Abstract
MeltpoolNet is a benchmarking framework and open-source dataset for predicting melt pool behavior in metal additive manufacturing (MAM). By aggregating data from over 80 research papers and employing physics-aware feature engineering, the framework enables high-accuracy prediction of melt pool geometry and defect classification, significantly outperforming traditional analytical models like the Rosenthal equation.
Introduction
In Metal Additive Manufacturing (MAM), a high-energy beam melts metal powder layer by layer. The size, shape, and stability of the resulting melt pool are considered the "DNA" of the final part. Deviations in the melt pool lead to critical defects: if the pool is too deep, "keyhole" pores form; if too shallow, layers fail to adhere (lack of fusion); and if too elongated, the metal beads up (balling).
Engineers traditionally rely on the Rosenthal equation to estimate these characteristics. However, this simplified physics model ignores complex fluid dynamics and material-specific laser absorption, often leading to poor real-world predictions. While machine learning (ML) offers a faster alternative, most existing studies are limited to small, single-machine datasets. MeltpoolNet addresses this by providing a universal, large-scale benchmark for melt pool prediction.
Methods
The MeltpoolNet framework treats melt pool characterization as two distinct tasks: Regression for physical dimensions (depth, width, and length) and Classification for identifying the melt pool mode (Desirable, Keyhole, Lack of Fusion, or Balling).
A key innovation is physics-aware featurization. Instead of using only raw parameters like laser power and speed, the authors calculated "Absorptivity Coefficients"—mathematical proxies for energy absorption based on melt pool shape. The framework benchmarks several architectures, including ensemble methods (Random Forest, XGBoost), Gaussian Process Regression (GPR), and deep learning (Neural Networks).
Datasets
The authors compiled a comprehensive dataset by extracting and cleaning data from over 80 MAM research articles. This "data salvage" resulted in approximately 2,200 experimental data points covering 29 different metal alloys and four major MAM processes: Laser Powder Bed Fusion (L-PBF), Electron Beam Powder Bed Fusion (E-PBF), Laser Engineered Net Shaping (LENS), and Electron Beam Additive Manufacturing (EBAM).
Key Insights and Results
Benchmarking results demonstrate that ML models significantly outperform traditional analytical methods:
- Geometry Prediction: Neural Networks and Gradient Boosting models achieved $R^2$ values between 0.95 and 0.99 for predicting melt pool dimensions. In contrast, the traditional Rosenthal estimation for melt pool length yielded an $R^2$ of -0.15, failing to capture the trend.
- Defect Classification: The Random Forest model achieved 88.4% accuracy (AUC-ROC of 0.98) in identifying defect-free "desirable" prints versus failure modes like balling.
- Feature Engineering: Incorporating hatch spacing and absorptivity coefficients improved model accuracy more significantly than increasing model complexity.
- Interpretability: The authors used constrained optimization to "discover" new physical equations that mimic the Rosenthal form but fit experimental data 20% better, offering a balance between ML accuracy and physical intuition.
Limitations and Future Work
Despite its success, the framework faces challenges with data scarcity for certain features like hatch spacing, which were infrequently reported in source literature. Furthermore, current models predict "steady-state" melt pools, whereas real-world pools fluctuate dynamically. Future work aims to integrate these models into real-time closed-loop control systems to adjust processing parameters on the fly, ensuring part quality throughout the build process.
Summary updated