Smart Process Design with Machine Learning for Quality Assurance in Metal Additive Manufacturing

Abstract

Additive Manufacturing (AM), particularly Laser Powder Bed Fusion (LPBF), has revolutionized the fabrication of complex metal components through its layer-by-layer construction capability. However, the widespread adoption of LPBF remains limited by recurring defects such as porosity, residual stresses, and deformation, which compromise the structural integrity and dimensional accuracy of printed parts. Addressing these challenges, this dissertation presents three complementary methodologies: (1) layer-wise printing parameter optimization, (2) reinforcement learning-enabled scan path planning, and (3) a multi-fidelity Bayesian optimization framework for efficient process parameter tuning. Collectively, these approaches enhance process control and reduce defects, advancing LPBF toward greater reliability and industrial applicability. The first methodology introduces a novel optimization framework that dynamically adjusts laser power on a layer-by-layer basis. By modeling laser power as a polynomial function of build height and reducing the optimization space using weighted basis functions, the approach accounts for thermal and geometric variations throughout the build. Employing Response Surface Methodology, the framework minimizes a composite defect index incorporating porosity, residual stress, and displacement. Case studies demonstrate significant improvements in defect mitigation, including a 19.57% reduction in porosity compared to constant parameter settings, validated through both simulation and experimental results. The second methodology, Reinforced Scan, leverages Deep Q-Learning to optimize laser scan sequences for improved thermal uniformity. The optimization is structured hierarchically and guided by a custom reward function that balances temperature variance, spatial discrepancy, and high-temperature cluster distance. Simulations verified using finite element models, along with experimental validation on Ti64 substrates, confirm that Reinforced Scan outperforms conventional approaches, including sequential scans and heuristic strategies such as SmartScan and Least Heat Influence. The third methodology presents a Multi-Fidelity Bayesian Optimization (MFBO) framework. By integrating low-fidelity thermal approximations with high-fidelity simulations via Gaussian Process-based surrogate models, the framework delivers substantial computational efficiency without compromising accuracy. A decoupled acquisition strategy—using Upper Confidence Bound for high-fidelity sampling and variance-based selection for low-fidelity exploration—enables effective parameter tuning. Applied to the regulation of end-of-cycle temperatures, MFBO demonstrates faster convergence and superior sample efficiency compared to single-fidelity optimization. Together, these methodologies offer a robust and scalable strategy for improving LPBF process reliability and part quality. Layer-wise optimization targets inter-layer thermal gradients and porosity, while Reinforced Scan enhances intra-layer thermal uniformity to minimize residual stress. The integration of machine learning and advanced optimization establishes a foundation for intelligent, adaptive LPBF systems, with significant implications for high-performance applications in aerospace, biomedical, and automotive manufacturing.