Visualizing the prediction of laser cleaning: a dynamic preview method with a multi-scale conditional generative adversarial network.

By producing high power per unit area and obtaining immediate feedback, lasers offer the potential for high-precision surface processing. However, due to having many complex and nonlinear physiochemical processes, unfitted values of parameters can easily result in undesired cleaning results: the substrate is either damaged or not cleaned. Thus, an accurate, flexible, and automatic image preview method is highly demanded in the cleaning practice. Here, we propose a multi-scale conditional generative adversarial network (MS-CGAN) that allows workers to view what a cleaned version of the surface would look like before actually cleaning, so that it can assist workers in adjusting laser parameters in advance to obtain desired cleaning results for resource saving and efficiency improvement. The advantage of this method is that it requires zero knowledge of the mechanism of a laser, and hence avoids the need for modeling the complex photon-atom interactions that occur in laser cleaning. Extensive experiments show that MS-CGAN can not only visualize the high-fidelity prediction of the laser cleaning effect, but also produce preview images that exactly match the cleaned surfaces. The method provides an option for industry benefits, such as minimizing the risk of wasting resources, sustaining the environment, and cutting the cost of labor.