Pick the template, input parameters, output fields, and which runs train or test.
Training and models
A solved case becomes a grid, then weights, then a fast prediction. Training reads the mesh and the time-directory fields. It does not read the viewer files.
Pipeline
1SolveMass run
2ArtifactNative zip
3StageCartesian grid
4Trainweights/
5Predictsurrogate engine
Job artifact
Each job ships a native case zip. One field file looks like this.
results/openfoam_native.zip ├── constant/polyMesh/ │ ├── points │ ├── faces │ ├── owner │ └── neighbour ├── 0.1/ │ ├── p (or p_rgh) │ ├── U │ └── alpha.water ├── 0.2/ │ └── … └── …
/* OpenFOAM time-dir field (p) — cell values */ dimensions [0 2 -2 0 0 0 0]; internalField nonuniform List<scalar> 4 ( 1.02e5 1.01e5 9.98e4 9.95e4 );
Stage resamples onto a grid
Cell centres become a regular lattice. The model needs that grid.
Unstructured cells
Cartesian lattice
A staged sample
One sample is one run at one time: inputs plus field planes on the grid.
Logical sample
{
"inputs": {
"damHeight": 0.35,
"Uinlet": 0.12
},
"simulation_time": 0.2,
"run_index": 7,
"grid": { "nx": 48, "ny": 32, "nz": 1 },
"fields": {
"p": [[1.02e5, 1.01e5, /* … nx */], /* … ny rows */],
"U": [[[0.1,0,0], [0.12,0,0], /* … */], /* … */],
"alpha.water": [[1, 1, 0, 0, /* … */], /* … */]
},
"bbox": { "min": [0, 0, 0], "max": [0.6, 0.4, 0.1] }
}On disk
{
"inputs": { "damHeight": 0.35, "Uinlet": 0.12 },
"simulation_time": 0.2,
"run_index": 7,
"grid": { "nx": 48, "ny": 32, "nz": 1 },
"storage": "npy",
"fields_file": "run_7_t0.2.npy",
"channels": ["p", "U_x", "U_y", "U_z", "alpha.water"],
"coords_file": "coords.npy"
}Same time, different input
Two input values at the same time produce different field maps. The picture is illustrative.
{ damHeight: 0.35 }{ damHeight: 0.42 }{
"engine": "surrogate",
"model_id": "model_dam_fno_1",
"inputs": { "damHeight": 0.42, "Uinlet": 0.10 },
"simulation_time": 0.2
}Train tensors
Samples pack into inputs and field channels, stored under the model id.
train.npz
x: (N, n_inputs)
y: (N, C, H, W)
models/{id}/staging/bundle.zip
models/{id}/weights/bundle.zip{
"x_keys": ["damHeight", "Uinlet", "simulation_time"],
"y_channels": ["p", "U_x", "U_y", "U_z", "alpha.water"],
"grid": { "ny": 32, "nx": 48 },
"n_train": 80,
"n_test": 20
}Checklist
Sweep the inputs. Keep the native case zip, not a viewer-only export.
A worker with the staging capability writes models/{id}/staging/.
A worker with the training capability writes models/{id}/weights/.
An optional hold-out checks the error.
A surrogate run uses the trained model. The viewer still reads fields.
Related screens in this console: Models and Architecture. Mass runs and predictions stay in Forge.