🤖 AI

LM6-CastBench: A Geometry-Aware Simulation Dataset for Transient Thermal Prediction in LM6 Sand Casting

This paper introduces LM6-CastBench, an open dataset comprising 450 geometry-process cases for LM6 aluminum sand casting that enables reproducible, geometry-aware benchmarking of machine learning models for predicting transient thermal fields, demonstrating that spatially structured U-Nets significantly outperform point-wise MLPs in capturing complex thermal behaviors.

Original authors: Suhani Kolhatkar, Anand Khandekar

Published 2026-11-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Suhani Kolhatkar, Anand Khandekar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When molten metal is poured into a mold to create a part, it does not cool evenly. The way heat moves through the liquid as it turns solid determines whether the final object is strong or riddled with hidden flaws like tiny holes or cracks. Engineers have long relied on powerful computer simulations to predict these temperature changes, but running these simulations for every new design is slow and expensive. It is like trying to solve a complex puzzle from scratch every time you change a single piece, rather than finding a pattern that applies to the whole picture. To speed up this process, researchers have begun using artificial intelligence to learn from past simulations and predict future outcomes instantly. However, a major hurdle has remained: most of these learning systems were trained on data that stripped away the shape of the object, treating the metal as a simple list of numbers rather than a physical form with curves, holes, and thicknesses. Without the shape, the computer struggles to understand how heat behaves in the real world.

A new study introduces a solution to this problem by creating a massive, open library of simulated cooling data specifically designed to keep the geometry intact. The researchers focused on LM6, a common aluminum alloy used in everything from car parts to machinery, and built a dataset called LM6-CastBench. Instead of flattening the data, they preserved the three-dimensional structure of the metal parts, including the exact arrangement of the computer mesh that divides the object into tiny pieces for calculation. They generated 450 distinct scenarios by combining three different shapes—a brake lever, a flange, and a stepped block—with ten variations of each shape, and then simulated each one under fifteen different temperature conditions. This approach ensures that the artificial intelligence sees the object exactly as it exists, with all its physical details, rather than just a abstract summary of numbers.

The team tested two different types of artificial intelligence models against this new dataset to see which could better predict the temperature fields. The first model was a standard point-by-point predictor, which looked at the coordinates of a single point in the metal and tried to guess its temperature based on the surrounding conditions. The second model was a spatially aware system, designed to look at the entire shape at once, much like how a human eye perceives a whole object rather than just isolated dots. The results were clear: the model that understood the shape of the object performed significantly better. While the point-by-point model managed to make rough guesses, it struggled immensely with parts that had sharp corners, holes, or sudden changes in thickness. In these complex areas, its predictions were often off by dozens of degrees, a difference that could mean the difference between a safe part and a failed one.

In contrast, the spatially aware model, which processed the entire 3D structure together, reduced these errors dramatically. For the most complex shapes, such as the flange with its ring structure and bolt holes, the new model cut the prediction error from over thirty-six degrees down to just one degree. It was able to track how heat moved across the surface and through the material, correctly identifying how the shape influenced the cooling process. The study showed that this advantage was not just a small improvement but a fundamental shift in accuracy, particularly for parts with localized features where heat gets trapped or escapes quickly. The researchers noted that while the spatial model took slightly longer to run a prediction, the gain in precision was substantial enough to make it a far more reliable tool for engineers.

This work does not claim to have solved every problem in casting or to have replaced the need for physical testing. The simulations were limited to specific shapes and temperature ranges, and the models were trained on data generated by a specific type of computer solver. However, the study successfully demonstrated that keeping the physical shape of the object visible to the artificial intelligence is crucial for accurate results. By providing a reproducible dataset where geometry and process conditions are linked, the researchers have given the scientific community a new standard for testing future tools. The findings suggest that the next generation of design software will need to see the world in three dimensions to truly understand how metal cools, moving beyond simple lists of numbers to embrace the complex reality of the physical forms they are meant to predict.

Technical Summary: LM6-CastBench

Problem Statement
Transient thermal fields during the solidification of aluminum castings are critical for predicting defect-prone regions, such as porosity and shrinkage. While finite-element (FE) simulations can resolve these fields, repeatedly solving the thermal problem across varying geometries and process conditions is computationally expensive, motivating the use of data-driven surrogate models. However, existing machine learning studies on casting thermal behavior rely on study-specific simulation databases that lack reproducible, geometry-level benchmarking. These prior works often utilize procedurally generated 2D geometries or fixed casting configurations where only process parameters vary, failing to provide a benchmark where 3D geometry and process variations are systematically controlled while retaining the underlying mesh and transient field data.

Methodology
The authors introduce LM6-CastBench, an open simulation dataset designed for geometry-aware learning of transient thermal fields in LM6 aluminum sand casting. The dataset is generated via an end-to-end pipeline:

Geometry Generation: Three geometry families (brake lever, flange, and stepped block) are parametrized. Ten distinct variants per family are generated using Latin Hypercube Sampling (LHS) within defined design spaces, exported as STEP models, and independently meshed with first-order tetrahedral elements. Local refinement is applied near holes, fillets, and curved features.
Simulation: Each of the 30 geometry variants is simulated under a grid of process conditions: five pouring temperatures (953–1033 K) and three mold temperatures (303–403 K), resulting in 450 total geometry-process cases. Simulations are performed using the Elmer FEM solver with an enthalpy-based phase-change formulation under a conduction-dominated assumption.
Data Organization: The dataset retains node coordinates, mesh connectivity, signed-distance values (SDF), process metadata, and time-resolved nodal temperatures. Crucially, the data is split at the geometry-variant level (80/10/10 for train/val/test) within each family. This ensures that timesteps from the same physical geometry never appear in both training and test sets, preventing data leakage and enabling true geometry-level generalization evaluation.
Benchmarking: Two surrogate models are compared under these geometry-level splits:
- A Point-wise MLP: Predicts nodal temperature using coordinates, SDF, process temperatures, and time as input.
- A Grid-based 3D U-Net: Predicts the temperature field on a $32 \times 32 \times 32$ voxel grid using occupancy, SDF, process, and temporal channels, then maps predictions back to FEM nodes.

Key Contributions

Controlled 3D Simulation Dataset: A dataset of 450 transient thermal cases spanning three geometry families, ten variants each, and fifteen process combinations, linking geometry, mesh, process conditions, and thermal response in a common representation.
Reproducible Geometry-Level Evaluation: A data organization strategy that separates geometries during evaluation, allowing researchers to test whether models can generalize to unseen shapes rather than just unseen time steps of the same shape.
Benchmark Results: A demonstration that retaining spatial structure (via the U-Net) substantially improves thermal-field reconstruction compared to point-wise approaches, particularly for geometries with localized features and abrupt section changes.

Results
Quantitative benchmarks on held-out geometry variants show that the U-Net consistently outperforms the MLP:

Error Reduction: The U-Net significantly reduces Mean Absolute Error (MAE). For the brake geometry, MAE drops from 8.38 K (MLP) to 2.19 K (U-Net). For the flange, it drops from 36.75 K to 1.03 K, and for the stepped block, from 80.71 K to 1.86 K.
Spatial Behavior: The MLP exhibits larger, persistent errors around geometric discontinuities (e.g., bolt holes, flange rings, and cross-sectional transitions) due to its lack of neighboring context. The U-Net tracks the simulated field more closely in these regions by processing spatial locations jointly.
Temporal Dynamics: While the MLP is faster at inference (0.51–1.20 s vs. 2.66–8.04 s for U-Net), the U-Net converges to low residual errors more rapidly, especially in the later stages of cooling where the MLP often plateaus at a persistent bias.

Significance and Claims
The paper claims that LM6-CastBench provides a reproducible dataset and benchmark for geometry-aware surrogate modeling of casting thermal behavior. The authors assert that the retained spatial structure in the dataset has a measurable, geometry-dependent effect on prediction accuracy. The results indicate that spatially structured prediction (U-Net) is superior to point-wise prediction (MLP), with the largest gains observed in geometries containing localized features and abrupt section changes. The dataset is intended to serve as a baseline for future comparisons involving graph networks, neural operators, and physics-informed models, offering a controlled space to evaluate these methods against the same underlying physics and geometry splits.

Limitations
The authors note that the current dataset covers only ten variants per family and two process parameters. The U-Net's accuracy is bounded by the $32^3$ voxel quantization, where nodes sharing a voxel receive identical predictions. The benchmark currently uses only two baseline architectures, and the test set (three held-out geometries) is treated as illustrative of the dataset's utility rather than a statistically exhaustive evaluation. Future extensions could include broader geometric coverage, parametrized gating/riser systems, and coupled filling-solidification physics.

Technical Summary: LM6-CastBench

More like this