Abstract
The increasing application of Additive Manufacturing (AM) in key industries requires reliable predictions of Material Properties (MP) to ensure consistent part quality and performance. The complex relationships between Process Parameters (PP) and MP, along with the inherent uncertainty in powder-based AM methods, render reliable property prediction challenging. This paper presents a novel Multistage Transfer Learning Model (MTLM) for predicting MP in powder bed fusion-based AM. A model is recommended that combines Crystal Graph Convolutional Neural Networks (CGNN) and Bayesian Neural Networks (BNN) to correlate PP and MP with final properties. The model is proved on Titanium alloy 6 − 4 (Ti-6Al-4 V) as the test material, and it shows better prediction accuracy compared with traditional Machine Learning (ML) algorithms, with Root Mean Square Errors (RMSE) of 11.7 megapascal (MPa) for Ultimate Tensile Strength (UTS), 8.9 MPa for Yield Strength (YS), and 0.021% for porosity. The model incorporates uncertainty quantification through Bayesian inference, providing confidence metrics important for industrial applications. Trained on 3,083 experimental samples and validated across different combinations of process parameters, the model demonstrates strong generalization, achieving MP prediction accuracies of over 93%. Real-time processing capabilities are proven by integrating big data analytics platforms to enable dynamic optimization of AM process parameters. The ability of the model to capture complex process-structure-property relationships, along with the quantification of prediction uncertainties, is a significant phase in computational materials science for AM.