Train on prepared point sets¶

Use Sampling → Interior data and Sampling → Boundary data when your collocation points already exist in CSV/NPZ files. They replace Interior sampler and Boundary points, respectively. The data blocks supply coordinates; PDE residual / Predefined PDE and Boundary condition still assign the physics. The existing sampler inspectors retain their older import controls for compatibility.

Start with a non-rectangular example¶

Open File → Examples → 2D prepared triangle · imported data. Click Validate, then Train PINN. Open Training Monitor or Comparison explicitly to see results; training alone opens no view and writes no file. The graph includes opt-in CSV and ParaView export blocks fed by Comparison.

Actual desktop graph. Coordinate bounds define the scaling/checking box; the triangle itself is represented by the supplied points. There is no CAD viewport or inferred mesh. Open full size.

The manufactured problem uses the triangle x≥0, y≥0, x+y≤1 and u=x²+y², so Δu−4=0. Its imported sets are:

There are 55 interior rows and 12 rows per boundary set. Boundary samples exclude vertices. The example separately evaluates 58 supplied locations with IDs, including the vertices, and uses an explicit reference expression. The fixture is independently generated demonstration data, not imported FEM results or a reproduced paper case.

Choose the domain mode¶

For non-rectangular scalar data, select the geometry block and choose Prepared 2D points. It becomes Coordinate bounds. Set x/y bounds enclosing every training/evaluation point. They control neural-network coordinate scaling and basic validation; they do not specify the actual domain boundary.

Wire the same bounds block into Interior data and Boundary data, then connect:

• Interior data → PDE residual (or a supported scalar Predefined PDE).
• Boundary data → Boundary condition.
• Physics, conditions and the chosen neural network → trainer or compact recipe.

Prepared 2D requires both imported sources and at least one named boundary set. It never fills missing sets with rectangle samples. Use a separate Evaluation points block on every requested prediction/comparison/export; a regular bounding box grid would include points outside an unknown domain and is rejected.

On built-in interval/rectangle domains, the data blocks also accept the existing named side imports. Those retain the earlier import semantics and can combine mapped sides with built-in samples for unmapped sides. The new free-label workflow is specifically the Prepared 2D scalar mode.

Map interior coordinates¶

Select Interior data and choose Map / preview interior coordinates…. CSV mappings name columns (e.g. x,y); NPZ mappings name an N×D array (e.g. points). Interior imports contain coordinates only. Normals, target values and additional pointwise PDE inputs are deliberately unavailable here.

The supplied rows are fixed. Point budgets, random seeds and sampler choices do not create or replace imported points. CSV/NPZ readers support at most 100,000 rows per source and a 64 MiB file/expanded-NPZ limit; NPZ uses allow_pickle=False.

Add named boundary sets¶

Select Boundary data, click Add named point set…, enter a name and map its file. Use Named set to select an existing mapping, preview/edit it, or remove it. These are undoable model edits. No file is deleted by removing a mapping.

Prepared names begin with an ASCII letter and contain letters, digits, _ or -, up to 64 characters; pde and total are reserved for loss records. Up to 64 sets and 100,000 boundary rows in total are supported. Names such as inlet, outer_wall or edge-17 have no implied normal direction or topology.

The existing mapping dialog previews supplied rows. This sloped edge maps x,y, nx,ny and value; its Neumann interpretation is chosen separately.

Boundary coordinates are required. Outward unit normals are optional for ordinary Dirichlet conditions, and required for Neumann or any Custom operator/target that uses a normal. Prepared 2D never substitutes a box normal, even if a set is named left. Normals must be finite, unit length and match the coordinate row count; the preparer is responsible for outward orientation.

Optional target values mean the value of the condition assigned downstream: u for Dirichlet, ∂u/∂n for Neumann, or the Custom operator's target. They are not universally solution values. In particular, Neumann is an outward derivative, not automatically conductivity-scaled flux.

Assign conditions downstream¶

Types and expressions remain in Boundary condition. A mapped target source is shown explicitly and the unused constant field is disabled.

Choose Dirichlet, Neumann or Custom for each named set. Without imported target values, use the constant or a restricted coordinate expression. With imported values, clear any target expression: using both is a validation error. Custom operators use the existing restricted expression language, including dn(u). At least one Dirichlet set is required for this scalar workflow.

Prepared validation checks finite dimensions/values, coordinate bounds, required normal mappings and incompatible target sources before starting a worker. It cannot verify an arbitrary shape from points alone, determine whether a point is truly interior, orient normals, detect every geometric gap, or construct cells. Keep coordinates and targets in consistent units before import.

Supported methods and persistence¶

The first non-rectangular workflow supports steady 2D scalar baseline, LRA, gPINN, RBA and fixed-point SA-PINN trainers, with the existing replaceable networks. Poisson, Helmholtz and reaction–diffusion presets support Prepared 2D; PDE residual also accepts restricted scalar expressions. No special network is required.

Hard box Dirichlet transforms, integral loss without quadrature weights, RAR/importance candidate generation on unknown geometry, Expert/Scale fresh sampling, arbitrary flow data and transient/physical-3D prepared geometry remain outside this milestone. The built-in geometry workflows keep their existing scope.

Model format 23 stores block-owned mappings and named conditions while reading formats 1–22. Saved runs retain evaluated rows and weights; archived results can be reopened/exported without training files. Resuming training requires the original numerical coordinates, targets and operator normals. Changing these rejects a checkpoint rather than continuing against different data.

GeoSampler and further work¶

The existing planar GeoSampler preparation helper converts supported labelled NPZ exports to explicit point/normal arrays. Map those prepared files into these blocks; the new named sets avoid pretending CAD labels are rectangle sides. Direct CAD loading, automatic manifest-to-graph wiring, mesh connectivity, unit conversion and region-field metadata remain future work. For output mappings and preserved IDs, see Evaluate supplied points.

Checks and recorded results¶

From the repository root, using its existing environment:

python -m examples.training_data --device cpu --output-dir /tmp/margam-training-data-cpu
python -m examples.training_data --device cuda --output-dir /tmp/margam-training-data-cuda
python -m examples.capture_training_data --overwrite

CPU and CUDA records cover 100 initial updates plus two resumed updates, exact same-device continuation, source hashes, row counts, loss terms and supplied-point CSV/VTU export. Errors are recorded without an accuracy threshold or tuning claim. Numerical tests also substitute the manufactured solution directly into the PDE and mapped boundary operators. The capture manifest identifies the actual UI screenshots. ParaView GUI playback was not part of this steady-data check; the file contains vertices, not triangular cells.

Keep data with the project¶

Choose Copy into project or Link existing file in the mapping dialog. Copies travel with the model's companion folder, recovery and run archives; links remain dependent on their original location. New sources default to Copy, while existing linked sources keep Link. No rows, IDs or field mappings are changed. See the illustrated copy/link guide.

Use View → Project data overview… to inspect row counts and storage for all imported sets, including disconnected blocks. It flags missing/invalid files and changes since its first session check. See the overview guide.