Modeling and prediction of 316 L stainless steel relative density in L-PBF process using machine learning

Area of Science:

• Materials Science and Engineering
• Additive Manufacturing
• Computational Materials Science

Background:

• Laser-Powder Bed Fusion (L-PBF) is a key additive manufacturing process for producing metal parts.
• Predicting and controlling the relative density of L-PBF parts is crucial for ensuring mechanical integrity and performance.
• Existing models often struggle to capture the complex, nonlinear relationships between L-PBF process parameters and part density.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting the relative density of L-PBF parts.
• To integrate multiple predictive models into a hybrid framework for enhanced prediction accuracy.
• To identify key process parameters influencing relative density in L-PBF.

Main Methods:

• A literature-based dataset of 287 experimental measurements was compiled.
• Six machine learning models were developed: k-nearest neighbours (KNN), adaptive boosting decision trees (AdaBoost-DT), and four multilayer perceptron (MLP) variants (SCG, LM, BR, RB).
• A committee machine intelligence system (CMIS) was used to integrate the predictions from the individual models.

Main Results:

• The hybrid CMIS framework demonstrated improved predictive performance compared to individual models.
• The models effectively captured the complex, nonlinear relationship between laser power, scanning speed, hatch spacing, and relative density.
• Statistical metrics (R², SD, APRE, AAPRE, RMSE) were used to evaluate model accuracy.

Conclusions:

• The proposed hybrid modelling framework provides an effective approach for predicting relative density in L-PBF parts.
• The study highlights the potential of machine learning in optimizing additive manufacturing processes.
• Dataset heterogeneity and outliers in literature data can introduce uncertainty in predictive models.