ML Enhances Predictions in Metal Additive Manufacturing

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 19 2024

In a paper published in the journal Additive Manufacturing, researchers introduced a comprehensive framework for benchmarking machine learning (ML) models in predicting mechanical properties of metal additive manufacturing (MAM). They compiled a vast dataset from over 90 MAM studies, including 140 data sheets with detailed processing conditions and material properties.

The framework featured physics-aware featurization, adjustable ML models, tailored evaluation metrics, and integrated shapley additive explanations (SHAP) analysis for model interpretability. Data-driven explicit models were also developed for better interpretability than conventional ML approaches.

Related Work

Past work has demonstrated the potential of ML in MAM, but often with limited datasets. Research has concentrated on predicting mechanical properties with different ML techniques, such as convolutional neural networks (CNNs), extreme gradient boosting (XGBoost), and smaller, focused datasets. These attempts could have been more robust in their generalisability and application due to the limited scope of data and processing parameters, notwithstanding their contributions.

Predicting Mechanical Properties

The mechanical properties network (MechProNet) framework integrates data collection, feature extraction, and ML models to predict the mechanical properties of additively manufactured components. Empirical data were gathered from manufacturing and materials journals and datasheets from AM companies.

Sources provided the primary data, with accuracy ensured using Plot Digitizer to extract information from figures and plots. The dataset includes processing parameters and material properties from experiments, which serve as input features for the ML models.

The benchmark dataset, derived from literature, consists of 1,600 data points detailing processing parameters, material characteristics, AM processes, machine types, post-processing conditions, and mechanical property orientations. Mechanical properties, such as yield and tensile strength, serve as labels.

This dataset encompasses a range of experiments with various materials and techniques across multiple AM processes. The dataset includes diverse processes like powder bed fusion (PBF) and directed energy deposition (DED), with specific sub-processes and machine types noted. It also captures the mechanical properties of metals in different post-processing conditions.

Featurization involves selecting and encoding input features for ML models, including processing parameters, material properties, and categorical features such as material type and machine type. Categorical features are one-hot encoded to convert them into numerical categories suitable for ML algorithms. Additionally, materials are featured based on chemical composition and elemental properties. Data points with incomplete feature information are excluded to ensure uniform input for the models.

Various ML models are applied to the dataset, including random forest (RF), gradient boosting (GB), support vector regression (SVR), Gaussian process regression (GPR), lasso and ridge regression, NN, and XGBoost. For forecasting mechanical qualities, each model has a special set of benefits, such as robustness, the ability to handle non-linear interactions, or the ability to quantify uncertainty.

Hyperparameter optimization improves model performance by applying techniques like random search, grid search, and Bayesian optimization. Bayesian optimization utilizes past evaluations to refine hyperparameters more effectively.

Results and Discussion

This section examines the predictive performance of various ML models applied to a benchmark dataset. Results are averaged across five iterations, with standard deviations as error bars. The evaluation includes a correlation matrix of mechanical properties and a drop-column feature importance analysis to identify key features for each dataset and task. Analysts explored how training set size affects model performance and identified data-driven explicit models for mechanical properties.

The effectiveness of ML models, including RF regressor, GP regressor, SV regressor, and others, is analyzed for predicting yield strength, ultimate tensile strength, elastic modulus, and elongation at break. The team initially used baseline featurization and explored additional methods like chemical composition and elemental featurization to enhance prediction accuracy.

RF models with optimized features consistently performed best, achieving high accuracy and low mean absolute error (MAE) across all tasks. The models’ performance was also assessed for Vickers and Rockwell hardness and roughness, demonstrating that feature selection and hyperparameter optimization significantly improved model outcomes.

A drop-column feature importance analysis identified post-processing conditions and thermal properties as crucial features influencing mechanical property predictions. XGBoost feature importance analysis and SHAP were employed to provide interpretable insights into feature contributions.

SHAP plots revealed detailed relationships between features and predictions, while the team examined the impact of dataset size on accuracy, showing that larger training sets lead to better model performance. Finally, data-driven models were identified to provide explicit, interpretable relationships among processing parameters, material properties, and mechanical properties.

Conclusion

To sum up, this study established a comprehensive ML benchmark for predicting the mechanical properties of additively manufactured parts using a diverse dataset covering various MAM processes, materials, and machines. Researchers evaluated several featurization techniques, finding that RF, GB, and NN consistently performed better.

They also enhanced interpretability with SHAP analysis and developed a data-driven model to establish explicit relationships between processing parameters and material properties. The MechProNet benchmark offers a standardized platform to optimize additive manufacturing processes and supports the MAM ML community.