Cherry-picking: automated model term selection for precise overlay and CD process control

In the lithography process, overlay and critical dimension (CD) process control requires modeling of measured signatures. Polynomial models such as Zernike or Legendre are commonly used, but selecting the correct terms is crucial to avoid overfitting and to ensure optimal performance. Traditional methods often include non-relevant terms, which can reduce control loop stability. This paper introduces an automated solution for identifying individual significant model terms from polynomial models based on metrology measurements utilizing lasso regression and metrology-specific cross-validation. This approach enhances modeling flexibility and accuracy, outperforming traditional full-order models and other feature selection algorithms in simulated run-to-run process control. The proposed solution is versatile and can be readily integrated into existing lithography workflows in semiconductor fabs. Additionally, our solution leads to stable process control and provides a simplified, more interpretable model that is more effective for general monitoring in other processes such as etch, deposition or polish.