The machining-induced subsurface damage (SSD) on fused silica optics would incur damage when irradiated by intense lasers, which severely restricts the service life of fused silica optics. The high absorption of fused silica to 10.6 μm makes it possible to utilize pulsed CO2 laser to remove and characterize SSD by layer-by-layer ablation, which improves its laser-induced damage threshold. However, thermal stress during the laser ablation process may have an impact on SSD, leading to extension. Still, the law of SSD morphology evolution mechanism has not been clearly revealed. In this work, a multi-physics simulated model considering light field modulation is established to reveal the evolution law of radial SSD during the laser layer-by-layer ablation process. Based on the simulation of different characteristic structural parameters, two evolution mechanisms of radial SSD are revealed, and the influence of characteristic structural parameters on SSD is also elaborated. By prefabricating the SSD by femtosecond laser, the measurements of SSD during CO2 laser layer-by-layer ablation experiments are consistent with the simulated results, and three stages of SSD depth variation under two evolution processes are further proposed. The findings of this study provide theoretical guidance for effectively characterizing SSD based on laser layer-by-layer ablation strategies on fused silica optics.

Experimental surface shape data

Intelligent manufacturing of ultra-precision optical surfaces is urgent but rather difficult to be break through due to the complex physical interactions. The development of data-oriented neural network provides a new way, but the existing networks cannot be adapted for optical fabrication with high feature dimensions & small specific dataset. In this letter, for the first time, a novel Fourier convolution-parallel neural network (FCPNN) framework with library matching was proposed to realize multi-tool processing decision, including tool size & material, slurry type and removal rate. The feature dimension requirement is reduced by 3-5 orders of magnitude to achieve supervised learning with hundred-level dataset.