Prior-primed deep neural network based EUV mask inspection

Actinic patterned mask inspection (APMI) is used to verify the quality of photomasks for EUV lithography by revealing eventual defects in the patterned mask layout. The current approach to APMI, based on conventional imaging, is expensive and challenging to scale to keep up with Moore's law. Ptychography offers a promising alternative for actinic EUV mask inspection by mitigating the need for expensive optics and providing better scalability compared to direct imaging approaches. However, the adoption of this lensless imaging method in semiconductor fabs is hampered by throughput challenges, which are due to the slow, iterative phase retrieval process and to the time-intensive data collection. In this study, we explore and demonstrate a rapid APMI method by exploiting a deep neural network (DNN) architecture which makes use of the extensive prior information available for photomask samples. Our aim is to achieve high-fidelity image reconstruction and identify defects in a photomask sample by processing only a small subset (less than 5% in this case) of the measured diffraction patterns using a network trained exclusively with synthetic data. We developed our DNN using both synthetic and experimental data, and finally, we tested the DNN with a completely synthetic dataset to ensure a clean split among training and test data and to prove that this approach can be used in a real situation with no external information on the mask defect content. Although the DNN was not able to accurately detect all the defects, we used the DNN prediction as a starting point for conventional ptychography and we demonstrated a significant improvement in reconstruction speed even with respect to the case where ptychography is initiated by an educated guess based on the prior knowledge of the mask layout. We conclude the paper by showing the outcome of a die-to-database inspection of a logic-like EUV mask pattern obtained with our approach.