The quality of metal parts manufactured by laser powder bed fusion (LPBF) is primarily governed by process-induced defects, including surface roughness across layers and scan tracks, as well as porosity originating from insufficient melting or gas entrapment during vaporization. These defects arise from inherent variability in melt pool formation caused by strong sensitivity to process parameters such as laser power, scan speed, and energy absorption. As a result, deterministic computational models face fundamental limitations in capturing scan-track irregularities and defect variability, while fully resolved powder-scale simulations remain computationally prohibitive for real-time or part-scale applications. This work presents a statistical physics-based digital shadow framework for efficient and predictive simulation of surface roughness and porosity in LPBF. The governing equations are solved using an in-house C++ finite volume solver AM-CFD, which captures melt pool flow driven by Marangoni effects while accounting for stochastic variations in energy input. A mechanistic reduced-order-based stochastic calibration strategy is introduced to capture the intrinsic variability of melt pool dynamics under multi-layer and multi-track scanning conditions. To enable fast and scalable predictions suitable for online-ready simulation and control, the framework leverages tensor decomposition to construct non-intrusive reduced-order surrogate models. The calibrated physics-based digital shadow model is demonstrated on the NIST overhang part (X4), achieving a 9.3% difference in predicted roughness compared to experimental measurements. By combining data generated from the physics-based model and experiments, a machine-learning surrogate is further trained for rapid inference, achieving prediction times on the order of 0.4 ms with high accuracy. Together, the physics-based and data-driven components form a digital shadow of the LPBF process that supports calibration, fast prediction, and defect-aware control. The proposed framework bridges high-fidelity physics-based modeling and near-real-time surrogate inference, providing a Software 3.0-ready computational foundation for digital twin development, process control, and online monitoring in additive manufacturing.