Automated Computer Vision-Enabled Manufacturing of Nanowire Devices

We present a high-throughput method for identifying and characterizing individual nanowires and for automatically designing electrode patterns with high alignment accuracy. Central to our method is an optimized machine-readable, lithographically processable, and multi-scale fiducial marker system—dubbed LithoTag—which provides nanostructure position determination at the nanometer scale. A grid of uniquely defined LithoTag markers patterned across a substrate enables image alignment and mapping in 100% of a set of >9000 scanning electron microscopy (SEM) images (>7 gigapixels). Combining this automated SEM imaging with a computer vision algorithm yields location and property data for individual nanowires. Starting with a random arrangement of individual InAs nanowires with diameters of 30 ± 5 nm on a single chip, we automatically design and fabricate >200 single-nanowire devices. For >75% of devices, the positioning accuracy of the fabricated electrodes is within 2 pixels of the original microscopy image resolution. The presented LithoTag method enables automation of nanodevice processing and is agnostic to microscopy modality and nanostructure type. Such high-throughput experimental methodology coupled with data-extensive science can help overcome the characterization bottleneck and improve the yield of nanodevice fabrication, driving the development and applications of nanostructured materials.

was later improved by including data rings 2 and colors in combination with position detection, 3,4 thus significantly increasing the information density. Perhaps the most common is the 2D barcode system QR code, where the outer three corners provide information about the orientation of the marker, and the inside of the marker contains a significant information payload. 5 The other most common fiducial markers are ARToolKit 6 and ARTag, which use a binary interior system to convey information. 7 However, the drawbacks of these systems are high false-positive detection rates 7 that increase with reducing marker size and are limited by the resolution of the imaging system, and image contrast. This problem has been addressed with further developments in fiducial marker designs, namely the CalTag system, which uses checker patterns for accurate subpixel calibration, 8 the AprilTag system, which uses lexicographic codes to demonstrate an improvement in detection of markers with reduced size, 9 and the RuneTag system, which uses patterns of circular dots of different sizes, 10 resulting in significantly improved detection rates. These tags have been designed for the applications that are visible with a naked eye, but the challenge of encoding and detecting location information on a sub-micron scale has not been addressed.
To demonstrate the challenges associated with fiducial marker patterning when approaching the lithography feature resolution limit we pattern a set of the most common fiducial markers.
Markers were fabricated using electron beam lithography (EBL) on an oxidized silicon substrate with and metallized with sputtered W followed by lift-off. Figure S2 shows SEM images of EBL patterned CCC, ARToolKit, ARTag, CalTag, AprilTag and RuneTag with a marker size of ~2x2 µm 2 , and Figure S3 with a marker size of ~20x20 µm 2 . It is clear that not all features of the fiducial markers are appropriate for nanofabrication techniques, as their pattern fidelity is not maintained during processing. ARToolKit, ARTag and AprilTag show damaged edges, which makes them unsuitable for the use in such small scale applications, since they rely on edge detection. The CCC and RuneTag are completely destroyed during lift-off, and also show bridging between features as a result of proximity effects or higher density of exposed features. Figure S2 shows non-lifted areas due to some exposed regions being enclosed, and the corners appear to be rounded due to a combination of effects, such as resolution of the resist and electron beam, proximity effects and minimum grain size of the sputtered W. Our system LithoTag has been designed to overcome these issues. Considering the minimum resolution of the lithography system, the smallest features of the marker must be larger than the minimum feature resolution. With the circle diameter being the limiting size factor, the smallest circle size for the LithoTag that can be achieved is about the same as the minimum feature resolution. The same minimum circle size in a RuneTag would mean the whole marker is almost twice the size of the LithoTag.
Previous fiducial markers such as AprilTag, ARToolKit and CalTag have been designed to retain detection accuracy for changes in projection angles, as they use sharp line edges and corners for detection. The LithoTag detection system does not use sharp edges as they are more likely to be damaged during processing. Instead, a convolutional technique is used for detection, which is not very robust to changes in viewing projection angle and so the marker can only ensure detection reliability if viewed from approximately above. However, for the purposes of nano-fabrication, the changes in projection angle viewing are a favorable trade-off against features such as oversaturation resistance and minimum resolution, where LithoTag has significant advantages.       Figure S6a shows the output CAD file that has been generated by computer-vision algorithm after obtaining the information on isolated nanowire locations, with the contacts being drawn with respect to each nanowire centre and rotated along its orientation direction. Figure S6b shows an SEM image taken of the same area after deposition of the contacts to show the automated fabrication accuracy.      (Figure S10). For an individual nanowire device with 1 µm channel length, the drain current drops by an order of magnitude after illumination, which can be predominantly attributed to InAs surface states. 11 Highly mobile electrons contribute to conduction along the channel in dark condition. Once the device is illuminated, electron-hole pairs are generated and the electrons can become trapped in the surface states in the native oxide, thus not contributing to the conduction. The photoexcited holes are left to recombine with equilibrium electrons, further reducing the number of free electrons in the channel and contributing to negative photoconductivity. Turning the light off again, the current gradually increases to approximately 60% of the dark current within 10 seconds. We demonstrate multiple NPC cycles by turning the light source on and off every 10 seconds for 100 seconds.
It has been shown that the current would be expected to reach the original dark current magnitude after turning the light source off given enough time for recovery. 12,13 Figure S10: Time-dependent conductivity measurements showing negative photoconductivity of automatically fabricated nanowire device under white light illumination (white) and dark conditions (grey shading).