Machine Learning-Driven Optimization of Burst Femtosecond Laser Processing for High-Performance Anti-Reflective Windows

Femtosecond laser processing of large-area micro- and nanostructures exhibits significant potential for applications in materials science, optical engineering, and biomedicine. However, existing methods for fabricating high-performance micro- and nanostructures heavily rely on empirical trial-and-error approaches, leading to cumbersome processes, high resource consumption, and low efficiency. To achieve the efficient manufacturing of anti-reflective microstructures with nearly perfect performance, we propose a strategy that utilizes machine learning (ML) to assist femtosecond lasers in real-time prediction and process optimization. A multilayer perceptron model was trained on simulation data derived from the finite-difference time-domain method, establishing a nonlinear mapping between microstructural morphology parameters and transmittance. By deploying the trained model in the fabrication system, transmittance spectra can be predicted within 0.004 s upon input of structural parameters, significantly enhancing process optimization efficiency. Ultimately, using ML-optimized processing parameters combined with a burst pulse and bow-tie scanning technique, large-area anti-reflective microhole arrays (12 × 12 mm²) with a periodicity of 2 μm were fabricated on the surface of magnesium fluoride (MgF₂) windows at a rate of 10,000 holes per second. The anti-reflective MgF₂ window achieved an average transmittance of 99.03% in the 3 to 5 μm range, maintaining stable transmittance across a broad angle range (0-50°) and demonstrating excellent infrared image capturing capabilities. This study facilitates the practical deployment of anti-reflective windows in extreme-environment imaging applications.