A novel approach to the e-beam mask writers’ local registration error offers critical sampling density at reasonable throughput.

*By Klaus-Dieter Roeth, Hendrik Steigerwald, Runyuan Han, Oliver Ache, Frank Laske (KLA-Tencor MIE GmbH, Germany)*

**Abstract
**Mask data are presented which demonstrate local registration errors that can be correlated to the writing swathes of state-of-the-art e-beam writers and multi-pass strategies, potentially leading to systematic device registration errors versus design of close to 2nm. Furthermore, error signatures for local charging and process effects are indicated by local registration measurements resulting in systematic error, also on the order of 2nm.

**Keywords:** LMS IPRO, metrology, registration, local e-beam error

**1. INTRODUCTION**

Mask manufacturers are facing major reticle specification challenges, driving the need for new and enhanced but cost-effective solutions that provide tighter registration specifications and on-device registration qualification.

E-beam mask writers’ local registration error may have a critical impact on the error contribution of the reticles to wafer overlay as it is very local and most likely is not revealed with standard quality control schemes and sampling. The reticle error signatures are, of course, writing-strategy-dependent; but may also be caused by residual deflector alignment instabilities, thus leading to a very local but potentially critical, non-correctable device overlay error on the wafer. Since the e-beam writer strategy does not differ significantly between ArFi masks and EUV masks, we expect a similar error signature for both mask types. On the other hand, future multi-beam mask writer system may require carefully verifying the stitching performance between the swathes.

Standard strategies for mask registration do not provide data sampling of sufficient areal density to reveal such local registration (LReg) errors, as the required number of measurements would impact cycle time and diminish throughput way below cost-effectiveness of a mask registration tool. Therefore, today’s e-beam writer control scheme to verify actual local e-beam writer-induced error and deflector alignment performance relies on special test masks which are manufactured on a weekly or even less dense schedule. The typically observed LReg error signature requires a region of interest of several tens of microns to manifest as local area of increased lateral shift of the reticle device features.^{1} The novel approach taken by LMS IPRO7 offers the critical sampling density at reasonable throughput, thus enabling actual e-beam writer performance tests on each critical product mask, for both VSB (variable shaped beam) type and MBMW (multi-beam mask writer) type.

**2. Experimental methods**

The LMS IPRO7 platform offers high precision and repeatability for standard mask metrology applications utilizing a sequential measurement methodology. Data points of a grid that spans over the entire active area of the photomask are acquired by subsequent stage movement, image acquisition and image processing.^{2} The standard metric used to quantify the throughput of a mask metrology system is the Move, Acquisition and Measurement (MAM) time that typically is a few seconds per measurement site on the grid. For the sampling density and the field size required to determine e-beam mask writer signatures on production mask patterns, a MAM time of this order of magnitude will lead to a total measurement time of several days.

For LReg measurement a novel approach is taken, where the image acquisition is not fully sequential but the entire field of view of the metrology system is used to determine the position of several sites at the same time. The time for a typical LReg measurement of a region of interest (ROI) on the reticle of 100µm in length and width is hereby reduced to less than one hour. Within this region of interest single features with critical dimensions (CD) below 200nm can be measured with high precision inside dense patterns. The cycle time extrapolated to single-site MAM is equivalent to a reduction by a factor of 10^{4}. Thus, LReg measurement as a method to qualify commercial reticles is enabled.

Since there may be various root causes for e-beam writer-induced LReg error, methodologies from metrology tool grid calibration are utilized to confirm that the observed error signatures are not an artifact from the local registration measurement itself. To demonstrate rotational invariance, the same region of interest on the reticle is measured in standard reticle orientation (0°) and in 90°-rotated orientation. It is verified that observed LReg error signatures do not change with respect to the mask coordinate system. To test translation invariance of the LReg measurement method, the region of interest is shifted by a distance smaller than the distance between periodic signatures inside the field of view and error signatures are compared. Also, LReg measurements of the same region of interest with different sampling density are performed to rule out effects of the sampling strategy on the LReg error signature.

Various reticles are investigated by LReg measurement, containing high volume manufacturing (HVM)-quality test reticles written by both vector e-beam writer vendors. HVM-standard ArFi reticles as well as state-of-the-art EUV test reticles are investigated. For LReg measurement a region of interest with maximum length and height of 100µm is defined on a dense, on-device pattern of the mask layout. LMS IPRO enables sampling with period lengths of < 300nm for both x-dimension and y-dimension. The LReg results contain a high information density, since for each of the several thousand measurement sites the LReg error component for x-axis and y-axis is determined.

**3. Experimental Results**

To visualize the lateral distribution of the LReg error from a test measurement, two heat maps are used. For the sake of simplicity, the x-component and the y-component for one LReg dataset is depicted separately. The LReg error is the deviation of each measured site versus the corresponding design coordinates. Figure 1 represents such a visualization of a LReg measurement on an ArFi test reticle with dense contact array and a CD of 150nm. The lateral distribution of the x-deviation is depicted in Fig. 1a as a heat map, with blue representing negative x-shift, green representing zero shift and yellow representing positive shift of the measured site with respect to the x-axis. The lateral distribution of the y-deviation is depicted in Fig. 1b with identical scaling. A grid-like, systematic LReg error signature with period length of 10µm is observed in Fig. 1a. The y-component of the LReg error shows a cut-line, parallel to the x-axis. The amplitude of both error components is larger than 4nm.

**Figure 1. A region of interest with the area of 100µm times 100µm is measured on a dense contact pattern with a CD of 150nm on the ArFi-HVM test reticle A. The x-component of the LReg error is depicted in Fig. 1a for a measurement with the test reticle in standard orientation (0°), and in Fig. 1b the y-component of the LReg error of the same measurement is depicted. A total number of 32231 sites are measured for this LReg dataset.**

A major objective of the tests conducted within this study is to identify the signatures that are caused by the e-beam writer itself. Thus, it must be confirmed that signatures observed in the LReg results do not originate from the LReg measurement method. Figure 2 depicts the LReg results of the same ROI on the reticle measured in two different reticle orientations. The results are displayed in the actual mask coordinate system. Thus, any error that is related to the LMS IPRO measurement tool, e.g., caused by optics or the lateral translation stage, would be rotated by 90° in Fig. 1a with respect to Fig. 2b. The amplitude of the LReg error is scaled in arbitrary units but for Fig. 2a and 2b the same scaling is applied for the color code. The lateral distribution of the x-deviation is depicted in Fig. 2a as a heat map, with blue representing negative x-shift, green representing zero shift and yellow representing positive shift of the measured site with respect to the x-axis. A signature of equidistant lines, parallel to the y-axis is observed. The period length of these lines is around 9µm for the x-deviation (Fig. 2) and around 27µm for the y-deviation (not depicted), both with an amplitude of more than 4nm of local lateral shift. Since both images depicted in Fig. 2 show a very similar LReg error signature in the coordinate system of the test reticle, rotational invariance is confirmed.

**Figure 2. A region of interest with the area of 100µm times 100µm is measured on a dense contact pattern with a CD of 200nm on a EUV test reticle. The x-component of the LReg error is depicted in Fig. 2a for a measurement with the test reticle in standard orientation (0°) and in Fig. 2b for a measurement of the reticle rotated by 90° around the z-axis. The result data in Fig. 2b is back-rotated, thus the x-axis and y-axis, respectively, in Fig. 2a and Fig. 2b correspond to each other. A total number of 4875 sites are measured for this LReg dataset.**

Translation invariance of the LReg measurement method is tested on a dense contact pattern with a CD of 200nm on a ArFi reticle by shifting the ROI by 1.6µm in positive direction for both axes. The x-deviation result is shown in Fig. 2 where this shift of 1.6µm corresponds to one pixel. The overall LReg signature in Fig. 3a with a period length of 8µm is shifted by one pixel towards the center of origin in Fig. 3b, confirming that the LReg signature remains in the mask layout coordinate system when the ROI is shifted. The y-deviation is on the same order of magnitude and has a similar period length.

**Figure 3. With respect to a), the region of interest in b) is shifted by 1.6µm in positive x-axis and 1.6µm positive y-axis, which corresponds to one pixel. Measurements on test reticle E for a) and b) are both performed with the same resolution and with the reticle in standard orientation. A total number of 3969 sites are measured for this LReg dataset.**

To confirm that the sampling strategy does not cause any artifacts in the observed LReg signature, the same ROI, at the same orientation is measured with different sampling rates on a EUV test reticle. The x-deviation of the LReg signature is depicted in Fig. 4. Even though the signature in Fig. 4b is quite blurry due to the reduced sampling rate, it is confirmed that the overall period length of 9µm with an amplitude of more than 4nm from Fig. 4a is maintained.

**Figure 4. X-Component of LReg measurement on EUV test reticle C. The sampling rate in b) is reduced with respect to a) for two subsequent measurements of the same region of interest. The characteristic LReg signature of lines parallel to the y-axis is observed in a) and b). The sampling rate for b) is decreased by a factor of 1.5 along the x-axis with respect to a). A total number of 4875 sites for a) and 3225 for b) is measured for this LReg dataset.**

Table 1 contains an overview of different test results. Not all the obtained LReg results for different test reticles are displayed as figures in this contribution. EUV reticles and ArFi reticles are analyzed that have been written by different generations of vector e-beam writers from both major vendors. It can be stated that for all test reticles, for at least one of the error components, i.e., for x-axis or y-axis, a periodic signature is observed. For all LReg signatures the amplitude exceeded 4nm in lateral shift. The observed LReg signature (reticle domain, Table 1) is scaled to the wafer-level (wafer domain, Table 1) to depict the actual impact of the wafer device overlay error locally caused by the LReg signature. Even for state-of-the-art 10nm node reticles the wafer overlay impact exceeds 1nm.

**Table 1. Overview of the LReg signatures for different test reticles. All reticles except test reticle C are ArFi HVM test reticles. Test reticle C is an EUV test reticle. All test reticles except test reticle E are written with different scanner generations from one of the major suppliers for vector e-beam writers. Test reticle E is written by a vector e-beam writer from another supplier. The third column lists the corresponding lithography node for the measured contact pattern.**

**4. Discussion**

The results shown in Fig. 2, Fig. 3 and Fig. 4 show that the observed systematic lateral shifts are actually LReg signatures on the mask that are caused by the specific vector e-beam writer. Error signature artifacts that may be related to the LMS IPRO measurement methodology itself are excluded by demonstrating translation invariance, rotation invariance and independence of the error signature of the sampling rate.

It is likely that the period length of 42µm for ArFi reticle B (Table 1) correlates with the sub-deflector field settings of the e-beam writer. Test reticles A, B and C were written on a state-of-the-art e-beam writer. The observed error magnitude for these reticles is above the critical value for state-of-the-art HVM reticles regarding overlay. Test reticle E is manufactured by an older generation of e-beam mask writer than A, B and C but from the same manufacturer, thus the magnitude of the LReg error is even larger. Reticle D is manufactured by vector e-beam writer from another vendor, thus significant LReg error is observed for both major vendors.

EUV test reticle C shows a similar LReg error signature as ArFi test reticles A and B. This is expected since there has been no major change of e-beam writing strategy for EUV lithography. The periodic nature of the LReg error and the observed period lengths lead to the conclusion that the LReg error can be addressed by e-beam writer corrections. Since LMS IPRO is a registration system with high accuracy, the necessary correction parameters can be derived from the presented LReg datasets. All data presented have been obtained in single measurements that each lasted less than one hour. Given the observed period lengths that would allow for a reduction of the ROI, cost-effective characterization of production reticles (ROI vs. signature) within 15 minutes is enabled.

For evaluation of the measured LReg signatures (reticle domain) with regard to the impact as a non-correctable wafer device overlay error, the columns under wafer domain are calculated (Table 1). Corresponding to the scanner de-magnification factor of four, for both ArFi and EUV, the period length and magnitude are divided by factor of four. The wafer overlay is calculated by the assumption that two lithography layers based on reticles with the same LReg error are used. Thus, for the mean wafer device overlay error the magnitude is multiplied by two and divided by the square root of two. This also means that the local maximum error is even larger, e.g., for the 14nm node it is larger than 1.5nm. The method presented within this contribution offers a procedure to characterize the systematics of the LReg error on any product reticle within less than 15 minutes.

A proposed scheme for a feed-forward routine, based on LReg measurement is shown in Fig. 5. In addition to the standard specification of production reticles, i.e., precision and accuracy of the pattern, a LReg measurement is performed. As indicated by the presented results, such a LReg test would only add 15 minutes to the total reticle qualification time. Based on the LReg results improved writer corrections and writer calibrations may be implemented that will improve the pattern overlay.

**Figure 5. Proposed flow for enhanced quality control (QC) for production reticles before shipment from the mask shop. The LReg test region of 100 x 100 µm² can either be positioned on appropriate on-device patterns or on an adjacent test pattern, located outside the functional patterns of the production reticle.**

**5. Summary**

Cost-effective and accurate LReg error signature measurement has been demonstrated on state-of-the-art reticles for both ArFi and EUV lithography. The reported magnitude of the error is above the critical value for state-of-the-art HVM reticles regarding overlay. High sampling rate enables measurement of the LReg error on dense, periodic patterns and the large ROI enables deriving the corresponding e-beam writer corrections to significantly improve layer-to-layer pattern overlay for subsequent lithography steps.

**References**

- Chalom, D., Klikovits, J., Geist, D., Hudek, P., Eder-Kapl, S., Daneshpanah, M., Laske, F., Eyring, S. Roeth, K.-D., “Investigation of local registration performance of IMS Nanofabrication’s Multi-Beam Mask Writer,” SPIE 965805, (2015).
- Roeth, K.-D., Laske, F., Heiden, M., Adam, D., Parisoli, L., Czerkas, S., Whittey, J., Schmidt, K.-H., “Experimental Test Results of Pattern Placement Metrology on Photomasks with Laser Illumination Source Designed to Address Double Patterning Lithography Challenges,” Proc. SPIE Vol. 7488 (2009).

This paper was first published in Proc. SPIE 10810, Photomask Technology 2018, 108100A (12 October 2018); doi: 10.1117/12.2506286

## Leave a Reply