In the fab environment, wafer value is at a premium and process innovation must be achieved with minimal wafer risk and resources. When working to improve or change a process outcome (e.g. reduce roughness or achieve a target etch depth), the experimentation process typically requires approval for a pre-determined number of wafers and time. When placed under such tight constraints, what is the most efficient approach? How can a process engineer take advantage of not only their subject matter expertise but also their historical data? Enter Bayesian optimization (a.k.a: Sequential Learning or Active Learning)
Bayesian optimization allows for an iterative and intelligent approach to identifying the best possible factor combinations to achieve a desired outcome (e.g. maximize yield, reduce defects, etc.). In this paper, we will use a semiconductor process engineering example where we analyze historic data to iteratively improve factor settings to achieve a new or improved outcome. We will show that whether there is minimal or months of historic data, Bayesian optimization will provide a series of parameter values to test, and with each result, improve the desired outcome. When wafers are at a premium and process change needs to be achieved accurately and with minimal wafer waste, Bayesian optimization can vastly reduce time and waste.
