Physics-Informed Multi-Task Learning for Material Removal Rate Prediction in Semiconductor Chemical Mechanical Planarization

The chemical mechanical planarization (CMP) process is a complex and critical operation in the semiconductor manufacturing industry, involving a wide variety of process parameters. As a universally employed technique in semiconductor fabrication, it has received extensive research attention over the past several decades. Despite the development of various physics-based and data-driven models aimed at assessing the quality and productivity of CMP tools, particularly through the metric of material removal rate (MRR), a universally accepted prediction method has yet to emerge, primarily due to the complicated chemical/physical process. To help remedy this situation, this paper explores multi-task learning and the integration of physics-based knowledge constraints into convolutional neural networks (CNNs) to enhance the predictive accuracy of CNNs. The hybrid approach aims to leverage the strengths of both data-driven models and physics-based relationships, offering a robust framework for predicting MRRs with higher confidence. The validation results show that the auxiliary tasks in multi-task learning improve the accuracy of the primary task. Also, the physics-informed strategy can clearly increase the generalization capability of the baseline CNN model.