Bayesian Optimization: A Goal-Oriented Approach for Process Development (2025-US-30MP-2554)

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 predetermined number of wafers and amount of 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 for identifying the best possible factor combinations to achieve a desired outcome (e.g., maximize yield, reduce defects, etc.). In this paper, we use a semiconductor process engineering example where we analyze historic data to iteratively improve factor settings to achieve a new or improved outcome. We 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.