OCD -Machine learning based metrology for copper pad surface for hybrid bonding integration

In the context of 3D hybrid bonding architectures, ensuring the flatness of metallic pads before bonding is critical. The control of this flatness is important and the dimension to be measured is very small (~nm) compared to the dimensions of the pattern (~µm). Traditional control methods, such as Atomic Force Microscopy (AFM), offer high precision but lack the throughput required for high-volume manufacturing. Spectroscopic reflectometry (SR) presents an attractive metrological method due to its speed and larger measurement coverage. This study investigates the sensitivity of SR to dishing of copper pad across wafers with different thicknesses. We evaluate the ability of using a Deep Neural Network (DNN) model trained on raw spectra to predict copper dishing, using AFM as the reference measurement. Additionally, we assess whether the model can be extended to new batches of wafers and capture process variations. This study examines the capabilities of the deep learning based-model robustness predict the dishing of copper pads across varying experimental conditions.