Additive manufacturing (AM) is an evolving manufacturing technology that has gained significant attention due to its unparalleled design freedom, material efficiency, and potential for producing complex, high-performance components. Among various AM processes, laser powder bed fusion (LPBF) stands out as a promising metal AM technique for industrial applications. However, despite its rapid development, LPBF is not yet fully mature for first-time-right, high-quality production, as multiple process-related factors can adversely affect part quality. These challenges are typically addressed through costly and time-consuming trial-and-error approaches for process optimisation. A more systematic and physics-based optimisation of LPBF requires a deeper understanding of the underlying process mechanisms, which can be achieved through high-fidelity numerical simulations. Such simulations must capture the complex, highly nonlinear, and strongly coupled multiphysical phenomena involved in LPBF. Accurately resolving the fast-evolving and highly localised phenomena during LPBF requires extremely fine spatial and temporal discretisation. As a consequence, high-fidelity LPBF simulations incur prohibitively high computational costs, often forcing researchers to reduce domain sizes. This work discusses the potential of scientific machine learning to overcome the computational challenges associated with high-fidelity LPBF simulations. By integrating physics-based models with machine learning techniques, it is possible to significantly accelerate simulations while preserving accuracy and physical consistency.