Physics-constrained sequential generative model for process recipe inverse design and in-line tuning of Trench MOSFET

This paper proposes a context-aware, Long Short-Term Memory (LSTM) based sequential generative surrogate for Trench MOSFET inverse design and in-line tuning. Unlike conventional proxy models treating the highly coupled fabrication flow as a static black box, the proposed framework explicitly integrates semiconductor manufacturing's sequential causality. The framework employs a “predict-and-fix” sequence generation strategy that effectively mitigates error accumulation, and it accurately synthesizes feasible process recipes by incorporating a physics-constrained loss function. The framework is comprehensively validated on an industrial 12-inch Trench MOSFET manufacturing process. Results demonstrate that the model achieves rapid recipe convergence through a closed-loop feedback mechanism for both new product development and mass-production in-line tuning. Supporting the use of simulation-augmented learning in practical fabrication and strictly satisfying discrete equipment constraints, this work establishes a robust, physically-aware artificial intelligence paradigm for semiconductor virtual manufacturing.