Uncertainty-aware hybrid framework for melt pool geometry prediction and process map development in Laser Powder Bed Fusion of Inconel 718

Melt pool geometry governs porosity formation, inter-layer bonding, microstructural evolution, and residual stress development in Laser Powder Bed Fusion (L-PBF), making its accurate prediction central to process qualification and process-window development. However, experimental qualification is resource-intensive, high-fidelity multi-physics simulations remain computationally expensive for large-scale parameter exploration, and reduced-order analytical models often lose accuracy near melting-regime transitions. Existing machine learning (ML) approaches provide deterministic predictions but do not explicitly address uncertainty or systematic prediction bias, limiting their reliability across different operating conditions and experimental datasets. This study presents an uncertainty-aware hybrid ML framework for melt pool geometry prediction and process qualification in L-PBF of Inconel 718 (IN718). Single-track and multi-track specimens were fabricated over a broad range of laser powers, scan speeds, hatch spacings, and layer counts and characterized using optical and scanning electron microscopy. The proposed framework combines k-fold ensemble regression with physics-informed directional classification, where the ensemble standard deviation serves as a physically interpretable uncertainty signal and is subsequently converted into a targeted bias-correction mechanism. The melt pool depth-to-width ratio (D/W) was adopted as a unifying descriptor linking melt pool morphology, thermal-gradient-driven grain growth direction, melting-regime transitions, and process-map development. The framework was trained using 68 experimentally measured melt pools and evaluated using an external dataset of 54 measurements compiled from published L-PBF IN718 studies, including validation and independent test datasets. Compared with established analytical models and polynomial curve-fitting baselines, the proposed approach improved predictive accuracy, enhanced cross-study generalization, and provided useful ensemble-based uncertainty estimates. Multi-track experiments further validated the maximum hatch spacing criterion for predicting lack-of-fusion onset from single-track measurements, achieving prediction errors below 6% and enabling the identification of defect-free processing windows. The results demonstrate that ensemble prediction uncertainty contains information related to regime-dependent model sensitivity and can be exploited to improve prediction reliability near conduction-transition-keyhole boundaries. Beyond the specific IN718 case, the proposed framework provides a potentially transferable strategy for uncertainty-aware process qualification and rapid process-window development and offers a computationally efficient foundation for digital twin development, in-situ monitoring, and future closed-loop control of metal additive manufacturing (AM).