Digital twin is generally defined as a digital representation of physical objects in the real world, stored in cyberspace and used to simulate processes and consequences of target phenomena. Recognizing the importance of this concept, we propose Tsunami Digital Twin (TDT) as a new paradigm in tsunami science and engineering aimed at enhancing tsunami disaster resilience. We report recent progress in TDT applications and practical implementations. Current TDT developments in Japan focus on multi-platform computing capabilities to extend tsunami forecasting technologies to other countries. As part of this effort, we have launched a new project, “TsunamiCast,” which aims to construct both a fully cloud-based and on-premises end-to-end tsunami inundation forecasting facility for at-risk coastal communities. The standard TsunamiCast infrastructure integrates two kinds of urgent computing platforms of cloud computing systems and on-premises servers with GPU computing capabilities. In Japan, activation of TsunamiCast begins with seismic information from the Japan Meteorological Agency’s Earthquake Early Warning (EEW) service, available within seconds of the first seismic observations. Refined source estimates from real-time GNSS analysis and offshore tsunami data assimilations follow this. Fault rupture parameters—including moment magnitude, fault geometry, focal mechanisms, and slip distribution—are derived using the RAPiD and REGARD algorithms. With these inputs, the TsunamiCast executes tsunami propagation and inundation simulations on the SX-Aurora supercomputer (AOBA). Outputs include travel and arrival times, inundation extent, maximum flow depth distribution, and loss estimates. The simulations use the TUNAMI-N2 code, based on nonlinear shallow-water equations discretized with a staggered leap-frog finite difference method. On SX-Aurora, tsunami source and inundation modeling each complete in ~10 minutes at 10-m resolution. The system has since been extended to diverse high-performance computing (HPC) platforms. Takahashi et al. (2024) migrated the code to modern CPUs and GPUs using a directive-based approach, maintaining the original structure while improving performance portability. Hardware-specific optimizations, including GPU load balancing, further enhance efficiency. A six-hour tsunami simulation with over 47 million grids now completes in under 2.5 minutes on 32 Intel Sapphire Rapids CPUs and 1.5 minutes on 32 NVIDIA H100 GPUs. These advances enab