Abstract
<title>Abstract</title> <p>Laser-based Direct Energy Deposition (DED-LB) is a key technology in both manufacturing and repair of components within metal Additive Manufacturing (AM). However, challenges such as heat accumulation and insufficient dynamic process control restrict its broader adoption. This work proposes a novel strategy to stabilize the deposition process by integrating a Machine Learning (ML) model as a ML-based surrogate controller, coupled with a Finite Element (FE) solver for run-time process control. The FE framework performs thermal analysis of the DED-LB process and generates High-Fidelity (HF) synthetic data through an offline optimization approach for training the ML model. The architecture of the ML model is optimized via hyperparameter tuning and validated on an independent case outside the training dataset. The resulting ML-based surrogate controller, coupled with the FE solver, analyzes melt pool morphology and thermal profiles in run-time and outputs corrective laser power adjustments to maintain a constant melt pool penetration depth. This closed-loop system enables autonomous control and rapid dynamic response, ensuring consistent thermal management throughout the deposition process. Three scanning-sequence scenarios are presented to evaluate the performance of the run-time control system. The results indicate that integrating this framework maintains a stable melt pool penetration depth, thereby enhancing geometric precision and reliability in DED-LB through tailored time-series power profiles, while reducing the computation time to one-third relative to the offline optimization approach. Among the three sequences, two were not included in the training dataset of the ML model, and the accurate control of these cases demonstrates the robust generalization capability of the ML-based surrogate controller and confirms its suitability for precise and scalable control of metal AM processes.</p>