Resist Model Calibration by Using CNN Backpropagation
Published as 2023 IEEE TSM, 2021 SPIE Advanced Lithography
New approach for resist modeling
Conventional resist model is "function-based"➡️use too many kernels
Use "free-form" kernels instead of "function-based" ones
Structure of CNN and resist model are similar➡️optimize kernels with CNN training method
44% faster lithography simulation can be done
OPC Using Bidirectional-RNN
Published as 2021 IEEE TSM (BPA), 2019 SPIE Advanced Lithography
First paper that used RNN for lithography applications
Previous ML-OPC methods consider each segment as independent
Actual OPC is iterative process; need to consider causality between segments
Use RNN to consider causality between segments
36% lower EPE than SOTA ML-OPC method
Synthetic Pattern Generation Using GANs
IDCT image cGAN output After post-processing
Published as 2020 IEEE TSM
First paper that used cGAN for synthetic pattern generation
Pattern diversity is limited for ML-litho applications
Generate new patterns using DCT-GAN and deblurring-cGAN
DCT-GAN: generate DCT signal vectors, deblurring-cGAN: generate deblurred pattern image from IDCT image
Synthetic pattern covers 76% more design space
Dynamic IR-drop Prediction Using CNN
Published as 2022 MLCAD, 2021 ISCAS
First paper that used U-net for dynamic IR-drop prediction
Previous ML-based IR-drop predicts each gate one-by-one➡️too slow
IR-drop occurs locally, so use image-to-image network (U-net)
Effect of decaps can be also considered by using recurrent U-net
16 times faster than commercial tool, 12% prediction error
Transient Clock Power Prediction with Pre-CTS Netlist
Published as 2020 ISLPED, 2018 ISCAS
First paper that used ML for clock power prediction
Clock power consumes up to 40% of total power, but hard to estimate without CTS
Using pre-CTS netlist, predict clock tree structure (including clock gating cells) using ANN
Also, use ANN to optimize CTS parameters to minimize clock power
Clock power prediction error < 5%
ECO Leakage Power Optimization Using GCN
Published as 2020 GLSVLSI (BPC)
First paper that used GCN for leakage power optimization
Conventional leakage opt. is too slow (iterative Vth update)
Vth change of one cell affect the change of other cells in common path
Use graph convolutional network to share appropiate Vth to neighbors
Similar leakage reduction with commercial tool, but 2x faster
