Data-driven optimization of maskless grayscale laser lithography

A data-driven optimization algorithm is described for maskless grayscale laser lithography (MGLL), providing effective correction for topography errors deriving from the non-linear response of the photoresist to exposure dose, proximity effects and sharp dose transitions near edges and vertices. The algorithm employs an artificial neural network (ANN) to represent the physical process, with the MGLL virtual photomask and its radial averages as input variables and the predicted surface topography as the output. Virtual photomask optimization is achieved by predicting the surface topography obtained with a given virtual photomask and implementing corrections based on the predicted error. An elaborate training model with a range of different surface features is employed to acquire a comprehensive dataset comprising approximately one million rows of data for training the ANN. Hyperparameter optimization with an independent test set sees greatest accuracy and ability to generalize achieved with a simple ANN comprising a Sigmoid activation function and a single hidden layer with 15 neurons. Optimized virtual photomasks are calculated for the training model and an ultrathin free-form micro-optical element (FFMO), with the average Euclidean distance between the experimental and target surfaces reduced from 3.6 mu m with an experimental contrast curve to 1.0-1.2 mu m with the optimization algorithm in a single build. The developed approach provides an effective, versatile and adaptable pathway towards first time right MGLL.