Casting solidification GNN (v5.1)

A surrogate model for the solidification of metal castings in a sand mold. It takes a tetrahedral mesh of a casting together with the alloy and mold properties, and predicts the time at which every mesh node solidifies. The nodes that solidify last are the hot spots, where feeding problems and shrinkage defects usually start.

A transient finite-element solve of the same field takes minutes to hours. The model returns it in seconds, which makes it useful for screening design variants before a full simulation. It does not replace a calibrated process simulation or physical validation, and a hot spot is a shrinkage-risk indicator, not a porosity label.

To look at predictions without installing anything, open the demo Space: it shows four parts in two alloys in 3D, with the hot spots and a shrinkage screening.

Contents

• Input and output
• How to use
• Model architecture
• Files
• Training data
• Training procedure
• Evaluation
• Intended use and limitations
• Reproducibility
• License and citation

Input and output

The model reads one casting at a time.

The mesh is a tetrahedral volume mesh in millimetres, in gmsh 2.2 ASCII format (.msh). Surface triangles that cool into the mold belong to a physical group named Cooling; in this release that is the whole outer surface. The model was trained on meshes of roughly 7,500 to 10,000 nodes, and that is the density it should be given. It does not use node numbering or absolute coordinates: node inputs are distances, flags and volume-type quantities in millimetres, and edges carry only relative displacements between neighbouring nodes.

Eight numbers describe the alloy and the mold: thermal conductivity, density, specific heat, latent heat, liquidus, solidus, pouring temperature and mold effusivity. The code ships two tabulated alloys, EN-GJS-400-15 (ductile iron, the default) and G20Mn5 (cast steel), and one mold, green sand. You can also pass your own eight values.

The output is one number per node: the time in seconds at which that node cools through the solidus. The maximum over all nodes is the total solidification time, and the 10% of nodes with the largest times are the hot spots.

How to use

This repository holds the weights and this card. The code that loads and runs them is on GitHub:

git clone https://github.com/eugenmik/castsolid-gnn
cd castsolid-gnn
python -m venv .venv && source .venv/bin/activate
pip install "torch==2.3.0" --index-url https://download.pytorch.org/whl/cpu
pip install torch-cluster -f https://data.pyg.org/whl/torch-2.3.0+cpu.html
pip install -r requirements.txt

Python 3.11 is recommended. For a GPU, replace cpu with your CUDA tag (for example cu121) in both URLs. Then, from the repository root:

from src.model_api import CastSolidModel

model = CastSolidModel.from_pretrained()         # downloads this repository's files once
t = model.predict("part.msh")                    # EN-GJS-400-15 in green sand
t_steel = model.predict("part.msh", alloy="G20Mn5")

t.max()                                          # total solidification time, s
hot = CastSolidModel.hotspots(t)                 # boolean mask, top 10% of nodes

from_pretrained() fetches config.json and the files it lists into the local Hugging Face cache, pinned to the v5.1 tag; later calls run offline. predict returns a NumPy array in the node order of the .msh file. To use another alloy, pass props_json="my_alloy.json" with exactly these keys (SI units, temperatures in °C; 1018 is the green-sand effusivity):

{"k": 32.9, "rho": 6800.0, "cp": 515.0, "L_kJ_kg": 220.0,
 "TL": 1160.0, "TS": 1120.0, "Tpour": 1260.0, "mold_effusivity": 1018.0}

To start from a STEP file instead of a mesh, run python -m src.predict_casting part.step in the GitHub repository. It heals the CAD, builds a mesh at the right density, runs the model and writes the result for ParaView.

Model architecture

The model is a hybrid of three trained parts. The GNN predicts the spatial shape of the field; two gradient-boosted heads predict its overall level and its spread; a fixed recomposition step combines them. The GNN alone is less reliable about absolute magnitude, which depends strongly on alloy and part size. On the original benchmark of held-out geometry families, adding the heads brought nodal MAE down from 5.77 s to 4.26 s.

flowchart LR
  M[Tetrahedral mesh] --> F[15 node features<br/>+ union graph]
  P[Alloy + mold<br/>properties] --> G[5-value conditioning vector]
  P --> D
  F --> N[12-block GNN<br/>field shape]
  G --> N
  F --> D[Part descriptors]
  D --> T[GBM total head]
  D --> S[GBM spread head]
  N --> R[Soft-anchor<br/>recomposition]
  T --> R
  S --> R
  R --> O[Time per node, s]

Graph network

Nodes are mesh vertices. Edges are the union of the tetrahedral mesh edges and the 16 nearest neighbours of each node. Each edge carries its displacement vector, its length and a flag that says whether it is a mesh edge.

Each node gets 15 features computed from the mesh and its surface tags:

• geodesic distance through the mesh to the nearest cooling surface and to the nearest adiabatic surface, plus straight-line distance to the surface and to adiabatic faces;
• cooling, adiabatic and interior flags;
• part volume, surface area and the fraction of surface that cools;
• local modulus, the volume-to-surface ratio inside geodesic balls of 5, 15 and 40 mm, which is a local version of Chvorinov's modulus;
• two boundary-condition channels: the mold effusivity on surface nodes (zero inside the part), and a boundary heat-input channel, which is zero because a sand-mold surface adds no heat.

The alloy and mold enter as five dimensionless numbers: Stefan number, normalized freezing range, thermal diffusivity ratio, normalized superheat and normalized mold effusivity. They condition every processor block through FiLM (h ← h·(1+γ(g)) + β(g)).

The network follows the encode-process-decode design of MeshGraphNets. MLP encoders embed nodes and edges, 12 message-passing blocks (hidden width 128, residual connections, LayerNorm, mean aggregation) process them, and a two-head decoder predicts the per-node deviation and a graph-level scale separately. The training target is the standardized log1p of the solidification time.

Gradient-boosted heads

Both heads are scikit-learn HistGradientBoostingRegressor models with one row per part: the alloy properties, the part's volume-to-area ratio, the mold effusivity, and boundary-condition descriptors that are fixed to the plain sand mold in this release. The total head (16 inputs) predicts log1p of the total solidification time. The spread head (17 inputs, absolute-error loss) predicts log1p of the standard deviation of the nodal field.

Soft-anchor recomposition

The GNN field is scaled, shifted and clipped at zero,

pred(a, m) = max(0, a · (gnn − mean(gnn)) + m)

with (a, m) chosen to minimize

J = w_total · ((q99.5(pred) − T_total) / T_total)²  +  w_std · ((std(pred) − S) / S)²

where T_total and S come from the two heads, w_total = 1 and w_std = 4. When both targets can be met exactly, this equals the closed-form affine solve. When clipping makes that impossible, the total anchor may move by a few percent so the spread stays close. The post-clip spread is not monotone in a, so the solver brackets the smallest root explicitly.

Files

None of the weight files can execute code when loaded. The GNN is stored as safetensors and the heads as skops. Before it opens a skops file, the loader on GitHub checks every type the file names against an allowlist of scikit-learn and NumPy types. Training code is not published.

Training data

Every reference field was computed with FEniCSx (dolfinx) using an enthalpy-method model of transient heat conduction with latent heat. There is no mold filling, flow, gas or stress in it.

Training meshes have roughly 7,500 to 10,000 nodes. The target at each node is the time its temperature drops through the solidus. Material properties come from handbook data. The raw simulation corpus (about 100 GB) is not distributed.

One data issue shaped this project more than any model choice. The solver stored temperatures in dolfinx degree-of-freedom order, while coordinates and geometric features were in gmsh node order, so about 60% of the nodal targets were silently attached to the wrong positions. Every early model failure traced back to it. The training set was rebuilt with a recovered permutation between the two orderings. If you train on solver output, check node identity across every solver and mesh interface before tuning a model.

Training procedure

v5.1 came out of a chain of fine-tunes, not a single training run.

1. SP1 trained a 12-block two-head GNN from scratch on the primitive set, with the union graph, local-modulus features and a loss on both nodal and total time.
2. FT2b fine-tuned it on the multi-alloy set. The FiLM layers started at zero, so training began from exactly the SP1 function.
3. A follow-up pass re-encoded one sparse boundary input as a field diffused along the mesh, because the network had learned to suppress it, and picked the checkpoint on gate metrics instead of validation loss alone.
4. FT3 fine-tuned on real production geometries (set E) with the normalization statistics frozen. Its epoch-21 checkpoint is the v5.1 GNN.
5. The total head (v3), the absolute-error spread head (v2) and the recomposition weights were selected on held-out real-part validation rows.

Training ran on one RTX 3060 (12 GB) with batch size 2, at 15 to 20 minutes per epoch. To be promoted, v5.1 had to pass four gates on held-out data: two gold-standard real parts, 24 held-out real parts, an extrapolation split and a check against collapsed field variance.

Evaluation

Do not compare numbers across regimes. Total times in the primitive regime are tens of seconds; on real geometry they run to hundreds or thousands.

Hot-spot overlap compares the 10% of nodes that solidify last in the prediction with the same set in the reference. 1.0 means the two sets are identical.

One of the six real parts reached a hot-spot overlap of only 0.035, so the median says nothing certain about any one part.

Important. The training data and the real-geometry references come from the same FEniCSx heat-transfer model, the same material database and the same boundary assumptions. These numbers measure how well the surrogate carries over to new geometry. They are not evidence of physical accuracy against instrumented castings.

Intended use and limitations

Good uses:

• comparing early casting concepts;
• finding candidate last-to-solidify regions;
• deciding which designs need a full process simulation;
• early quoting and design discussions, as long as the limitations stay visible.

Out of scope:

• final process sign-off or contractual acceptance;
• guaranteed defect or porosity prediction;
• mold filling, flow, gas, stress or distortion;
• unattended decisions on geometry outside the training envelope;
• safety-critical decisions without an independent engineering review.

Known limitations:

1. The physics is pure conduction with latent heat, and the reference solver shares the model's property and boundary assumptions.
2. This release covers one setup: a single casting whose whole outer surface cools into a green-sand mold, in EN-GJS-400-15, G20Mn5 or a custom property set.
3. Meshes much coarser or finer than the training density (about 7,500 to 10,000 nodes) are outside what the model has seen. At that density, large thin-walled parts can be under-resolved.
4. The training envelope in ood_ranges.json only covers part size and distance to cooling. It is not a calibrated uncertainty estimate.

Treat each prediction as a screening hypothesis: where should an engineer look first, and is the concept ready for a full simulation? Keep the mesh, the inputs and the model version with every result, and send high-consequence cases to a calibrated simulation and a physical review.

Reproducibility

The published files were converted from the internal training checkpoints and checked against them:

The published inference code was also run side by side with the internal pipeline on five production parts (two of them with both alloys), plus a fixed-size mesh with a custom property set: every output field matched exactly. The heads were trained with scikit-learn 1.5.2, and other versions may refuse to load them or change their output. Other versions used: Python 3.11, PyTorch 2.3.0, PyTorch Geometric 2.7.0, NumPy 2.x.

License and citation

Code and weights are released under the Apache License 2.0. The tables in data/properties/ of the GitHub repository are modelling inputs, not certified material specifications.

@software{miknevic_casting_gnn_2026,
  author  = {Miknevic, Eugen},
  title   = {Casting Solidification GNN: a graph neural network surrogate for
             rapid solidification screening of metal castings},
  version = {5.1},
  year    = {2026},
  url     = {https://huggingface.co/eugenmik/castsolid-gnn}
}

Contact: Eugen Miknevic, eugenmiknevic@gmail.com. If you have failure cases, questions or measured data from real castings, please post them in the Community tab.