Manufacturing-Oriented Dual Neural Surrogate for Trench MOSFET Fabrication Process and Device Optimization

This paper presents a dual-recurrent neural network (RNN) surrogate model for end-to-end semiconductor manufacturing process optimization targeting 150 nm gate-length N-type trench MOSFETs. The forward RNN accurately predicts device electrical characteristics with 98% accuracy, while the inverse RNN infers critical process parameters with 95% precision, effectively avoiding the local-optima issues of conventional multi-objective optimization. Validated against simulated and experimental data of a 150 nm trench MOSFET, our approach establishes a robust inverse-design methodology for global optimization of multi-stage fabrication parameters. This data-driven framework enables precise process control and improved device yield, highlighting its significance for advanced manufacturing.