A machine learning pipeline for silicon wafer warpage prediction in diamond coated wire slicing

High-tech companies that manufacture silicon wafers are nowadays continuously engaged in the effort to produce wafers with increasing quality levels, to fulfill the ever-tightening requests from customers, typically manufacturers of microchips and electronic devices. One of the main measures of the quality of wafers is their shape, specifically their flatness (intended as the characteristic of being as similar as possible to an ideal plane), as flatter wafers guarantee more effective lithographic processing. In this paper, we consider one of the main silicon wafer production steps, that is, the wire sawing of silicon ingots into wafers, which affects wafer warpage. Focusing on Diamond Coated Wire (DCW) slicing, we develop a Machine Learning (ML) pipeline to predict warpage based on several machine parameters’ measurements. We test different ML techniques, namely Linear Regression, Random Forest, Light Gradient Boosting Machine, Multi-Layer Perceptron, and the recently released Tabular Prior-data Fitted Network (TabPFN) on an in-house collected dataset consisting of 1098 5-h cuts, for a total of 2.9M records. Across the tested ML models, TabPFN provides better predictions of the average warpage, even with limited pre-processing, while being less effective at predicting the warpage standard deviation. Overall, this study represents a first attempt at using ML to predict wafer warpage based on DCW slicing machine parameters, showing the potential to enable, in future work, closed loops where warpage reduction drives the machine settings.