This thesis focuses on interactively visualizing, and ultimately simulating, cumulus clouds both in virtual reality (VR) and with a standard desktop computer. The cumulus clouds in question are found in data sets generated by Large-Eddy Simulations (LES), which are used to simulate a small section of the atmosphere over a period of several hours. These data sets are large, 3D, multi-variate, and time-varying, which pose several difficult visualization challenges. In order to overcome these challenges and gain insight into such complex data, several techniques are developed and employed together. At a high level, the research presented in this thesis can be divided into four categories: analyzing and visualizing interesting features in the data, giving the user sufficient control over the visualization, keeping the visualization interactive, and interactively simulating the cumulus clouds. The first step to understanding the data was to find and track the features, i.e. the clouds, in the data. Through the use of a connected component labeling algorithm, individual clouds in the data could be identified. These clouds were then visualized in a VR visualization environment, which allowed atmospheric scientists to visually identify interesting clouds for further study. The next step along the way was to improve the visualization experience. One aspect of this was to develop data “reprocessing”. This allowed the atmospheric scientists to generate derived data from the raw simulation data for inclusion in the visualization environment. Another aspect of this was to improve the user interface with a more intuitive interaction technique and through the use of familiar 2D widgets. The next challenge to address was the data access bottleneck. In order to move more data from disk to main memory and from main memory to the GPU, a lossy vector compression method based on quantization is developed. By analyzing and bounding the angular error introduced by quantization, unit vectors can be quantized using 16 bits with less than 0.4 degrees of angular error. This can reduce the size of mesh geometry, and, when also quantizing vector length, can be used to compress vector fields. In order to help understand the relationship between the clouds and the air around them, interactive particle tracing was the next research area. Using the power of the GPU, millions of particles could be advected interactively. These particles could be visualized as particles in different visual styles or as flow curves such as stream-lines or streak-lines. By combining vector-field compression with a multi-resolution advection scheme, users could interactively seed and advect particles around selected clouds of interested through time. Most of the techniques developed in this thesis have been integrated into Cloud Explorer, which is an experimental VR visualization platform for interactively visualizing and studying the cumulus cloud data. Several techniques have also been integrated into stand-alone applications. The final research area addressed in this thesis was interactive simulation of the cumulus clouds. GALES, a GPU-based, atmospheric, Large-Eddy Simulation, was the result of this effort. GALES runs 16x faster than the Dutch Atmospheric Large-Eddy Simulation, which is a Fortran-based LES, while maintaining comparable numerical accuracy. This speedup enables GALES to interactively run and visualize simulations that fit into GPU memory. This opens new and exciting possibilities for future computational steering research.