Improving optical critical dimension metrology via systematic screening of TEM reference data

Le résumé fourni par la source

BackgroundIn integrated circuit manufacturing, transmission electron microscopy (TEM) has long served as the golden reference for optical critical dimension (OCD) measurement. However, the reliability of TEM data is frequently challenged by the sample destructive preparation, and subjective human interpretation is usually affects by the boundary selection between features in the TEM measurement. Such TEM reference uncertainty will introduce significant challenges when tuning the OCD models based on the reference data, thereby compromising the OCD measurement accuracy.AimThis work proposes a framework designed to systematically validate the TEM data, reduce the TEM measurement uncertainty, and thereby enhance OCD measurement accuracy.ApproachA comprehensive TEM-reference diagnostic framework was developed, integrating optimized TEM sampling, machine learning-based Auto-TEM measurement, statistical regression diagnostics using Cook’s distance and standardized residual analysis, and machine learning sensitivity analysis based on a leave-one-out strategy. The framework was validated using SiGe etch structures in a 28 nm planar process flow.ResultsThe optimized sampling strategy and Auto-TEM workflow reduced random and operator-induced variability in TEM reference extraction. Regression diagnostics identified abnormal TEM reference sites that strongly affected the OCD–TEM calibration. After excluding these unreliable references, the conventional model-based OCD calibration improved, with R2 increasing from 0.76 to 0.86 and the calibration slope changing from 0.76 to 0.82. The ML-based OCD workflow independently identified the same abnormal reference sites through leave-one-out sensitivity analysis. After reference screening, the ML-based correlation improved from R2=0.77 to 0.88, with the regression slope changing from 1.09 to 1.18. These results indicate that abnormal TEM reference data can introduce systematic OCD-to-TEM calibration errors on the order of 10%.ConclusionsThe results demonstrate that the integrity of TEM reference data should be explicitly evaluated rather than assumed a priori in OCD calibration workflows. The proposed framework effectively reduces both systematic and random uncertainty originating from TEM measurements, thereby improving OCD calibration robustness and model reliability. In addition, the Auto-TEM workflow further suppresses operator-induced variability and enables more stable reference extraction for advanced-node OCD metrology.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.