Introduction of physical knowledge in kriging-based meta-modelling approaches applied to Non-Destructive Testing simulations

In this paper, we propose a new way to introduce physical knowledge into the kriging method to improve regression accuracy and obtain real-time realistic simulations. Our technique incorporates a coarse physical model into the correlation function by spatial warping. Tested in the context of the simulation of temporal ultrasonic signals for Non-Destructive Testing (NDT) applications, the obtained results are more accurate than ordinary kriging especially for small datasets. Integrating a physical model also makes the resulting meta-model more generic: using physics, more system use cases are covered without increasing the data requirements. Eventually, applied in the case of noisy experimental data, our enhanced kriging can be tuned to split the deterministic content from the random output fluctuations. As a result, the meta-model can be established directly on real measurements to get a higher degree of realism.