Machine Learning-Based Process Optimization for Directed Energy Deposition of Aerospace Components

To address the high experimental costs and data scarcity inherent in Directed Energy Deposition (DED), this study proposes a data-efficient hybrid optimization framework for the precision manufacturing of Inconel 718 aerospace components. The framework leverages a two-stage strategy to bridge traditional experimental design with advanced machine learning, ensuring robust process optimization even with limited datasets. In the first stage, the Taguchi method (L16 orthogonal array) was employed for coarse-grained screening to identify influential control factors. In the second stage, a Fully Connected Neural Network (FNN) coupled with Bayesian Optimization (BO) was deployed. Crucially, this machine learning component functions as an optimization-oriented trend surrogate rather than a global regressor, successfully guiding the optimization under extreme data scarcity. The optimized process window yielded exceptional structural integrity, achieving a porosity as low as 0.03%. To thoroughly validate its practical efficacy, tensile testing (ASTM E8/E8M) and Rockwell hardness measurements (ASTM E18) were systematically conducted on the optimized specimens. The mechanical characterization demonstrated an average tensile strength of approximately 1358 MPa and a hardness of ~40 HRC. Finally, the framework was successfully validated through the robotic DED fabrication of a complex-geometry aerospace engine combustion chamber casing, bridging laboratory-scale optimization with authentic industrial applications.