Source-linked AI summary
MeltpoolNet: Melt pool Characteristic Prediction in Metal Additive Manufacturing Using Machine Learning
Parand Akbari, Francis Ogoke, Ning-Yu Kao, Kazem Meidani, Chun-Yu Yeh, William Lee, Amir Barati Farimani
TL;DR
MAM meltpool geometry and flaws are difficult to predict because the process is complex and defects affect printed-part quality. The paper introduces MeltpoolNet, a benchmark combining a broad experimental dataset, physics-aware featurization, machine-learning models, and interpretable model identification. The benchmark supports geometry and defect prediction, with feature engineering and selected models producing strong reported performance while the training sets require complete feature combinations.
Problem
MAM meltpool geometry and flaws are difficult to predict because process behavior is complex, while defects affect printed-part quality.
Method
MeltpoolNet benchmarks machine-learning models using experimental meltpool data, physics-aware featurization, and explicit data-driven model identification.
Results
Neural networks, Gradient boosting, and random forest outperform other benchmarked models for regression and classification, while feature engineering improves accuracy and generalizability.
Takeaways & Limitations
The benchmark provides a basis for meltpool control and process-parameter optimization, including prediction of printed-part porosity.
Takeaways & Limitations
Training and prediction use only samples containing all features required by a given feature combination.
Abstract
from arXiv · showhide
Characterizing meltpool shape and geometry is essential in metal Additive Manufacturing (MAM) to control the printing process and avoid defects. Predicting meltpool flaws based on process parameters and powder material is difficult due to the complex nature of MAM process. Machine learning (ML) techniques can be useful in connecting process parameters to the type of flaws in the meltpool. In this work, we introduced a comprehensive framework for benchmarking ML for melt pool characterization. An extensive experimental dataset has been collected from more than 80 MAM articles containing MAM processing conditions, materials, meltpool dimensions, meltpool modes and flaw types. We introduced physics-aware MAM featurization, versatile ML models, and evaluation metrics to create a comprehensive learning framework for meltpool defect and geometry prediction. This benchmark can serve as a basis for melt pool control and process optimization. In addition, data-driven explicit models have been identified to estimate meltpool geometry from process parameters and material properties which outperform Rosenthal estimation for meltpool geometry while maintaining interpretability.
1. Introduction
MAM quality remains difficult to control because defects arise from complex multiscale physics and strongly affect part integrity. MeltpoolNet addresses this challenge by combining experimental data, machine learning, and physics-aware features for meltpool characterization.
- MAM defects threaten printed-part structural integrity, while their complex multiscale physics makes quality improvement difficult.
- Meltpool characteristics are central to monitoring because lack of fusion, keyhole, and balling originate in meltpool dynamics.
- Machine learning can use experimental data to predict meltpool behavior and help identify efficient processing parameters.
- AM machine-learning development is constrained by scarce, expensive, heterogeneous data and the need to account for process–material interactions.
- MeltpoolNet unifies disparate meltpool-characterization sources into a larger repository spanning broader processing parameters and alloys than prior limited studies.
- The framework predicts meltpool geometry and flaw types using experimental data across varying parameters, materials, and PBF or DED processes.
2. Methodology
MeltpoolNet organizes raw experimental features, featurization, machine-learning models, and prediction tasks into one benchmark framework.
- The MeltpoolNet framework connects dataset raw features and featurization to implemented machine-learning models and associated tasks.
2.1. Data collection
The dataset was assembled from published experimental studies, with meltpool geometry and flaw or mode labels extracted primarily from figures and tables.
- MeltpoolNet collects meltpool geometry, flaw, and mode data from published experimental manufacturing and materials studies.
- Most source measurements were extracted from figures and tables using Plot Digitizer.
2.2. Datasets
The datasets span multiple AM processes, alloys, parameters, meltpool geometries, and defect modes, with mode labels linked to beam energy density and meltpool behavior.
- The literature-derived dataset contains about 2200 points with process parameters and material properties as inputs and meltpool geometry or mode labels as outputs.
- Beam input energy density, combining power, scanning speed, and diameter, distinguishes expected conduction and keyhole meltpool modes.
- Lack of fusion occurs when meltpool or overlap depth is below powder-layer thickness, while balling is associated with unstable or insufficiently bonded tracks.
- Regression predicts meltpool depth, width, and length, whereas classification predicts meltpool defect mode.
- The benchmark evaluates geometry and defect or mode classification across 29 alloys and multiple AM processes.
- PBF generally uses lower beam powers and higher scanning speeds than DED in the benchmark distributions.
2.3. Featurization
MeltpoolNet featurizes process, material, elemental, and absorptivity information to represent the physical factors governing melt pool predictions.
- Feature construction: One-hot encoding converts categorical AM processes into binary features so models can distinguish heat sources and feedstock relationships.
- Feature construction: Missing features are handled by removing samples lacking one or more variables required for a specified feature combination.
- Feature construction: Baseline features include beam power, scanning speed, beam diameter, and material properties such as density, heat capacity, conductivity, and melting temperature.
- Elemental features: Elemental featurization incorporates alloy composition using properties from 19 unique elements and a linear mixture rule.
- Absorptivity coefficients: Absorptivity coefficients are added as features for different melt pool regimes, and this addition significantly improves accuracy in some tasks.
2.4. Dataset splitting and Metrics
The benchmark normalizes heterogeneous inputs, evaluates models on held-out data using cross-validation, and uses task-specific regression and classification metrics.
- Preprocessing: Heterogeneous input parameters are normalized using their mean and standard deviation.
- Dataset splitting: Models are trained on a training partition and evaluated on a separate test partition reserved for performance assessment.
- Dataset splitting: Five-fold cross-validation averages model accuracy across five iterations, each using one group for testing and the others for training.
- Metrics: Regression uses mean absolute error and R2, while classification uses accuracy and AUC-ROC.
2.5. Models
MeltpoolNet benchmarks multiple machine learning models for melt pool regression and classification, including ensemble, probabilistic, and margin-based methods.
- Model benchmark: The benchmark evaluates models on melt pool geometry regression and defect or mode classification tasks.
- Random Forest: Random Forest combines independently trained decision-tree outputs by averaging regression predictions or taking majority votes for classification.
- Gaussian Processes: Gaussian Process models provide probability distributions over possible outputs and can perform well on small datasets.
- Support Vector Machines: Support Vector Machines classify by maximizing the margin between the separating hyperplane and nearby data points, and also support regression.
2.5.4. Ridge Linear Regression ’Ridge’
The benchmark includes regularized linear regression and tree-based ensemble models to address overfitting and nonlinear prediction in melt pool tasks.
- Ridge regression: Ridge regression applies L2 regularization by shrinking parameters to reduce model complexity and overfitting on small datasets.
- Regularization: LASSO applies L1 regularization and can nullify parameters to reduce model complexity and overfitting.
- Gradient boosting: Gradient boosting builds decision trees sequentially, with each new tree correcting errors from the preceding model.
2.5.7. Logistic Regression ’LR’
The section describes machine-learning models used for melt-pool regression and classification, including logistic regression, neural networks, and XGBoost. Model performance is evaluated with MAE, R2, AUC-ROC, and accuracy.
- Logistic regression predicts classes by applying a sigmoid function to a weighted combination of input features and comparing probabilities with a threshold.
- Neural networks support both regression and classification through input, hidden, and output layers whose connections are modified during training.
- XGBoost generates predictions with regularized gradient-boosted trees and is used for both regression and classification tasks.
- Model performance is evaluated using MAE, R2, AUC-ROC, and accuracy.
2.5.10. Description of Hyperparameter optimization
The study optimizes model hyperparameters that control complexity, regularization, tree count, kernels, and classification margins. Automatic selection is intended to support robust updates as data are added.
- Hyperparameters control model configuration and influence the parameters learned during training, making their selection important for prediction performance.
- Random forest and gradient boosting models vary n-estimators from 1 to 500, trading improved prediction performance against longer computation time.
- SVM optimization studies four kernels and varies regularization parameter C from 1 to 1000, controlling the tradeoff between classification errors and decision-boundary margin.
- Neural networks optimize layer width across 32, 64, 128, 256, and 512 neurons and tune alpha from 1e-7 to 1e-1 to constrain weight complexity.
- The optimization process can automatically select model configurations for ongoing dataset updates and additions.
3. Results and Discussion
MeltpoolNet benchmarks regression and classification models across melt pool geometry and defect tasks, showing that feature choices and model families strongly affect performance. Physics-informed features and explicit constitutive equations improve prediction while preserving interpretability.
- Regression: 93.32% accuracy was achieved by Random Forest with baseline features for melt pool depth, while Neural Network reached an MAE of 29.80 µm with baseline plus hatch spacing.Adding hatch spacing increased accuracy and reduced MAE relative to baseline for the reported depth results.
- Regression: 95.80% R2 accuracy was achieved by Neural Network using baseline features, the second absorptivity coefficient, and hatch spacing for melt pool depth.This configuration also produced an MAE of 20.54 µm.
- Regression: 97.73% accuracy was achieved by Gradient Boosting with baseline plus elemental features for melt pool width, while the minimum MAE was 18.56 µm with baseline plus hatch spacing.For the LPBF width dataset, Neural Network reached 97.32% accuracy with the second absorptivity coefficient, and the minimum MAE was 6.92 µm with that coefficient plus hatch spacing.
- Regression: 99.62% accuracy and an MAE of 13.25 µm were achieved by Gradient Boosting with baseline features plus hatch spacing for melt pool length.Adding hatch spacing raised accuracy and decreased MAE for the length prediction task.
- Classification: 88.42% accuracy and an AUC-ROC of 0.98 were achieved by Random Forest with both absorptivity coefficients and hatch spacing for melt pool classification.The baseline plus material one-hot configuration produced 85.78% accuracy before absorptivity features were added.
- Model comparison: Random Forest, Gradient Boosting, XGBoost, and Neural Networks consistently outperformed the other explored algorithms on datasets with complex decision boundaries and many features.The benchmark also investigated training-set size, while reported benchmark results were averaged over five runs with standard deviations.
- Model identification: The identified explicit constitutive models obtained higher R2 values than Rosenthal equations while retaining an interpretable explicit form.The melt pool depth equation also had a form similar to the Rosenthal estimate, consistent with the underlying physics.
4. Conclusion
The work establishes a comprehensive benchmark for melt pool geometry and defect prediction, combining diverse experimental data, feature engineering, machine-learning models, and evaluation practices. It also develops interpretable data-driven geometry models that outperform Rosenthal estimation.
- The benchmark covers melt pool geometry and defect-type prediction across diverse additive-manufacturing experiments.The collected data span varied alloys, processes, and processing parameters.
- Neural networks, gradient boosting, and random forest outperform the other evaluated machine-learning models for regression and classification.
- Feature engineering is important for obtaining accurate and generalizable machine-learning models in additive manufacturing.
- The identified explicit models estimate melt pool geometry more interpretably and more accurately than Rosenthal geometry estimation.