|
Narrowed CD specifications coupled with very tight cycle time requirements have resulted in search for improvement opportunities in CD stability and tuning options for mask fabrication unit processes, including pattern generation, resist development and etch, which may yield narrower scattering band of CD off-target (CDO) of final products. Targeting models are already in productive use at AMTC, accounting for different mask and blank types, clear field, resist type, pattern type and many other parameters. This targeting model is static however, and changes in the CD performance of contributing factors must be adjusted manually when CD drift inevitably occurs. In the past, several approaches to introduce time-based corrections were pursued. Correction of step function of the resulting CDO caused by e.g. resist lot change is the easier task, due to the fact that such factors can be closely analyzed prior to productive use by test, and offset accounting for the individual factor can be introduced. More troubles cause factors, whose effects on CDO is smooth and can be observed as long-term drift in the CDO. The CD drift is frequently of very different origin and effects of several factors are overlapping in time. By measuring the final CD on the products, we can see only the ‘envelope’ of all the effects. To target such factors, we need to identify their root cause and ideally an easy-to-monitor indicator. In this paper we show an analysis approach to identify the most significant and vital indicators to process bias. Analysis of production data covering several manufacturing steps including metrology over more than three years was performed. Using machine learning methods, a “big data” set is reduced, and the most appropriate model is selected using statistical methods. Criteria for selection of factors were significance level in analysis of variance and the distribution of residuals was used for model comparison. Based on these factors a model of the etch contribution to the CD was established, describing the variation of the etch process for a virtual mask with constant clear field, resist sensitivity and absorber composition and thickness. This model is based on the process data collected at the etch process during processing of each mask processed with the same recipe. Monitoring this time trend of the “modelled etch bias” gives very fast feedback about the stability of the etch process and evolution of the etch contribution to CD. This data is used to trigger appropriate corrective actions to further stabilize the manufacturing process.
|
