Physics-informed convolutional neural networks for fluid flow through porous media

In three linesCNN encoder-decoder framework predicts pore-scale velocity fields in porous media directly from geometry. Custom loss function enforces velocity reconstruction, incompressibility, no-flow conditions, and physical constraints. Tested on out-of-distribution geometries and accelerates Lattice-Boltzmann simulations (90% of cases).

Summary generated by Claude — human-verified