A grid mapping-based U-net algorithm for photolithography overlay virtual metrology

Photolithography is a critical process in semiconductor manufacturing that forms IC patterns on wafers. As ICs grow more complex, inter-layer misalignment—known as overlay error—has emerged as a major cause of yield degradation. While overlay metrology is essential for ensuring alignment accuracy, its high costs and impact on production efficiency have driven increasing demand for VM models. This paper proposes a new high-accuracy overlay error VM algorithm with three components. First, we integrate multi-modal metrology data, including alignment, leveling, lens, and APC, into a 3D tensor through grid-based representation. Second, we introduce TwinUNets, an architecture specifically designed for effective multi-modal data learning, consisting of two U-Nets and a feature fusion layer. Third, we introduce a continual learning based model update algorithm that efficiently retrains using minimal data while adapting to temporal changes in sensor data and process parameters. Experimental results demonstrate that the proposed algorithm achieves better overlay prediction accuracy compared to alternative ones while maintaining stable performance over long periods without degradation.