Full-chip application of machine learning SRAFs on DRAM case using auto pattern selection

As technology continues to scale aggressively, Sub-Resolution Assist Features (SRAF) are becoming an increasingly key resolution enhancement technique (RET) to maximize the process window enhancement. For the past few technology generations, lithographers have chosen to use a rules-based (RB-SRAF) or a model-based (MB-SRAF) approach to place assist features on the design. The inverse lithography solution, which provides the maximum process window entitlement, has always been out of reach for full-chip applications due to its very high computational cost. ASML has developed and demonstrated a deep learning SRAF placement methodology, Newron™ SRAF, which can provide the performance benefit of an inverse lithography solution while meeting the cycle time requirements for full-chip applications [1]. One of the biggest challenges for a deep learning approach is pattern selection for neural network training. To ensure pattern coverage for maximum accuracy while maintaining turn-around time (TAT,) a deep-learning-based Auto Pattern Selection (APS) tool is evaluated. APS works in conjunction with Newron SRAF to provide the optimal lithography solution. In this paper, Newron SRAF is used on a DRAM layer. A Deep Convolutional Neural Network (DCNN) is trained using the target images and Continuous Transmission Mask (CTM) images. CTM images are gray tone images that are fully optimized by the Tachyon inverse mask optimization engine. Representative patterns selected by APS are used to train the neural network. The trained neural network generates SRAFs on the full-chip and then Tachyon OPC+ is performed to correct main and SRAF simultaneously. The neural network trained by APS patterns is compared with those trained by patterns from manual selection and multiple random selections to demonstrate its robustness on pattern coverage. Tachyon Hierarchical OPC+ (HScan+) is used to apply Newron SRAF at full-chip level in order to keep consistency and increase speed. Full-chip simulation results from Newron SRAF are compared with the baseline OPC flow using RBSRAF and MB-SRAF. The Newron SRAF flow shows significant improvements in NILS and PV band over the baseline flows. This whole flow including APS, Newron SRAF and full-chip HScan+ OPC enables the inverse mask optimization on full-chip level to achieve superior mask performance with production-affordable TAT.