This work presents a comprehensive multiscale modelling framework to predict the compressive response of lattice structures produced by Powder Bed Fusion, with a focus on the degrading influence of internal manufacturing defects. The approach is applied to AlSi10Mg octet truss lattices and integrates detailed defect characterization, micromechanical modelling, and machine learning acceleration to enable efficient, defect-aware predictions. Starting from micro-CT scans of lattice specimens, the internal defect population is statistically characterized, including not only defect size and location but also aspect ratio, orientation, and surface roughness. Finite Element Analyses (FEAs) are performed at the microscale on Representative Volume Elements (RVEs), modelling struts with individual defects, to determine their local compressive stress–strain response. These responses are then used to inform mesoscale lattice models, where each strut’s behaviour reflects a stochastic assignment based on the defect-informed database. To address the computational cost of high-fidelity microscale simulations, a machine learning surrogate model is introduced. Trained on the FEA dataset, it accurately predicts the mechanical response of RVEs based on defect geometry and surface features, enabling rapid generation of effective material behaviour for use in the larger-scale model. The framework captures the statistical variability induced by internal defects and provides physical insight into the role of defect morphology and roughness on structural performance. Validation against experimental compression tests on 2 × 2 × 2 and 3 × 3 × 3 lattice specimens confirms the model’s predictive capabilities. This integrated approach offers a powerful tool for understanding and mitigating defect sensitivity in metal additive manufacturing, supporting defect-tolerant design and structural optimization.