Deep learning-based prediction of interfacial properties in multi-material laser powder bed fusion structures

• A synthetic dataset is generated using a deep autoencoder network. • Proposed MatAttn-Net achieved a low MSE of 0.29 and 0.34 for interfacial properties prediction. • Traditional ML models are suitable for a limited real dataset. • MatAttn-Net outperformed ML techniques by 40–50% on the synthetic dataset. • Optimal model selection reduces prediction errors by up to 60% in LPBF processes. Advancing multi-material Laser Powder Bed Fusion (LPBF) requires precise control of process parameters; however, optimization remains challenging due to the high experimental costs and time associated with data collection. These limitations hinder the development of data-driven additive manufacturing (AM) systems powered by artificial intelligence (AI). This work introduces a deep learning (DL) framework that integrates two core components: a synthetic material autoencoder (SynthMat-AE) that generates high-fidelity training data and a material attention network (MatAttn-Net) developed to predict key interfacial properties in Al6061/AlSi10Mg bimetallic structures. MatAttn-Net leverages convolutional layers for spatial feature extraction alongside a custom attention mechanism to capture interfacial complexity. Validation through hold-out and cross-dataset evaluations demonstrates that the proposed approach outperforms conventional machine learning (ML) methods, such as Random Forest (RF), Gradient Boosting (GB), and Support Vector Regression (SVR), in predicting porosity and interfacial shear strength. This data-efficient framework minimizes experimental dependency, enhances predictive accuracy, and supports the optimization of multi-material LPBF processes in intelligent manufacturing systems.