A Geometric Surrogate for Simulation Calibration

Computer simulations are amply utilized in scientific, industrial, and business environments for analyzing complex real-world phenomena. For a simulation to match realistic scenarios, it is often necessary to conduct time-consuming parameter calibration, i.e., finding the optimal simulation parameters given a set of observations. In this work, we employ a machine learning-based approach to perform calibration faster and more accurately, with two components: a surrogate model of the simulation that is easy to obtain but not physically interpretable and a bridge model that maps the surrogate to the calibrated parameters. This bridge is learned using a dataset of simulations with known parameters. We propose a geometry-based surrogate, tangent slope-intercept (TSI) descriptors, that can effectively represent the simulation with low dimensionality. This property allows the bridge model to avoid the curse of dimensionality, learning without needing large datasets and avoiding overfitting. We showcase the effectiveness of TSI descriptors through experiments with synthetic signals and physical simulations of turbulent flow dynamics.