Scientists often desire interactive visual analytics services to efficiently and effectively study their large-scale scientific data generated from simulations or observations. However, as the volume of scientific data is growing exponentially, it becomes increasingly difficult to achieve this goal for a typical interactive visual analytics system nowadays. The bottlenecks in visual analytics processes manifest in fetching time series data in a continuous manner. Since the changes in scientific datasets over a period of time are usually small and continuous, it is possible to learn an optical flow based representation of such dynamics. Therefore, the intermediate time steps of data can be efficiently inferred at run time. However, the existing optical flow determination methods cannot be directly applied to scientific datasets due to the highly complex non-rigid transformations in the feature space of scientific datasets. In this paper, we present a new method, named particle flow, that can capture the inherently complex dynamics of scientific datasets. We can effectively reconstruct any intermediate frames by interpolating the starting and ending frames using the resulting particle flow. We have also demonstrated that our approach can be effectively applied in data reduction for scientific datasets. Extensive experiments are conducted to show the accuracy and the efficiency of our approach over existing methods.