Abstract
As an interposer, Through-Glass Vias (TGVs) play a critical role in advanced packaging such as Co-packaged optics (CPO). Currently, due to the complex influence of laser wet-etching process parameters, the precise bidirectional prediction of TGV parameters and the etching morphology still remains a challenge. In this paper, a bidirectional design method for TGVs is proposed, which is based on the cross-modal learning method. By integrating a Cellular Automaton Etch-Diffusion (CAED) physical model with a Stable Diffusion (SD) architecture, accurate forward prediction from laser parameters to TGV morphology is realized successfully. In addition, the Contrastive Language-Image Pre-training (CLIP) model is also applied to achieve an efficient inverse design of TGVs. Furthermore, the generalization ability is examined in this paper, demonstrating significant robustness and stability of the generative model. The results provide an efficient method for enhancing TGV quality within a deep learning framework.