Predicting the Critical Dimensions of Micron and Sub-micron Structures Using Joint Training Models and Electromagnetic Simulation Tools

This paper presents an enhanced non-destructive optical measurement method crucial for 3D IC technology. Utilizing a deep neural network and electromagnetic simulation tool, an inverse model is established to predict the critical dimensions (CD) of micron and sub-micron structures. The forward model accelerates data generation, overcoming limitations of the finite-difference time-domain (FDTD) method. Compared to FDTD simulations of reflectance spectra which take about 50 min, the model's prediction time for reflectance spectra is only 0.01 s. The three key dimensions with better prediction results from the inverse model are top radius (R_top), height, and bottom radius (R_bot). In the proposed joint training model, by simultaneously training forward and inverse models and adjusting the weights of forward and inverse loss functions, the mean absolute percentage error (MAPE) for R_top, height, and R_bot is reduced from 0.26 to 0.08%, 0.69 to 0.44%, and 4.01 to 0.6% respectively.