Abstract
Recent advances in HPC-oriented technologies have made large-scale numerical simulations feasible on an extreme scale and enabled their use in solving increasingly complex real-world scientific problems. However, the analysis of such real-world phenomena requires not only highly accurate computations but also realistic turnaround execution time. In this context, research on surrogate models has recently increasingly attracted attention as an effective alternative to traditional numerical simulations. This is mainly because it can significantly reduce computational time by replacing numerical simulations with approximate models, such as machine learning models. However, when applied to large-scale simulations, the output data can proportionately become large, thus leading to potential difficulties in later visualization and analysis tasks. Therefore, a great reduction in the overall processing time can be expected if we can build an image-based, or visualization, surrogate model by also taking visualization processing into account, that is, to directly output rendering images rather than simulation results.