Modeling the laser direct writing grayscale exposure process and predicting photoresist 3D morphology based on UNet/CGAN networks

The laser direct writing (LDW) grayscale exposure process is crucial for precision manufacturing of large-aperture diffractive optical elements (LA-DOEs), yet it faces challenges like photoresist nonlinear response, multi-parameter coupling, and time-varying development kinetics. Traditional orthogonal experiments and physical modeling are limited by high cost and oversimplified assumptions. This study proposed an experimental data-driven framework: standardized processes (AZ4620 photoresist spin-coating, parametric LDW exposure, laser confocal morphology measurement) generated 91 valid samples, expanded via random cropping into 91,000 set data pairs. UNet and CGAN were validated for photoresist 3D morphology prediction. After 10 epochs, UNet achieved a validation set root mean square (RMS) loss of 0.0035, with a ∼1% mean absolute error between predicted and real morphologies. CGAN exhibited comparable low-frequency accuracy and additional roughness-like high-frequency information, but required 50 epochs and suffered training instability. This work establishes an intelligent mapping among process parameters, grayscale masks, and 3D morphologies, enhancing LDW process study efficiency and providing a paradigm for LA-DOE fabrication.