Despite significant advancements in Additive Manufacturing (AM) over the past decade, numerous scientific challenges remain to be addressed to establish AM as a reliable manufacturing process and facilitate its widespread industrial adoption. The AM process involves a multitude of complex parameters that must be precisely monitored and controlled to achieve the desired level of accuracy. In this context, Artificial Intelligence (AI), particularly Machine Learning (ML) techniques, offer promising solutions for various aspects of AM, including design optimization, process parameter selection, performance enhancement, in-situ process monitoring and control, real-time anomaly detection, and post-process evaluation. The integration of AI-driven methodologies has the potential to significantly enhance the overall design and manufacturing workflow in AM. Our current research focuses on developing advanced AI-driven tools and methodologies for predicting temperature field and history at the scale of the manufactured part, representing a crucial initial step toward optimizing and controlling the AM process. The proposed model aims to deliver significantly faster temperature history predictions compared to conventional finite element models while maintaining a high level of accuracy. We propose a hybrid modelling approach that combines Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs) to predict temperature histories across a range of process parameters and part geometries. GNNs are particularly well-suited for representing physical fields defined over complex, irregular meshes, such as temperature distributions on part geometries. RNNs are employed to ensure stable and accurate temporal predictions over long durations. The model is initially trained on simple and small-scale two-dimensional part geometries, where the short simulation time enables the generation of a diverse dataset with limited computational cost, thereby reducing constraints related to data availability. It is then employed to predict temperature histories for larger and more complex geometries, demonstrating its ability to generalize beyond the original training set. Subsequent fine-tuning on a subset of complex geometries further enhances the model’s predictive accuracy and robustness.