Despite the recent breakthroughs in protein structure prediction, our understanding of protein function remains limited, as biological activity is inherently linked to dynamic behavior. Although traditional physics-based simulations can capture these dynamics, they are computationally expensive and scale poorly. Recently, generative models have emerged as promising surrogates that can accelerate simulations by learning from existing trajectories. However, their ability to generalize to unseen systems and capture long-timescale motions remains limited. During the PhD, we are going to explore scalable generative frameworks capable of generating long simulations for unseen systems while preserving physical realism.