Injection molding is a widely used manufacturing process for producing polymer components with complex geometries and tight dimensional requirements. Predictive simulation of the process is essential for design and process optimization, but remains computationally expensive due to the strong coupling of multiple physical phenomena. Accurate modeling requires resolving non-isothermal flow of non-Newtonian materials with moving free surfaces, together with pressure-volume-temperature relations and solidification driven shrinkage and warpage. High resolution discretizations of thin-walled geometries further increase computational cost and limit large-scale design exploration. This work presents a data-driven modeling framework based on Anchored-Branched Universal Physics Transformer [1, 2] for fast prediction of injection molding processes across a wide range of geometries, materials, and processing conditions. The proposed approach learns the evolution of key physical quantities during filling and packing, while capturing thermal effects and deformation mechanisms associated with cooling. The method supports heterogeneous inputs including three dimensional geometries, full material characterization data, and process parameters such as melt temperature, mold temperature, flow rate, and gate location. The model is trained using a large dataset (order 100 terabytes) of high fidelity numerical simulations spanning extensive variability in geometry and process conditions, with mesh resolutions reaching up to one million nodes. Results demonstrate robust generalization to previously unseen configurations and accurate prediction of flow and thermal fields. The approach provides simulation results up to two orders of magnitude faster than conventional numerical solvers. The proposed framework enables scalable modeling of complex injection molding processes and supports accelerated design iteration and optimization in computational mechanics applications.