Performance comparison between model-based and machine learning approaches for the automated active alignment of FAC-lenses

Due to their short focal lengths, FAC lenses significantly influence the performance of high-power diode laser systems. In addition to the shape, coating and surface quality, high demands are placed on the assembly accuracy for these microoptical components. In order to optimally align and position the lenses despite varying properties (e.g. focal length), active alignment strategies are used. The automation of the active alignment process for production offers enormous potential. Compared to manual processes, the reproducibility and accuracy of the alignment is increased. For the automation of the active alignment process, a deep understanding of the system behaviour is necessary. To control a diversity of variants cost-effectively and robust, new approaches must be taken into account. Concepts of AI or machine learning are great for this kind of generalization and adoption and they have many advantages for the active alignment of systems like DOEs or free-form-optics, with a complex system behaviour. In this publication, we want to compare the performance of a classically model-based algorithm and a machine learning approach for the automated active alignment of FAC-lenses. The model-based algorithm uses a physical model of the metrology system (including the FAC to be aligned) to estimate a misalignment in 4-DOF. The machine learning algorithm consist of a deep neuronal network which was trained with image data.