Defect Learning with Predictive Sampling for Process Improvement

As technology nodes advance, the need for higher sensitivity optical inspection to identify critical defects has become extremely important for technology development. However, more sensitive optical inspection can induce more nuisance and hence more SEM non-visual (SNV) defects during review sampling. High SNV in the defect Pareto hinders the ability to get a true picture of the actual distribution of defect types on a wafer, and defect-of-interest (DOI) types that are crucial for process diagnostics can be missed. The culprit of this problem is the method of review sampling. Traditional review sampling consists of two parts: binning and defect selection. Binning is defined as a set of rules and conditions determined by human experience and judgment to categorize different DOI types. Then, defects are selected from each bin and reviewed by SEM. Due to the nature of high SNV from optical inspection, the random selection of defects will end up with high SNV in the defect Pareto. A defect Pareto with high SNV provides little value to yield learning. Because SEM review plus classification is limited by time and economic budget, improving the ability to predict whether a defect is DOI or SNV before SEM review is valuable. This paper introduces a machine learning based method suitable for high volume manufacturing that can increase the probability of finding DOIs during review sampling by integrating all available data sources, such as historical defect attributes from optical inspection, context information of the inspection recipe, design hotspots and metrology measurements. In addition to review sampling, this paper also illustrates other applications based on machine learning defect prediction, such as virtual process window discovery, and predicted defect types for trend monitoring. A predictive analytics platform was employed to allow defect type prediction based upon multiple inputs.