Random Fourier Surrogate for simulation calibration

Abstract

In this paper we propose Random Fourier Surrogate (RFS) as a method for simulation calibration. Computer simulations are actively used in various fields to model and analyze complex real-world phenomena, enabling practitioners to explore situations inaccessible by other means. To create a simulation that accurately reflects the desired real-world scenario requires calibrating the simulation parameters, a non-trivial process that requires domain expertise and iterative trial-and-error. The model-bridge paradigm has been proposed to automatize this process, where a more computationally efficient surrogate model is used in lieu of the real simulation, and a bridge model is used to map the noninterpretable surrogate parameters to calibrated simulation parameters, which are physically interpretable. Recently an efficient solution called tangent slope-intercept (TSI) descriptors has been proposed, where tangent lines of a principal curve are used to represent the simulation in low dimensionality. However, we find that it has two key problems: it is prone to overfitting and is sensitive to initialization. We address these issues using Random Fourier Features to construct a surrogate model, thereby reducing calibration time by avoiding complex computations, while retaining high accuracy. We evaluate the effectiveness of our method with different experiments on synthetic signal simulations and physical simulations of turbulent flow dynamics.