Neural network based uniformity profile control of linear chemical-mechanical planarization

Abstract

In this paper, a neural network based uniformity controller is developed for the linear chemical-mechanical planarization (CMP) process. The control law utilizes the metrology measurements of the wafer uniformity profile and tunes the pressures of different air-bearing zones on Lam linear CMP polishers. A feedforward neural network is used to self-learn the CMP process model and a direct inverse control with neural network is utilized to regulate the process to the target. Simulation and experimental results are presented to illustrate the control system performance. Compared with the results by using statistical surface response methods (SRM), the proposed control system can give more accurate uniformity profiles and more flexibility.Index Terms-Chemical-mechanical planarization (CMP), neural networks, uniformity, run-to-run control.

CMP process run-to-run control to adjust margin in control limit

“…The removal-rate profile shift in one product lot work is evaluated between the before and after polishing operation with only the product lot result because the product type could change from lot to lot. For the evaluation of the shift, since the removal-rate difference is calculated for each measurement site on a product wafer, the removal rates before polishing are calculated using expression (5). The within-wafer removal-rate profile update method, as we call the distribution shift, is calculated as follows:…”

“…If the film-thickness cannot be measured for all wafers in a product lot due to the throughput of the measurement apparatus, the film-thickness of the sampled wafer may be used to cancel the lotby-lot film-thickness variation. The removal rate is calculated from the within-wafer removal-rate profile using the site coherence method, expression (5). PES is also determined for all measurement sites by inverse calculation using expression (1) using the actual data of film-thickness and removal rate with the site coherence method.…”

“…It is to be noted that the ranges between the minimum and maximum are 0.786 for QC1 and 1.201 for QC3. For these distributions, the removal rates at the measurement sites of the three types of products are calculated using expression (5). The results are shown in Table 2.…”

“…Based on the distribution, a neural network-based profile control was proposed [5]. However, thus far, no methods to maintain filmthickness after CMP to within the control limits have yet been proposed for high product-mix manufacturing.…”