Physics-Informed Bayesian Optimization Framework for Etching Rate and Surface Roughness Co-Optimization

Abstract

Precise etching control, balancing etching rate (ER) and surface roughness (R a), is critical yet challenging in advanced semiconductor manufacturing. Traditional co-optimization relies heavily on costly and time-consuming trial-and-error experiments. This study introduces a Physics-Informed Bayesian Optimization (PIBO) framework to efficiently optimize ER and Ra during silicon plasma etching. PIBO overcomes extrapolation limitations of conventional Gaussian Process Regression by integrating physical prior knowledge - specifically, qualitative relationships of source power (PW) and pressure (P) to the process metrics. It utilizes a 3D etching profile model to generate data across parameter spaces. The framework employs an iterative feedback loop: starting from limited experimental data, it recommends new parameter sets for evaluation; these results then refine the model and subsequent recommendations. This approach systematically minimizes the number of required experiments. By embedding engineering understanding of the PW and P effects and leveraging computational modeling, the PIBO framework bridges the gap between physical intuition and datadriven optimization. It significantly reduces reliance on extensive trial-and-error, enabling faster and more efficient acquisition of optimal etching parameters (PW, P) that achieve the desired ER/Ra balance, ultimately accelerating process development.