<p>Laser powder bed fusion (LPBF) remains limited by the absence of integrated control systems capable of predicting and mitigating defect formation during fabrication. This study presents a physics-informed, feed-forward machine learning control framework that combines synchronized acoustic emission and pyrometer data for defect-aware process adjustment in LPBF. The proposed architecture integrates multimodal sensor acquisition, variational autoencoder-based feature encoding, sensor fusion, monitoring, predictive control, and decoder-based simulation within a unified workflow. Acoustic and thermal signals are transformed into reduced-dimensional latent representations, which are used to estimate defect likelihood and to generate candidate process updates for laser power and scan speed. These candidate process-structure-property states are evaluated through a simulation-assisted validation loop before machine write-back. Experimental validation was conducted on Inconel 625 using defect-inducing and nominal parameter sets across multiple geometry conditions, including thin walls, overhangs, enclosed features, and supports. The framework demonstrated stable convergence toward defect-reduced operating regions, strong multimodal agreement between acoustic and thermal responses, and inference speeds compatible with the controller. The results establish the feasibility of integrating synchronized multimodal sensing, machine learning, and physics-based constraints into a predictive control strategy for LPBF.</p>