Accurate optical proximity correction using convolutional neural network with feature aggregation

Optical proximity correction (OPC) as an important means to alleviate optical imaging distortion has become an indispensable component in the lithography process for integrated circuits at 90 nm and below technology nodes. As the pattern density of circuits continues to increase, conventional rule/model-based OPCs face the disadvantages of inaccurate optimization and high computational overhead, which have become unsustainable. We introduced a deep learning approach embedded with feature aggregation for OPC. The approach incorporates several enhancements to the backbone network: (1) Enhanced convolutional block composed of FFT convolution module and origin convolution module are employed to supplant conventional convolutional blocks. (2) Wavelet pooling layers are utilized in place of standard pooling layers to reduce feature dimensionality while mitigating overfitting. (3) Inception module substitutes traditional convolutional modules within the bottleneck layer to preserve extracted feature information effectively. Experimental results demonstrate that our proposed framework achieves comparable PVB performance to Neural-ILT 2.0 and DevelSet-ILT methods while reducing pattern errors by 3.1% and 5.5% and edge placement errors by 5.6% and 12.7%, respectively.