Improvement of Optical Critical Dimension Metrology Accuracy Through Data Denoising

In the AA (Active Area) dry etching process at the 28 nm node, the accuracy of the OCD measurement model is significantly constrained by the uncertainty of TEM slice data. Experimental data indicate that the range of three TEM slice depths at the Center-Middle-Edge points within the same STI measurement mark reaches 12.6%, highlighting the risk of model bias stemming from anomalous TEM data in the traditional modeling process. If such data noise is not effectively mitigated, it will directly increase the OCD prediction error for trench depth and critical dimensions (CD), jeopardizing the reliability of process monitoring. This study proposes a dualthreshold collaborative optimization strategy for the OCD modeling stage: employing Cook's distance $\left( {{D_i} > \frac{4}{n}} \right)$ to identify influential TEM data and combining standardized residuals (|r<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i</inf>|>2) to filter out outliers, thereby enhancing the model's generalization capability. Based on the 28 nm process dataset, the OCD prediction RMSE of trench depth and CD in the optimized model is reduced to 21.3%, and the sensitivity to TEM fluctuations within the same measurement mark is decreased by 34%. To facilitate the rapid implementation of this method by measurement personnel, an open-source Python analysis tool has been developed, which can automatically perform outlier diagnosis, model iterative optimization, and visualization report generation, significantly streamlining the process of constructing high-precision OCD models for AA etching process control.