Transient electrodynamic simulation of laser beam reflection on inclined surfaces using a generative adversarial network-based deep learning model

Abstract

In laser-material interactions, the temporal dynamics of beam reflection are crucial as they affect both laser absorption and material response. Currently, electrodynamic simulation is the most accurate method for calculating transient laser reflection and absorption patterns, but it is computationally very expensive. In this study, we present the first transient electrodynamic simulation model based on a generative adversarial network, with the time information embedded as color in the input image. The model accurately predicts reflection patterns of a laser beam on inclined surfaces using an input image containing geometric and time information of the domain. The finite-difference time-domain method was used to generate 16 data by changing the surface inclination angle from 0o to 75o. The model was based on a generator made by stacking a deep residual network and a discriminator. Although trained on limited data, the model predicted transient beam reflection patterns without overfitting, and the average structural similarity index measure and R-squared accuracy were 96.7 % and 98.0 %, respectively. Another deep learning model was developed to predict the laser beam absorptance from a laser beam reflection image. The ground truth absorptance values were obtained from the Fresnel equations, and the average R-squared accuracy was 99.8 %. Robustness evaluation was additionally performed to examine if the model can be used to predict reflection patterns from angled and curved surfaces. Reasonably accurate results were obtained when the angle difference was not greater than 3 degrees for angled surfaces and the normalized curvature was 0.032 or less for curved surfaces.