Application of Reduce Order Modeling to Time Parallelization

We recently proposed a new approach to parallelization, by decomposing the time domain, instead of the conventional space domain. This improves latency tolerance, and we demonstrated its effectiveness in a practical application, where it scaled to much larger numbers of processors than conventional parallelization. This approach is fundamentally based on dynamically predicting the state of a system from data of related simulations. In earlier work, we used knowledge of the science of the problem to perform the prediction. In complicated simulations, this is not feasible. In this work, we show how reduced order modeling can be used for prediction, without requiring much knowledge of the science. We demonstrate its effectiveness in an important nano-materials application. The significance of this work lies in proposing a novel approach, based on established mathematical theory, that permits effective parallelization of time. This has important applications in multi-scale simulations, especially in dealing with long time-scales.