A Learning Acceleration of Image-based Surrogate Models Using Pixel Shuffling

Abstract

Research on surrogate models, which are expected to significantly reduce computational time by replacing large-scale numerical simulations that require a large amount of computational time with approximate models such as machine learning models, has been attracting attention. However, if the numerical data generated by the simulations is large, the data generated by the surrogate model will also be large, which may cause time consuming in visualization processing during analysis. If a surrogate model, which we call image-based surrogate model, that also takes visualization processing into consideration can be constructed for large-scale simulation results, it will be possible to directly generate visualization results without simulation data, which is expected to greatly improve efficiency in analysis through visualization. In this study, an image-based surrogate model is constructed by learning multiple visualization images of numerical data and the simulation parameters at that time. The learning model to be developed is constructed based on the adversarial generative network model, and pixel shuffling is applied to the generators that are part of the model to make feature extraction more efficient, thereby speeding up the convergence of learning loss. In our experiments, we applied this method to actual numerical simulations and succeeded in speeding up the process by a factor of approximately 2.7 while maintaining the same level of prediction accuracy as existing learning models.