<p>Additive manufacturing (AM) particulate composites offer customizable properties and lightweight structures, making them valuable across engineering applications. However, accurately predicting their mechanical behavior remains challenging due to the complex interplay between manufacturing parameters, reinforcement particles, and matrix materials. This study proposes a deep learning framework for predicting and optimizing the mechanical properties of AM particulate composites using a temporal attention recurrent graph convolution network (TAR-GCN). The input dataset, sourced from publicly available composite materials data, includes various formulations and their corresponding mechanical properties. A regularized bias-aware ensemble Kalman filter (RBAEKF) is first applied for pre-processing to reduce noise and handle missing values. The refined data are then used to train the TAR-GCN model, which captures both spatial and temporal dependencies among AM parameters. The model is evaluated using accuracy, precision, and error rate, and its performance is benchmarked against baseline models, including ANN, traditional ML, and PIML methods. The proposed approach demonstrates superior predictive capability, offering a scalable solution to reduce experimental costs and enhance composite design.</p>