Exascale computing, i.e., supercomputers that can perform 1018 math operations per second, provide significant opportunity for improving the computational sciences. That said, these machines can be difficult to use efficiently, due to their massive parallelism, due to the use of accelerators, and due to the diversity of accelerators used. All areas of the computational science stack need to be reconsidered to address these problems. With this dissertation, we consider flow visualization, which is critical for analyzing vector field data from simulations. We specifically consider flow visualization techniques that use particle advection, i.e., tracing particle trajectories, which presents performance and implementation challenges. The dissertation makes four primary contributions. First, it synthesizes previous work on particle advection performance and introduces a high-level analytical cost model. Second, it proposes an approach for performance portability across accelerators. Third, it studies expected speedups based on using accelerators, including the importance of factors such as duration, particle count, data set, and others. Finally, it proposes an exascale-capable particle advection system that addresses diversity in many dimensions, including accelerator type, parallelism approach, analysis use case, underlying vector field, and more.