PorousDiff

Generative Surrogate Modelling for Mechanical Deformation in Aqueous Al-Ion Battery Electrodes.

PorousDiff is a conditional 3D diffusion model designed to act as a Virtual Compression Rig for anisotropic porous media.

Developed by the BASE Laboratory for the Isambard-AI supercomputer.

---

🔬 The Problem

Optimising electrode density requires understanding how fibrous microstructures buckle and close pores under compression.

• Synchrotron Experiments: Accurate but prohibitively expensive and slow.
• DEM/FEM Simulations: Computationally intractable for full-scale Representative Elementary Volumes (REVs).
• Standard GANs: Fail to capture long-range anisotropic connectivity (through-plane tortuosity).

PorousDiff solves this by learning a continuous mapping between a scalar compression ratio $C$ and the resulting 3D microstructure.

🛠️ Technical Stack

This project will use a high-performance stack designed for the NVIDIA Grace-Hopper Superchip (GH200).

• Core Architecture: 3D Denoising Diffusion Probabilistic Model (DDPM)
• Conditioning: Feature-wise Linear Modulation (FiLM) for scalar physics injection.
• Framework: MONAI Generative Models (PyTorch).
• Configuration: Hydra (structured .yaml configs).
• Experiment Tracking: Weights & Biases.
• Validation: OpenImpala (Tortuosity/Transport).

🚀 Getting Started (For Developers)

1. Clone the Repository

git clone [https://github.com/BASE-Laboratory/PorousDiff.git](https://github.com/BASE-Laboratory/PorousDiff.git)
cd PorousDiff