AERO

Solve an urban wind case in OpenFOAM 14 twice over: once the conventional way, and once with the fine-mesh initial field predicted by AERO's neural multi-fidelity model. The governing equations, mesh discretisation, k-ε parameters and residual tolerances are identical in both runs — only the starting point differs — so the iteration count and wall-clock are directly comparable.

Recent runs

Case	Cells	Baseline solver		AI 1		Saved by AI 1		Status
Case	Cells	iters	time	iters	time	iterations	wall-clock	Status

Choose a case

Drop a case archive here

.zip, .tar.gz, .tgz, .tar.bz2, .tar.xz · up to … MB

What the archive must contain

case/
├─ params.json          optional — inflow parameters are fitted from the inlet profiles if absent
├─ fine/                the target mesh and boundary conditions
│  ├─ 0/                U, p, k, epsilon, nut
│  ├─ constant/polyMesh
│  └─ system/           controlDict, fvSchemes, fvSolution
└─ coarse/              the 2x-coarser twin — required by AI 1
   └─ 0/ constant/ system/

The mesh must be a uniform 0.05 m voxel grid — a blockMesh with the buildings removed by topoSet/subsetMesh, which is what the AERO corpus generator produces. Graded and body-fitted meshes (snappyHexMesh and friends) are rejected: the network's input tensor, the slice figures and the 3D view all read the case as voxels. A bare OpenFOAM case (system/ + constant/ + 0/) is accepted and can be run with the baseline solver; AI 1 additionally needs the coarse twin to predict from. Cases carrying #codeStream, #calc or coded* entries are rejected, because those make OpenFOAM compile and run code from the archive.

or pick a sample district

Choose the solver

Baseline solver (cold start)

Stock OpenFOAM foamRun — incompressible SIMPLE with k-ε — started from a uniform initial field, exactly as it ships.

Converged solution, used as the reference
Iteration count and wall-clock

AI 1 (Warm start model) 3D U-Net

Hierarchical multi-fidelity initialisation: solve the coarse twin, let the network map it — with the geometry, wall distances and inflow parameters — to the production-mesh fields, seed 0/U and 0/p, then run the same solver.

Everything the baseline reports, plus
Prediction vs converged truth, against a coarse-upsample baseline

Both (full comparison) recommended

Mesh the case once, then solve it twice from the identical mesh — cold, then with AI 1. One run, every plot populated on both sides.

Iterations and wall-clock side by side, with the speed-up factors
Convergence histories of both arms on one axis

Access token