through a reservoir, or how light scatters off a dune.

However, we face a "data bottleneck." Obtaining high-quality 3D geometry for sand requires Synchrotron Microcomputed Tomography (SMT)—an expensive, slow process that yields only a few hundred or thousand grains per scan. Previous mathematical shortcuts to "fake" sand (like using simple ellipsoids or Fourier descriptors) often fail to capture the sharp edges and irregular "DNA" of real grains. This paper aims to bridge that gap by using generative AI to create an infinite supply of realistic synthetic sand.

Methods

The authors propose a two-stage deep learning approach:

  1. The Point Cloud Autoencoder: First, they take 3D point clouds (600 points per grain) and compress them into a lower-dimensional "latent space." To help the AI understand how points relate to each other, they use an algorithm called node2vec to encode topological information (adjacency) into the input. This ensures the model doesn't just see a cloud of dots, but understands the "surface" of the grain.
  2. Latent Denoising Diffusion: Instead of generating 3D points directly (which is computationally heavy), they train a Denoising Diffusion Probabilistic Model (DDPM) to work within the compressed latent space. The model learns to take pure "noise" and iteratively refine it into a valid latent representation of a sand grain.
  3. Physics Assembly: Once the grains are generated and decoded back to 3D, the authors use a rigid-body physics engine (Blender) to simulate "pouring" the sand into containers, creating stable, settled assemblies ready for virtual testing.

Datasets

The primary source of "ground truth" is the F50 sand database. This consists of 1,542 real SMT scans of natural silica sand. While 1,500 grains is a lot for a human to scan, it is a small dataset for deep learning. The authors used data augmentation (random rotations) to effectively double this for training the autoencoder.

Key Insights and Ideas

  • Bypassing Hand-Picked Descriptors: Traditionally, scientists had to decide which features mattered (e.g., "roundness" or "aspect ratio") and try to program them. This model is "unconditional"—it learns the entire statistical distribution of shapes on its own, capturing nuances that humans might not think to measure.
  • Latent Space Efficiency: By performing the "diffusion" (the noisy-to-clean transformation) in a compressed 28x28 latent space rather than 3D space, the process is significantly faster and more stable.
  • The "Digital Twin" Capability: Because the model learns the underlying distribution, researchers can "sub-sample" the results. For example, they can ask the model to generate 50,000 grains but then only pick the ones that are exceptionally elongated to see how a "flaky" sand assembly behaves compared to a "round" one.

Key Results

  • Massive Scaling: The authors successfully generated 50,000 synthetic grains from a starting pool of just 1,542 real ones.
  • Statistical Fidelity: The synthetic grains were nearly indistinguishable from the real ones when measured by standard geological metrics. For instance, the $D_{50}$ (median grain size) of the database was 0.2269 mm, while the synthetic batch hit 0.2214 mm.
  • Morphological Accuracy: Using ternary plots for compactness, flakiness, and elongation, the authors showed that the "shape space" occupied by the synthetic sand perfectly overlapped with the real F50 sand.
  • Open Science: The authors have made the 50,000 synthetic grains and the original 1,542 scans available in an open-source repository for the research community.

Limitations and Future Work

While the shapes are realistic, the current model focuses on the macro-shape (the general silhouette). It does not yet capture micro-roughness or surface texture at the micron level, which can significantly impact friction between grains.

The authors also note that while the Blender physics engine is great for creating stable "piles" of sand, it isn't a substitute for a full Discrete Element Method (DEM) simulation that can calculate complex stress-strain relationships. Future work will likely involve integrating these AI-generated grains into high-performance geomechanical solvers to see if "AI sand" behaves exactly like "real sand" under high pressure.

Summary updated

Abstract

Abstract The shapes and morphological features of grains in sand assemblies have far‐reaching implications in many engineering applications, such as geotechnical engineering, computer animations, petroleum engineering, and concentrated solar power. Yet, our understanding of the influence of grain geometries on macroscopic response is often only qualitative, due to the limited availability of high‐quality 3D grain geometry data. In this paper, we introduce a denoising diffusion algorithm that uses a set of point clouds collected from the surface of individual sand grains to generate grains in the latent space. By employing a point cloud autoencoder, the three‐dimensional point cloud structures of sand grains are first encoded into a lower‐dimensional latent space. A generative denoising diffusion probabilistic model is trained to produce synthetic sand that maximizes the log‐likelihood of the generated samples belonging to the original data distribution measured by a Kullback‐Leibler divergence. Numerical experiments suggest that the proposed method is capable of generating realistic grains with morphology, shapes and sizes consistent with the training data inferred from an F50 sand database. We then use a rigid contact dynamic simulator to pour the synthetic sand in a confined volume to form granular assemblies in a static equilibrium state with targeted distribution properties. To ensure third‐party validation, 50,000 synthetic sand grains and the 1542 real synchrotron microcomputed tomography (SMT) scans of the F50 sand, as well as the granular assemblies composed of synthetic sand grains are made available in an open‐source repository.
BibTeX
@article{vlassis2024synthesizing,
  author = {Nikolaos N. Vlassis and WaiChing Sun and Khalid A. Alshibli and Richard A. Regueiro},
  title = {{Synthesizing realistic sand assemblies with denoising diffusion in latent space}},
  journal = {International Journal for Numerical and Analytical Methods in Geomechanics},
  year = {2024},
  doi = {10.1002/nag.3818},
}

Loading related content...

Figure