In metal additive manufacturing (AM), material microstructures with tailored mechanical properties can be fabricated by appropriately controlling the laser scan strategy (SS) [1]. Therefore, clarifying the relationship between SS and the resulting microstructure is essential for achieving high-precision fabrication. However, because the possible combinations of SS are virtually infinite, it is impractical to evaluate them exhaustively through experiments alone, making numerical simulation indispensable. For grain-scale microstructure prediction in AM, both the cellular automaton (CA) method and the multi-phase-field (MPF) method have been employed. The CA method enables large-scale simulations and has been applied to evaluate microstructural differences arising from various SS [2]. However, its predictive accuracy remains limited. In contrast, although the MPF method provides high accuracy, its high computational cost has made large-scale microstructure prediction difficult. We have enabled large-scale microstructure simulations by combining multi-GPU parallel computing with a moving frame algorithm that restricts the computational domain to the vicinity of the meltpool. Using this approach, we have evaluated microstructural differences arising from two different SS [3]. However, the computational cost remains high, and systematic comparison of SS effects under multi-layer and multi-track conditions is still challenging. In this study, to enable microstructure prediction under different SS using the MPF method, we develop a subdomain moving frame algorithm that significantly improves computational efficiency by restricting the calculation domain to the vicinity of the meltpool boundary. Using the proposed method, microstructure predictions are performed for multiple SS, and the resulting microstructural differences under multi-layer and multi-track conditions are systematically evaluated.