Transfer learning based on 1D-CNN for critical dimension Predication of HAR grating structures

Abstract

This study investigates the prediction of four critical dimension (CD) parameters—top space, bottom space, depth, and pitch—of high aspect ratio (HAR) structures using simulated deep ultraviolet (DUV) reflectance spectra based on transfer learning without collecting all necessary data. A large dataset was generated through COMSOL Multiphysics simulations and used to train a one-dimensional convolutional neural network (1D-CNN). Under a transfer-learning scheme in which the network was pre-trained on 5,000 ideal (smooth-sidewall) grating spectra and then fine-tuned with 1,200 scalloped (non-ideal) grating spectra, a deep learning model trained only with θ_i = 35° spectra achieved significant improvements with R² values of 0.982 (top space), 0.9556 (bottom space), 0.9877 (depth), and 0.9745 (pitch), respectively. The corresponding mean absolute errors (MAE) were 0.0053, 0.0082, 0.0223, and 0.0268, while the mean absolute percentage errors (MAPE) were 0.89%, 1.36%, 0.74%, and 1.07%. These results validate the effectiveness of the CNN-based approach for rapidly and precisely characterizing the dimensional properties of HAR structures. Importantly, these results confirm the value of transfer learning: fine-tuning significantly improves prediction performance for CD estimation in HAR grating structures while reducing the required number of non-ideal (scalloped) spectra for fine-tuning to 1,200 in this study. Additionally, uncertainties arising from the intended measurement configuration and practical implementation conditions can be systematically identified and characterized using data-driven approaches. Consequently, simulation-generated data can provide a distinctive and robust framework for advanced process monitoring and can be readily integrated with measurement data in future deployment. In summary, the proposed method requires significantly less training data than the three existing comparative approaches. This strategy greatly reduces the burden of data collection and labeling, enhancing modeling efficiency. Furthermore, to assess feasibility under fabrication-induced profile non-idealities, the forward surrogate spectral prediction model is trained on ideal structures and subsequently adapted to simulated Bosch-inspired scalloped sidewalls via transfer learning, thereby reducing the need for extensive non-ideal training data and lowering the data-collection burden.