Process data-driven machine learning for non-uniformity prediction and virtual metrology in chemical mechanical planarization

Abstract Chemical mechanical planarization is an integral part of semiconductor industry. It provides wafersurface smoothing at the nanometer scale and is highly monitored. In a typical industrial fabrication facility, millions of data points are generated annually for process control, fault detection and classification purposes. By leveraging this sensor and process data for the training of machine learning models, a foundation for virtual metrology can be established. Utilizing real-world data from a high-mix, high-volume fab, we developed robust and application-oriented machine learning models capable of predicting non-uniformity and spatially resolved material removal rates on wafer surfaces. Experimental results demonstrate the accuracy of these predictions, with a high degree of precision in estimating spatially resolved material removal rates. Furthermore, the analysis of the feature importance via Shapley Additive Explanations of the models reveals that polishing time and carrier rotation are among the most critical factors influencing MRR variability.