Defect Metrology

Detecting and classifying DRAM contact defects in real time

The ability to identify yield-limiting factors rapidly and accurately in order to determine and eliminate their causes is critical to optimizing advanced IC device processes. Because of the complexity and wide variation of defects at the pilot run stage, the inspection strategy used at that point should provide classification data as quickly as possible and facilitate the ranking of detected defects according to their impact on yield. Troubleshooting using short-run experiments or split runs can then be prioritized based on the potential for yield improvement, thereby minimizing the time required to optimize the process step. Compared with manual defect review, real-time automated defect classification may also increase the wafer-inspection sampling rate, reduce the amount of work in process at risk, and reduce cycle times.

Real-Time Automated Defect Classification

To assess its capabilities, an automated wafer inspection system was implemented on the 300-mm line at PowerChip Semiconductor (Hsinchu, Taiwan) to perform a bit-line contact defect check. As DRAM contact critical dimensions (CDs) shrink to 150 nm or less, any overetch of bit-line contacts may cause a pattern collapse that results in an open via, which is a potential killer defect. The laser-based multiperspective system from Applied Materials (Santa Clara, CA) has a powerful real-time classification feature termed On-The-Fly Grouping, which identifies and segregates defects into categories during an inspection run with no loss in throughput. Initial studies were used to establish and optimize the inspection recipe. Based on those results, two versions of the system were evaluated: one with a 5X objective and one with a 10X objective. For purpose of comparison, test wafers were also evaluated using a single-perspective UV bright-field (BF) inspection system with a 0.25-µm-pixel size.

Figure 1: The multiperspective inspection tool with dark-field, bright-field, and gray-field detectors that collect light scattered within three angular regions.

The multiperspective inspection tool incorporates six collection optics with different angular fields of view. As shown in Figure 1, dark-field (DF) detectors collect the light scattered from the wafer plane into the lower elevation angles (3°–35°). A BF detector collects the light reflected at normal incidence (90°), while a gray-field (GF) detector collects light scattered into the higher elevation angles (70°–87°). The GF detector is doughnut shaped and collects light over a full 360° periphery. The multiperspective inspection system offers variable scan objectives, each with a different pixel size to provide for sensitivity and throughput optimization. A variable angle-selection control determines the collection areas of the DF detectors, and a polarization control can be used to reduce reflected noise from rough areas. The multiperspective data collected include details on more than 40 attributes by which defect types can be characterized and differentiated.

Figure 2: Total defect counts on wafers inspected using the single-perspective BF 0.25-µm-pixel system and multiperspective tools with 5X and 10X objectives.

Comparative Inspection Results

Study results revealed that the multiperspective inspection system with a 10X objective had the best sensitivity and, in fact, could detect defects that were not found by the UV BF tool. This finding is surprising because BF inspection is commonly believed to be the most sensitive method for detecting small defects. Figure 2 compares the total defect counts on five wafers inspected by the single-perspective BF system and the multiperspective tools with 5X and 10X objectives. Figure 3 shows typical wafer defect maps based on data from the same tools.

Figure 3: Inspection wafer maps for (a) the BF system, (b) the multiperspective tool with 5X objective, and (c) the multiperspective tool with 10X objective. The strong defect signature shown on the 10X map is that of contact residue defects.

Following the automated inspections, further defect review was performed using a scanning electron microscope (SEM). This high-resolution analysis aided in verifying and refining the automated grouping classifications. Two hundred random defects were reviewed, and the BF tool results and those from the multiperspective system with the 5X objective were plotted for three defect types, as presented in Figures 4 and 5, respectively. As the figures show, the results from the two tools were comparable. Even the blind contact defect counts, which fluctuated significantly over time, exhibited a similar trend with both inspection methods.

Figure 4: BF inspection system results for three defect types, based on a review of 200 randomly selected defects and normalization with total defect counts.

Defect Classification

A variety of defect types were detected by the multiperspective inspection tools, including blind contact defects (single and clustered), surface particles, and embedded particles. Identifying and eliminating blind contact defects is particularly important to process optimization because they can cause open contacts between top and bottom metal layers. Contact residues, which are also possible killer defects, were detected only by the multiperspective inspection tool with the 10X objective. The strong defect signature seen on the 10X map in Figure 3 is of the contact residue defect type. Figure 6 shows SEM images of six defect types.

Figure 5: Results from the multiperspective tool with a 5X objective for three defect types, based on real-time automated defect classification of all defects.

Classification accuracy and purity were determined using the inspection recipe from the multiperspective 10X tool. In this context, accuracy is the percentage of the defects caught and classified automatically, and purity is the exactness of the classification. For reliable monitoring and analysis, it is critical to have high purity, since this means that the defects that are classified as a given type are indeed that type of defect. Both accuracy and purity were 80–100% for the defect types encountered.

The results from one set of inspections are shown in Table I and represent typical data. In this example, the automated classification feature detected 20 single blind contacts, 9 cluster blind contact defects, 2 particles, and 371 contact residue defects. In the case of the single blind contact defects, subsequent manual review revealed that 16 of the 20 were indeed single blind contacts (row 1 in column 1), 1 was a cluster blind contact defect (row 2 in column 1), 2 were particles (row 3 in column 1), and 1 was contact residue (row 4 in column 1). Manual review also showed that one single blind contact had been misidentified and classified as a particle (row 1 in column 3). Thus, the classification accuracy for this defect type was 94.1% (16/17), and the classification purity was 80% (16/20).

As the table shows, both purity and accuracy results were excellent for all four defect types and provide high confidence in the automated classifications. The average accuracy was 98.1% ([16 + 8 + 23 + 371]/[17 + 9 + 28 + 372]) and the average purity was 98.6% ([16 + 8 + 23 +371]/[20 + 9 + 24 + 371]). These high percentages make it possible to eliminate the manual review that is typically performed during production inspections.

Defect detection and classification are only the first steps for achieving yield improvement, however. They must be followed by defect root-cause analysis and corrective action. In this study, the blind contact defects that had been classified automatically were determined to be caused by the CD exceeding the margin tolerance. The corrective action taken improved CD margin control during lithography and contact etching. The contact residue defects were traced to etch residue caused by an abnormal chamber environment during contact etching. A modification of the etch process wet clean recipe eliminated this residue.

Sampling Rate

To assess the productivity throughput of the multiperspective inspection system, its sampling rate was compared with that of the conventional production mode: BF inspection followed by manual review and classification. In the case of blind contact defects, the multiperspective tool with the 5X objective had a throughput of nine wafers per hour with a sensitivity exceeding that of the UV BF 0.25-µm-pixel inspection system. For contact residue defects, the tool with the 10X objective not only demonstrated the capability to double the sampling rate achieved with the UV BF system, but also detected defects that had not been seen by the BF tool. As depicted in Figure 7, implementing automated classification for bit-line contacts would allow the sampling rate to more than double.

Figure 7: Typical inspection throughput improvement achievable with automated classification using the multiperspective tool.

Conclusion

Inspection of DRAM contact layers using a multiperspective system with real-time classification can provide specific defect information immediately after wafer inspection without the need for follow-up manual review.

This study showed that defects were classified with high accuracy and purity (>80%), leading to rapid root-cause analysis and the subsequent elimination of killer defects. The tool's wafer-inspection sampling rate was found to be significantly higher than that of conventional inspection systems; in one instance, wafer hold time was cut from 1 hour 25 minutes to only 25 minutes. Such fast wafer disposition reduces cycle times and accelerates the yield-enhancement learning curve. By extending the inspection methodology to DRAM gate and word-line contact layers, PowerChip Semiconductor expects to reduce the overall inspection and response time from 12 hours to 30 minutes, thereby improving productivity greatly.

Luke Lin, PhD, is a deputy manager in the yield-enhancement section of the technology integration department at PowerChip Semiconductor (Hsinchu, Taiwan). He works in the areas of surface analysis, 200- and 300-mm defect metrology, and particle/yield improvement in the 300-mm DRAM production line. Before joining the company, he was active in the field of semiconductor CVD processing at Taisil Electronic Material in Taiwan. He received a BA from Tamkang University in Taipei and a PhD in physical chemistry from the University of Tsing-Hua in Hsinchu. (Lin can be reached at +886 3 5795000, ext. 2229, or lhlin@psc.com.tw.)

Maggie Cheng is a defect analysis engineer at PowerChip Semiconductor. She works in the areas of inspection, yield improvement, and defect analysis methods. Before joining the company, she was an application engineer working in several different areas of semiconductor fabrication at KLA-Tencor in Hsinchu. She received a BA in computer science from the department of information engineering at I-shou University in Kaohsiung County, Taiwan. (Cheng can be reached at +886 3 5795000, ext. 6536, or maggie@psc.com.tw.)

Timothy Han is a technologist at Applied Materials in Hsinchu, Taiwan. He works in the areas of inspection metrology and alternative yield-enhancement methods. Before joining the company, he worked in several areas of semiconductor fabrication and equipment business at PowerChip and KLA-Tencor. He received a BA from Tamkang University in Taipei and a master's degree in materials science and engineering from the National Ching-Hwa University. (Han can be reached at timothy_han@amat.com.)