In order to simulate and visualize natural phenomena, especially fluid behavior such as smoke and fire, many novel studies have recently been conducted. Usually these methods use CFD (computational fluid dynamics), which calculate Navier-Stokes equations in real-time to generate realistic fluid motion and interactions, as well as high-performance GPU technologies. We proposed a new approach to the visual simulation of fluid flow by combining the use of pre-calculated CFD data with the real-time processing of such data. As the domain-specialized CFD solver predicts detailed fluid dynamics to an accuracy of a guaranteed error range, we could provide nearly actual behaviors of a fire-driven fluid flow. Moreover, this CFD data includes physical quantities such as temperature distribution, which can provide useful information to the training evaluation process. However, the data-driven method requires appropriate data processing techniques to create and manage large data sets. In this study, we developed a firefighter training simulator to demonstrate our proposed methods and explore related research issues.