AUTOMATED ATOMIC SCALE FABRICATION

Abstract

A method for autonomously applying a dangling bond pattern to a substrate for atom scale device fabrication includes inputting the pattern, initiating a patterning process, scanning the substrate using a scanning probe microscope (SPM) to generate an SPM image of the substrate, feeding the SPM image into a trained convolution neural network (CNN), analyzing the SPM image using the CNN to identify substrate defects, determining a defect free substrate area for pattern application; and applying the pattern to the substrate in that area. An atom scale electronic component includes functional patches on a substrate and wires electrically connecting the functional patches. Training a CNN includes recording a Scanning Tunneling Microscope (STM) image of the substrate, extracting images of defects from the STM image, labeling pixel-wise the defect images, and feeding the extracted and labeled images of defects into a CNN to train the CNN for semantic segmentation.

Claims

1. A method for autonomously applying a dangling bond pattern to a substrate for atom scale device fabrication: inputting the pattern to be applied to the substrate; initiating a patterning process; scanning the substrate using a scanning probe microscope (SPM) to generate an SPM image of the substrate; feeding the SPM image into a trained convolution neural network (CNN); analyzing the SPM image using the CNN to identify defects on the substrate; determining a suitable defect free area on the substrate for application of the pattern; and applying the pattern to the substrate in the suitable defect free area.

2. The method of claim 1 wherein the substrate is hydrogen-terminated Si(100) surface.

3. The method of claim 1 wherein the SPM image has a resolution of 10 pixels/nm.

4. The method of claim 1 wherein identifying defects on the substrate includes characterizing defects on the substrate.

5. The method of claim 1 wherein identifying defects on the substrate includes locating defects on the substrate.

6. The method of claim 1 wherein analyzing the SPM image includes pixelating the image.

7. An atom scale electronic component comprising: a plurality of functional patches on a substrate, each of the plurality of functional patches containing a dangling bond pattern; and a plurality of wires electrically connecting the plurality of functional patches.

8. The electronic component of claim 7 wherein each of the plurality of functional patches is formed in a defect-free are of the substrate.

9. The electronic component of claim 7 wherein the plurality of functional patches are disconnected on the substrate.

10. The electronic component of claim 7 wherein the plurality of wires are routed to avoid defects between the plurality of functional patches.

11. A method for training a convolution neural network (CNN) to assess the quality of a surface of a substrate for atom scale device fabrication, said method comprising: recording a Scanning Tunneling Microscope (STM) image of the surface of the substrate; extracting a plurality of images of defects in the surface from the STM image; labeling pixel-wise each of the plurality of images of the defects; and feeding the extracted and labeled plurality of images of defects into a convolution neural network one image at a time to train the CNN for semantic segmentation.

12. The method of claim 11 wherein the STM image recorded is of a hydrogen-terminated Si(100) surface.

13. The method of claim 11 wherein the STM image is recorded at a sample bias of 1.3 V or 1.4 V.

14. The method of claim 11 wherein the STM image is recorded with a tunneling current of 50 μA.

15. The method of claim 11 wherein the STM image has a resolution of 10 pixels/nm.

16. The method of claim 11 wherein the CNN is trained to identify H—Si (100) surface defects.

17. The method of claim 11 wherein the CNN is trained to locate H—Si (100) surface defects.

18. The method of claim 11 wherein the SPM image recorded is 40×40 nm.sup.2.

19. The method of claim 11 wherein the SPM image recorded is 100×100 nm.sup.2.

20. The method of claim 11 wherein the convolution neural network includes five convolution encoder layers and five convolution decoder layers.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention is further detailed with respect to the following drawings that are intended to show certain aspects of the present invention but should not be construed as a limit on the practice of the present invention.

[0018] FIGS. 1A-1L show 4×4 nm.sup.2 images taken with an STM with a sample bias of 1.4 V and 50 μA each showing empty state images of various defects found on a typical H—Si(100) surface;

[0019] FIG. 2 shows an embodiment of a CNN architecture according to the present invention, the CNN including 5 convolutional encoder layers followed by 5 convolutional decoder layers, each of the encoder and decoder layers having a ‘relu’ activation function applied, and wherein the final set of images is passed through one final convolutional layer followed by a softmax activation giving 7 separate images corresponding to each of the 7 labels with the output displaying the clean H—Si areas in black;

[0020] FIG. 3 shows original input 40×40 nm.sup.2 images (constant current of 50 μA and 1.4 V) with traces of the outputs outlined, where T2 defects are traced in purple, DBs in blue, dihydrides in yellow, step-edges in red, impurities in green, and clustered defects in orange;

[0021] FIG. 4 shows original input 100×100 nm.sup.2 images with traces of the outputs outlined, where T2 defects are traced in purple, DBs in blue, dihydrides in yellow, step-edges in red, impurities in green, and clustered defects in orange;

[0022] FIG. 5 shows original input 100×100 nm.sup.2 images with traces of clean H—Si outlined in teal;

[0023] FIG. 6A shows an image with clean H—Si areas traced in teal, they are identified by the square inset showing the cleanest area that is suitable for pattern application;

[0024] FIG. 6B shows an image of the surface shown in FIG. 6A after the pattern is applied within the square inset;

[0025] FIGS. 7A-7C show a logic flow chart of a patterning program according to embodiments of the present invention;

[0026] FIG. 8 shows a logic flow char of a tip shaping sub routine according to embodiments of the present invention; and

[0027] FIG. 9 shows a confusion matrix mapping the Dice coefficient for each of the predicted defects.

DESCRIPTION OF THE INVENTION

[0028] The present invention has utility as a system and method for automated atomic scale fabrication capable of characterizing and locating defects on the H—Si(100)-2×1 surface to allow for a rapid and commercially feasible way to mass produce electronic components with atomic precession with limited human intervention. The present invention implements an encoder-decoder type convolutional neural network (CNN).sup.28,29,30 to locate and classify features on the substrate surface. By using semantic segmentation,.sup.31,32 the neural network is trained to recognize a variety of defects commonly found on the H—Si(100) surface. After implementing the model with existing patterning,.sup.13 and probe tip forming suites,.sup.23 full automation of the patterning process is achieved.

[0029] It is to be understood that in instances where a range of values are provided that the range is intended to encompass not only the end point values of the range but also intermediate values of the range as explicitly being included within the range and varying by the last significant figure of the range. By way of example, a recited range of from 1 to 4 is intended to include 1-2, 1-3, 2-4, 3-4, and 1-4.

[0030] Crystalline silicon is tetravalent and forms a diamond lattice; each silicon atom shares 4 bonds, two above and two below the atom. At the surface, two of these bonds are unsatisfied so the crystal reorganizes to a lower energy configuration. The addition of atomic hydrogen to the silicon surface during the annealing process results in the formation of one of three possible phases. The likelihood of forming these phases can be controlled by the annealing temperature at which the sample is prepared. The 2×1 phase forms at ˜377° C., the 3×1 phase forms at ˜127° C., and the 1×1 phase forms below ˜20° C..sup.33,14,34 The most regularly used for DB patterning is the 2×1 phase reconstruction where each surface atom pairs with a neighboring surface atom to create a dimer pair. The dimer pairs form in rows which run parallel to each other across the surface. Each silicon atom at the surface is left with a single unsatisfied bond which extends out into vacuum and can either be terminated with hydrogen or left vacant creating a dangling bond. Although the preparation of the H—Si(100)-2×1 phase is well understood, it is often difficult to create a perfectly clean, defect free surface. These defects as well as clean H—Si(100) can be imaged using a STM.

[0031] In order to train the CNN to recognize these surface defects, the defects are labeled pixel-wise in the STM images. The neural network is trained with seven different classes of labels. The first is regular, clean H—Si(100)-2×1, as shown in FIG. 1A. FIG. 1B shows two types of charged defects labelled ‘type 2’.sup.35, the origin of which is unconfirmed. FIG. 1C shows ‘DB’ or dangling bond.sup.19,36. The remainder of the known surface defects are thought to exist in a neutral charge state and consist of diversely reconstructed H—Si, adatoms, and adsorbed molecules. FIG. 1D shows a ‘dihydride’ in which two silicon atoms each bind to 2 hydrogen atoms instead of forming a dimer pair.sup.37. FIG. 1E shows a ‘step-edge’ which is a drop in the surface height by one atomic layer. Dimer rows run perpendicular to the original direction above the step and the boundary of the step edge is often marked with 1×1 or 3×1 reconstruction..sup.38,39,40 FIGS. 1F-K show several different defects that either appear infrequently or are often found close to other defects, rendering them difficult to properly isolate for labelling. These defects and any others that were unknown were given the label ‘clustered’. FIG. 1L shows the final label class, an adsorbed species, molecule, or cluster of atoms of unknown origin labelled as ‘impurity’. These defects are thought to be something other than H—Si and can usually be reduced by eliminating any potential contaminants during sample preparation.

[0032] After the training data is acquired and labeled, as described in the experimental methods below, it is used to develop and train the CNN for semantic segmentation. Semantic segmentation allows for both the localization and classification of objects in images. This can be used in many applications where the network must make a distinction between different objects in an image including use in self-driving cars.sup.41,42,43 and medical image analysis..sup.44,45,46 Here, a distinction is made between the pixels that make up each of the labelled defects. The various CNN architectures are trained. The CNN architecture that shows the greatest performance in recognizing the defects is implemented, as shown in FIG. 2. An image of the surface is given as the input with a resolution of ˜10 pixels/nm. The image is then passed through 5 different decoder layers. Each layer consists of a convolutional filter (3×3 kernel). The number of filters is doubled for each layer starting with 32. A batch normalization is then applied to maintain a mean activation threshold of 0. An activation is applied using a rectified linear unit (relu) which sets all pixels with a negative value to 0. A max pooling downsampling (2×2 kernel) is finally applied which reduces the area of the image by a factor of 4. The smallest downsampled data set consists of 512 images that are 0.4% the size of the original image. The images are then passed through 5 subsequent encoder layers which bring the image size back to the same number of pixels as the original input. Each layer is structured similarly to the decoder layers except the downsampling function is replaced with an upsampling function which uses nearest-neighbour sampling (2×2 kernel) to increase the image size by a factor of 4. Label segmentation may be slightly improved if the upsampling is unpooled using learned indices from the max pooling of the same image size. The final set of 32 images is passed through a final convolution layer of seven filters each corresponding to one of the labelled classes. The resulting labelled images are passed through a softmax activation filter which normalizes the pixel values to correspond with a probability of label accuracy.

[0033] The network training data set is made from 28 images (100×100 nm.sup.2 and a resolution of 1028×1028). Each of the 28 images is divided into 64 smaller images (128×128). Each of the smaller images is rotated by 90°, 180°, and 270° as well as flipped along its axis and rotated again increasing the data set by a factor of 8 resulting in a total of 14336 images. The images are divided into training, testing, and validating images at a ratio of ˜2/3:1/6:1/6 respectively (corresponding to 9560:2384:2392 images). The Adam optimization algorithm.sup.47 is utilized with an exponentially decaying learning rate to achieve better convergence on a local minimum. The loss function used is a soft Dice loss function.sup.48. The soft Dice loss function is preferably used above other loss functions.sup.49,29 because it removes the need to weight different labels that appear less frequently. This solves any class imbalance which may otherwise be caused given that clean H—Si appears much more frequently than all other classes. FIG. 9 shows a confusion matrix mapping the Dice coefficient for each of the predicted defects.

[0034] A subset of the outputs of the fully trained CNN is shown in FIG. 3. The clean H—Si label is left out of FIG. 3 because it makes it difficult to see the boundaries of the defects. More examples of the predicted output labels including H—Si can are shown in the FIG. 4 and FIG. 5. The overall Dice coefficient score of the model is recorded at 84%. A large portion of the 16% inaccuracy is explained by subjectivity in labeling data without the use of a standard defect size established. The majority of the defects are seen to be correctly identified at the center of the defect, but correct identification is reduced further from the defect center. Accuracy of prediction is increased when the criteria of defects are standardized for the various labels applied. Generally, the size scale of the DB patterns is much larger than the size scale of any variations in the defect labels.

[0035] With the successful development of the neural network, specifically in locating charged defects, as well as clean H—Si, the CNN is implemented in the automation of DB patterning. FIGS. 6A and 6B demonstrate the current automation process. According to embodiments, a user inputs a specific pattern they wish to create, for example an OR gate pattern is input. The patterning process is then initiated. The SPM controller is then programmed to take a scan of the sample with a resolution similar to the training data. The image is fed to the neural network and the output image containing the location of clean H—Si is returned. The program then finds the largest patch of clean H—Si that would support the pattern, as shown in the inset of FIG. 6A. Once found, the program begins patterning until it is complete, as shown in FIG. 6B. A full flow chart of the patterning program is shown in FIGS. 7A-7C and FIG. 8. The same procedure is also applied for more complicated fabrication schemes.

[0036] Accordingly, the present invention provides a routine that can assess the quality of a sample and executes a device manufacturing procedure in an area that is free of defects to fully develop atomically-precise fabrication tools. This routine relies on the use of a CNN which uses semantic segmentation to identify and locate certain defects that inhibit the manufacturing process. The neural network is trained with images of defects commonly found on the H—Si(100)-(2×1) surface. It is envisioned that defect-free regions, or patches, adequate for fabrication of functional logic units comprised of roughly one hundred atoms will exist and that interconnections between such units will be custom routed so as to avoid defects. According to embodiments, defect-free regions, or patches, are somewhat disconnected by unused and or defective areas. Small functional circuit units are made in the patches identified by the machine and have wires interconnecting the patches to make larger function circuits. The wire paths may be determined on the fly, with each wire routed to avoid microscopic defects between good patches. In this way, defect-free surface areas are connected to form larger, effectively defect-free circuit blocks. The techniques shown here have applications in device fabrication using any form of scanning probe microscopy as well as subsets of semiconductor device fabrication where the quality of the materials used must be assessed to optimize the fabrication process.

EXAMPLES

[0037] The following experiments are performed using an Omicron LT STM operating at 4.5 K and ultrahigh vacuum (4×10.sup.−11 Torr). Tips are electrochemically etched from polycrystalline tungsten wire and resistively heated in ultrahigh vacuum to remove surface adsorbates and oxide, and sharpened to a single atom tip using field ion microscopy.sup.46. In situ tip processing is performed by controlled tip contact with the surface.sup.6,47,48. Tip shaping parameters are the same as those detailed in incorporated Reference 29 below.sup.29.

[0038] Samples used are highly arsenic doped (1.5×10.sup.−19 atoms/cm.sup.3) Si(100). Samples are degassed at 600° C. overnight followed by flash annealing at 1250° C. The samples are then terminated with hydrogen by exposing them to atomic hydrogen gas at 330° C.

[0039] Image and data acquisition are done using a Nanonis SPM controller and software. All training data is acquired at an imaging bias of either 1.3 V or 1.4 V with a tunneling current of 50 μA. The patterning automation routine is programmed in Python and Labview using the Nanonis programming interface library.

[0040] The CNN is implemented using Keras (2.1.3) with TensorFlow backend. Data was labelled using LabelMe (1.0) software.

[0041] References and patent documents cited herein are incorporated by reference to the same extent as if each reference was individually and explicitly incorporated by reference.

[0042] As a person skilled in the art will recognize from the previous detailed description and from the figures and claims, modifications and changes can be made to the preferred embodiments of the invention without departing from the scope of this invention defined in the following claims.

REFERENCES

[0043] [1] Floris A. Zwanenburg, Andrew S. Dzurak, Andrea Morello, Michelle Y. Simmons, Lloyd C L Hollenberg, Gerhard Klimeck, Sven Rogge, Susan N. Coppersmith, and Mark A. Eriksson. Silicon quantum electronics. Reviews of Modern Physics, 85(3):961{1019, 2013. [0044] [2] Dieter K. Schroder. Semiconductor Material and Device Characterization. John Wiley & Sons, Inc., Hoboken, N.J., USA, October 2005. [0045] [3] Enrico Prati and Takahiro Shinada. Atomic scale devices: Advancements and directions. Tech-nical Digest—International Electron Devices Meeting, IEDM, 2015-Febru(February): 1.2.1 {1.2.4, 2015. [0046] [4] Alexander Ako Khajetoorians, Jens Wiebe, Bruno Chilian, and Roland Wiesendanger. Realizing All-Spin-Based Logic Operations Atom by Atom. [0047] [5] Martin Fuechsle, Jill a Miwa, Suddhasatta Mahapatra, Hoon Ryu, Sunhee Lee, Oliver Warschkow, Lloyd C L Hollenberg, Gerhard Klimeck, and Michelle Y Simmons. A single atom transistor. Nature nanotechnology, 7(4):242{6, 2012. [0048] [6] Taleana R Huff, Hatem Labidi, Mohammad Rashidi, Mohammad Koleini, Roshan Achal, Mark H Salomons, and Robert A Wolkow. Atomic White-Out: Enabling Atomic Circuitry through Mechanically Induced Bonding of Single Hydrogen Atoms to a Silicon Surface. ACS Nano, 11:8636{8642, 2017. [0049] [7] F. E. Kal, M. P. Rebergen, E. Fahrenfort, J. Girovsky, R. Toskovic, J. L. Lado, J. Fernandez-Rossier, and A. F. Otte. A kilobyte rewritable atomic memory. Nature Nanotechnology, 11(11):926{929, November 2016. [0050] [8] A. Imre, G. Csaba, L. Ji, G. Orlov, G. H. Bernstein, and W. Porod. Majority Logic Gate for Magnetic Quantum-Dot Cellular Automata. Science, 331(5758):205 {208, June 2006. [0051] [9] Kai Sotthewes, Victor Geskin, Rene Heimbuch, Avijit Kumar, and Harold J W Zandvliet. Research Update: Molecular electronics: The single-molecule switch and transistor. APL Materials, 2(010701), 2014. [0052] [10] Niko Pavlicek, Zsolt Majzik, Gerhard Meyer, and Leo Gross. Passivation of dangling bonds on hydrogenated Si(100)-2$$1: a possible method for error correction in hydrogen lithography. June 2017. [0053] [11] Roshan Achal, Mohammad Rashidi, Jeremiah Croshaw, David Churchill, Marco Taucer, Taleana Huff, Martin Cloutier, Jason Pitters, and Robert A. Wolkow. Lithography for robust and editable atomic-scale silicon devices and memories. Nature Communications, 9(1):2778, December 2018. [0054] [12] M. Baseer Haider, Jason L. Pitters, Gino A. DiLabio, Lucian Livadaru, Josh Y. Mutus, and Robert A. Wolkow. Controlled Coupling and Occupation of Silicon Atomic Quantum Dots at Room Temperature. Physical Review Letters, 102(4):046805, January 2009. [0055] [13] G. P. Lopinski, D. D. M. Wayner, and R. A. Wolkow. Self-directed growth of molecular nanostructures on silicon. Nature, 406(6791):48{51, July 2000. [0056] [14] Xiao Tong, Gino A Dilabio, and Robert A Wolkow. A Self-Directed Growth Process for Creating Covalently Bonded Molecular Assemblies on the HSi(100)-31 Surface. Nano Letters, 4(5):979{983, 2004. [0057] [15] Alexander B Sieval, Ricarda Opitz, Huub P A Maas, Michael G Schoeman, Geert Meijer, Frank J Vergeldt, Han Zuilhof, and Ernst J R Sudholter. Monolayers of 1-Alkynes on the H-Terminated Si(100) Surface. Langmuir, 16(26):10359{10368, 2000. [0058] [16] J. W. Lynding, T-C. Shen, G. C. Abelin, C. Wang, and J. R. Tucker. Nanoscale patterning and selective chemistry of silicon surfaces by ultrahigh-vacuum scanning tunneling microscopy. Nanotechnology, 7:128{133, 1996. [0059] [17] S R Schofield, N J Curson, M Y Simmons, F J Rue, T Hallam, L Oberbeck, and R G Clark. Atomically Precise Placement of Single Dopants in Si. Physical Review Letters, 91(13), 2003. [0060] [18] Robert. Wolkow. A Step Toward Making and Wiring Up Molecular-Scale Devices. Japanese Journal of Applied Physics To, 40:4378, 2001. [0061] [19] J. R Tucker and T-C. Shen. Prospects for Atomically Ordered Device Structures Based of STM Lithography. Solid State Electronics, 42(7-8):1061 {1067, 1998. [0062] [20] J N Randall, J R Von Ehr, J B Ballard, J H G Owen, and E Fuchs. Automated Scanning Tunneling Microscope image analysis of Si (100):H 2×1 surfaces. Microelectronic Engineering, 98:214{217, 2012. [0063] [21] Mohammad Hadi Modarres, Rossella Aversa, Stefano Cozzini, Regina Ciancio, Angelo Leto, and Giuseppe Piero Brandino. Neural Network for Nanoscience Scanning Electron Microscope Image Recognition. Scientic Reports, 7(1):13282, December 2017. [0064] [22] Maxim Ziatdinov, Artem Maksov, and Sergei V. Kalinin. Learning surface molecular structures via machine vision. npj Computational Materials, 3(1):31, December 2017. [0065] [23] Maxim Ziatdinov, Ondrej Dyck, Artem Maksov, Xufan Li, Xiahan Sang, Kai Xiao, Raymond R. Unocic, Rama Vasudevan, Stephen Jesse, and Sergei V. Kalinin. Deep Learning of Atomically Resolved Scanning Transmission Electron Microscopy Images: Chemical Identication and Tracking Local Transformations. ACS Nano, 11(12):12742 {12752, December 2017. [0066] [24] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. arXiv, 2016. [0067] [25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3431{3440. IEEE, June 2015. [0068] [26] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning Deconvolution Network for Semantic Segmentation. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 1520{1528. IEEE, December 2015. [0069] [27] Jo˜ao Carreira, Rui Caseiro, Jorge Batista, and Cristian Sminchisescu. Semantic Segmentation with Second-Order Pooling. pages 430{443. 2012. [0070] [28] Ross Girshick, JeDonahue, Trevor Darrell, and Jitendra Malik. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 580{587. IEEE, June 2014. [0071] [29] Mohammad Rashidi and Robert A Wolkow. Autonomous Scanning Probe Microscopy in situ Tip Conditioning through Machine Learning. ACS Nano, 12:56, 2018. [0072] [30] J. J. Boland. Structure of the H-Saturated Si(100) Surface. Phys. Rev. Lett., 65(26), 1990. [0073] [31] Y J Chabal, Krishnan Raghavachari, and A Td'z. New Ordered Structure for the H-Saturated Si(1QQ) Surface: The (3×1) Phase. Technical Report 10, 1985. [0074] [32] Kitiphat Sinthiptharakoon, Steven R Schoeld, Philipp Studer, Veronika BrazdovaBr, Cyrus F Hirjibehedin, David R Bowler, and Neil J Curson. Investigating individual arsenic dopant atoms in silicon using low-temperature scanning tunnelling microscopy. Journal of Physics: Condensed Matter J. Phys.: Condens. Matter, 26:8, 2014. [0075] [33] S R Schoeld, P Studer, C F Hirjibehedin, N J Curson, G Aeppli, and D R Bowler. Quantum engineering at the silicon surface using dangling bonds. Nature Communications, 4:1649+, 2013. [0076] [34] Yuji Suwa, Masaaki Fujimori, Seiji Heike, Yasuhiko Terada, Yoshihide Yoshimoto, Kazuto Akagi, Osamu Sugino, and Tomihiro Hashizume. First Principles Study of Dihydride Chains on H-Terminated Si(100)-2{circumflex over ( )}A1 Surface. Japanese Journal of Applied Physics, 45(3B):2200{2203, 2006. [0077] [35] Masaaki Fujimori, Seiji Heike, Yuji Suwa, and Tomihiro Hashizume. Initial-Stage Dihydride Formation on Si(100)-2×1-H Surface. Japan Journal of Applied Physics, 42(11B):1387 {1390, 2003. [0078] [36] Chris Pearson, Brian Borovsky, Michael Krueger, Robert Curtis, and Eric Ganz. Si(001) Step Dynamics. Physical Review Letters, 74(14):2710{2713, 1995. 3 [0079] [37] German Ros, Sebastian Ramos, Manuel Granados, Amir Bakhtiary, David Vazquez, and Antonio M Lopez. Vision-based Oine-Online Perception Paradigm for Autonomous Driving *. Technical report, 2015. [0080] [38] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes. Technical report, 2016. [0081] [39] Marvin Teichmann, Michael Weber, Marius Zollner, Roberto Cipolla, and Raquel Urtasun. MultiNet: Real-time Joint Semantic Reasoning for Autonomous Driving. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 1013{1020. IEEE, June 2018. [0082] [40] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), pages 565 {571. IEEE, oct 2016. [0083] [41] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen A. W. M. van der Laak, Bram van Ginneken, and Clara I. Sanchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60{88, December 2017. [0084] [42] Eli Gibson, Wenqi Li, Carole Sudre, Lucas Fidon, Dzhoshkun I. Shakir, Guotai Wang, Zach Eaton-Rosen, Robert Gray, Tom Doel, Yipeng Hu, Tom Whyntie, Parashkev Nachev, Marc Modat, Dean C. Barratt, Sebastien Ourselin, M. Jorge Cardoso, and Tom Vercauteren. NiftyNet: a deep-learning platform for medical imaging. Computer Methods and Programs in Biomedicine, 158:113{122, May 2018. [0085] [43] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. December 2014. [0086] [44] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M. Jorge Cardoso. Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations. July 2017. [0087] [45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. May 2015. [0088] [46] Moh'd Rezeq, Jason Pitters, and Robert Wolkow. Tungsten nanotip fabrication by spatially controlled eld-assisted reaction with nitrogen. Journal of Chemical Physics, 124(20), 2006. [0089] [47] Sam Jarvis, Adam Sweetman, Joseph Bamidele, Lev Kantorovich, and Philip Moriarty. Role of orbital overlap in atomic manipulation. PHYSICAL REVIEW B, 85:235305, 2012. [0090] [48] Hatem Labidi, Mohammad Koleini, Taleana Huff, Mark Salomons, Martin Cloutier, Jason Pitters, and Robert A Wolkow. Indications of chemical bond contrast in AFM images of a hydrogen-terminated silicon surface. Nature Publishing Group, 8, 2017.