The aim of this project is to develop smarter methods for planning and controlling metal 3D printing, with a focus on the Freemelt ONE electron beam melting system hosted at LiU (IEI). By combining physics-informed graph neural network models for thermal prediction with optimisation-based approaches for controlling energy input, the project aims to support more precise and stable manufacturing. Existing process simulation approaches are either not fast enough to be useful in the manufacturing process stage or show systematic deviations from real world measurements. The goal is therefore to learn data-driven corrections while preserving speed en the thermal simulations. The current baseline is a non-autoregressive GNN using one graph per build layer in the printer. The next step is to develop a spatio-temporal, autoregressive GNN in which corrected temperatures are propagated between layers, allowing the model to capture heat accumulation, cooling history, and geometry-dependent thermal behaviour. The model will be evaluated using build-level and geometry-level cross-validation to assess its ability to generalise to unseen builds and geometries.