Solidification grain structure has significant impact on the final properties of welded parts using fusion welding processes. Direct simulation of grain structure at industrial scale is yet rarely reported in the literature and remains a challenge. A three-dimensional (3D) coupled Cellular Automaton (CA) – Finite Element (FE) model is presented that predicts the grain structure formation during multiple passes Gas Tungsten Arc Welding (GTAW) and Gas Metal Arc Welding (GMAW). The FE model is established in a level set (LS) approach that tracks the evolution of the metal-shielding gas interface due to the addition of metal. The FE method solves the mass, energy and momentum conservation equations for the metal plus shielding gas system based on an adaptive mesh (FE mesh). Fields are projected in a second FE mesh, named CA mesh. A CA grid made of a regular lattice of cubic cells is created to overlay the fixed CA mesh. The CA model based on the CA grid simulates the melting and growth of the grain boundaries in the liquid pool. In order to handle large computational domains while keeping reasonable computational costs, parallel computations and dynamic strategies for the allocation/deallocation of the CA grid are introduced. These strategies correspond to significant optimizations of the computer memories that are demonstrated. The 3D CAFE model is first applied to the simple configuration of single linear passes by GTAW of a duplex stainless steel URANUS 2202. It is then applied to a more persuasive example considering GMAW in spray transfer mode during multiple passes to fill a V-groove chamfer. Simulations reveal the possibility to handle domains with millions of grains in representative domain sizes while following the formation of textures that result from the growth competition among columnar grains