Poreformer

Poreformer is a term used in the field of computational materials science and machine learning to describe transformer-based models applied to porous media. The aim is to learn representations of pore structures and use them to predict transport properties such as permeability, diffusivity, and capillary pressure, often with higher efficiency than traditional pore-scale simulations.

Porous media, including soils, rocks, catalysts, and ceramics, exhibit complex geometries across multiple scales. Conventional approaches rely on pore network models, lattice Boltzmann simulations, or direct numerical simulation on CT images, which can be computationally intensive. Poreformer seeks to learn from labeled data generated by these methods or experiments to generalize to new materials and conditions.

Architecturally, poreformer models typically combine representations of pore geometry, such as graph structures of pores and throats or voxel-based images, with transformer blocks that model long-range interactions. They may include auxiliary inputs for physical conditions like pressure, temperature, and fluid properties. The output can be predictions of effective properties, or surrogate simulations of flow and diffusion under specified boundary conditions.

Training data requirements include diverse pore structures, high-fidelity ground truth from experiments or simulations, and careful handling of scale and heterogeneity. Evaluation often uses metrics for permeability prediction accuracy, error in diffusion coefficients, and generalization across porosity ranges.

Applications span hydrogeology, petroleum engineering, catalyst design, and materials science, where rapid screening of porous materials is valuable. Poreformer can enable faster design cycles and integrated simulations, though challenges remain in interpretability, data availability, and extrapolation beyond the training domain.

See also: porous media, transformer (machine learning), pore-scale modeling.