The efficient modeling of spatio-temporal dynamics is a relevant topic in scientific computing and an enabling methodology in personalized medicine. This talk focuses on recent developments in Latent Dynamics Networks (LDNets), which provide lightweight, accurate, and scalable numerical simulations for complex dynamical systems with spatio-temporal features. These advancements bridge traditional physics-based and data-driven approaches, enabling real-time, scalable simulations for computational cardiology modeling and beyond. The key application of this talk is in cardiac digital twins, where LDNets facilitate real-time whole-heart electromechanical simulations. By learning a compact representation of cardiac models, we demonstrate through meaningful examples how these models dramatically reduce computational costs while retaining high fidelity. This enables efficient global sensitivity analysis, parameter estimation, and uncertainty quantification, all on standard computational hardware. Building upon this foundation, recent extensions to 2nd-order latent dynamics neural networks (LDNets2) further enhance the modeling of temporal dependencies by introducing 2nd-order latent operators that capture acceleration-like effects in the system’s hidden state evolution. These higher-order dynamics significantly improve long-term predictive stability and accuracy, particularly in the context of cardiac electrophysiology, where accurate modeling of rapid electrical waves is essential. The integration of LDNets2 into cardiac digital twins opens new perspectives for robust, high-fidelity forecasting of arrhythmias, strengthening the potential of these models to support personalized diagnosis, risk assessment, and therapy optimization. We acknowledge the HORIZON-EUROHPC-JU-2023-COE-03 project dealiiX "an Exascale Framework for Digital Twins of the Human Body" (no. 101172493), 2024-2026.