LM6-CastBench tests how spatial context improves casting temperature predictions

A benchmark for cooling, not defect detection

Cooling during solidification matters because thermal behaviour is connected to casting defects such as porosity and shrinkage-related discontinuities. Repeated numerical simulations across different shapes and process settings, however, are computationally demanding.

In their preprint, Suhani Kolhatkar and Anand Khandekar introduce LM6-CastBench as a resource for developing surrogate models: machine-learning systems trained to approximate simulation outputs. Its target is the temperature field throughout an LM6 aluminium casting over time. It does not simulate mould filling or directly predict defects.

The authors position the dataset as an alternative to study-specific databases that make geometry-level comparisons difficult. Their central contribution is to preserve the relationship between a casting’s shape, its numerical mesh, its process settings and its simulated cooling history.

What the 450 cases contain

The dataset includes three shape families: brake levers, flanges and stepped blocks. Each family has ten controlled geometric variants. Every variant is simulated under fifteen process combinations, formed from five pouring temperatures between 953 and 1033 kelvin and three mould temperatures between 303 and 403 kelvin. Together, these produce 450 geometry–process cases, not 450 distinct shapes.

Each shape is independently divided into a mesh of tetrahedral elements. The Elmer finite-element solver then calculates its transient thermal response using a conduction-dominated model. An enthalpy-based formulation accounts for phase change during solidification; metal flow and convective transport within the molten alloy are excluded.

The released cases retain CAD geometry, mesh connectivity, node coordinates, signed-distance descriptors, process parameters and time-resolved temperatures at mesh nodes. Metadata connects these assets to case identifiers and split assignments. The authors report releasing the dataset and code under CC BY 4.0, alongside benchmark scripts and a fixed split manifest.

Testing unfamiliar geometry variants

The benchmark separates data by geometry variant rather than randomly mixing individual time samples. Within each family, eight variants are used for training, one for validation and one for testing. Consequently, different times from the same physical shape cannot appear in both training and test sets.

This distinction matters: predicting another moment in a familiar shape is not the same task as predicting cooling in an unseen variant. Nevertheless, the test set contains only three held-out geometries overall. The authors therefore describe the comparison as illustrative rather than statistically exhaustive.

Two models are evaluated. A multilayer perceptron, or MLP, predicts each node’s temperature from its coordinates, signed-distance value, process temperatures and time, without directly processing neighbouring nodes together. A three-dimensional U-Net instead predicts a temperature field on a 32 × 32 × 32 voxel grid using spatial and process inputs. Its outputs are mapped back to mesh nodes for evaluation. Separate U-Nets are trained for the three geometry families.

Lower errors, slower inference

The authors report lower mean absolute temperature errors for the U-Net across all three families. For the brake geometry, error falls from 8.382 to 2.187 kelvin. For the flange, it falls from 36.748 to 1.033 kelvin; for the stepped block, from 80.707 to 1.861 kelvin. These are aggregate errors across test nodes, times and process cases, not guarantees for every individual prediction.

The authors associate the largest improvements with localized features and abrupt changes in cross-section. Their error maps show persistent MLP errors around flange rings, bolt holes and stepped transitions. They interpret the U-Net’s advantage as evidence that neighbouring spatial context helps reconstruct these fields.

Accuracy comes with a timing tradeoff. Reported MLP inference times range from about 0.51 to 1.20 seconds, compared with 2.66 to 8.04 seconds for the U-Net, including voxelisation overhead. These measurements compare the two learned models; they do not establish a speedup over the finite-element solver.

What remains limited

Coverage is restricted to three shape families, ten variants per family and two process parameters. Training uses a single seed per model rather than repeated-seed averages. The U-Net also has a resolution limit: nodes assigned to the same voxel receive identical predictions.

The source additionally lists three mould-temperature settings while specifying a fixed external boundary temperature of 353 kelvin; the presented equations do not clarify how the varied mould setting enters that boundary treatment. These constraints support reading LM6-CastBench as a controlled simulation benchmark, not evidence of demonstrated prediction accuracy in physical foundry trials.