Porosity remains one of the most consequential defects in laser powder bed fusion (LPBF) because even small internal voids can degrade fatigue resistance, fracture performance, leak-tightness, and long-term reliability in safety-critical mechanical components. Purely empirical machine-learning models have shown promise for defect detection, yet they often depend heavily on machine-specific datasets and can lose interpretability outside the training domain. This manuscript develops a physics-informed machine-learning (PIML) framework for porosity prediction in LPBF by integrating process parameters, in-situ monitoring signals, and physically meaningful descriptors such as line energy, volumetric energy density, scan-overlap ratio, Peclet number, cooling-rate proxy, and mechanism-sensitive indicators for lack-of-fusion and keyhole instability. A hybrid learning objective is formulated to combine statistical prediction accuracy with soft physics penalties that discourage mechanistically inconsistent outputs. The paper further proposes a risk-based decision logic suitable for critical-component qualification, where predicted porosity must be interpreted together with uncertainty, part zoning, and inspection requirements. An analytical demonstration is included to show how the framework distinguishes stable, lack-of-fusion-prone, and keyhole-prone process windows. The contribution of the study is therefore methodological: it offers a publication-ready formulation that connects LPBF process physics, monitoring, and machine learning in a single, interpretable workflow that can later be validated with computed tomography, metallography, and mechanical testing data.Porosity remains one of the most consequential defects in laser powder bed fusion (LPBF) because even small internal voids can degrade fatigue resistance, fracture performance, leak-tightness, and long-term reliability in safety-critical mechanical components. Purely empirical machine-learning models have shown promise for defect detection, yet they often depend heavily on machine-specific datasets and can lose interpretability outside the training domain. This manuscript develops a physics-informed machine-learning (PIML) framework for porosity prediction in LPBF by integrating process parameters, in-situ monitoring signals, and physically meaningful descriptors such as line energy, volumetric energy density, scan-overlap ratio, Peclet number, cooling-rate proxy, and mechanism-sensitive indicators for lack-of-fusion and keyhole instability. A hybrid learning objective is formulated to combine statistical prediction accuracy with soft physics penalties that discourage mechanistically inconsistent outputs. The paper further proposes a risk-based decision logic suitable for critical-component qualification, where predicted porosity must be interpreted together with uncertainty, part zoning, and inspection requirements. An analytical demonstration is included to show how the framework distinguishes stable, lack-of-fusion-prone, and keyhole-prone process windows. The contribution of the study is therefore methodological: it offers a publication-ready formulation that connects LPBF process physics, monitoring, and machine learning in a single, interpretable workflow that can later be validated with computed tomography, metallography, and mechanical testing data.