Learning the Inverse Solution of Laser Drilling Model

Laser drilling cooling holes into turbine components is well established in industries. However, there are drawbacks related to setting and maintaining the appropriate manufacturing conditions, which can be formulated as an inverse problem. The inverse solution of a physical model describing long-pulse laser drilling of sheet metal is learned through an artificial algorithm. The trained feed- forward neural network predicts process parameters of desirable production outcomes. Here, the trained network predicts beam radius and pulse power to drill through getting a hole with specified conicity. The hyperparameters of the neural network are trained using algorithmic differentiation. Using a physical model improves the solvability of the inverse problem, since all trials during training belong to the applicable range of laser drilling and improves the training procedure going in a physically informed direction.