Physics-guided machine learning surrogate for layout optimization to reduce on-state resistance in power MOSFETs

Continuous scaling of advanced MOSFETs has minimized intrinsic silicon resistance into the sub-milliohm levels, making layout-dependent metal interconnect resistance a significant contributor to total on-state resistance. Accurate evaluation of these parasitic effects typically relies on computationally expensive 3D field solvers, limiting their use in rapid design exploration. As a result, exploring the layout design space and identifying optimal geometric configurations that minimize total metal resistance becomes challenging and time-consuming task. Nevertheless, such optimization is essential, as improved design reduces conduction losses and enables higher efficiency of the devices. Herein, a physics-guided machine learning (ML) framework is proposed for the rapid prediction of layout-dependent metal resistance. Instead of directly predicting resistance, the model estimates the effective length-to-width ( L / W ) ratio of multi-layer interconnect networks. This approach allows the model to be technology-independent and therefore metal resistance to be reconstructed using known sheet resistance values. The ML model (XGBoost) is trained on physics-based simulation data generated using the Synopsys R3D field solver, ensuring consistency with underlying current transport behavior. The model achieves high accuracy, with average R 2 > 0.9 and low prediction errors (RMSE ∼10 −3 ). The proposed framework reduces evaluation time from hours to seconds, enabling fast design-space exploration, efficient estimation of metal resistance, early-stage design screening, and efficient layout optimization.