Inkjet-printed unclonable quantum dot fluorescent anti-counterfeiting labels with artificial intelligence authentication — TEI-tagged view

Yang LiupersName, Fei HanpersName, Fushan LipersName, Yan ZhaopersName, Maosheng ChenpersName, Zhongwei XupersName, Xin ZhengpersName, Hailong HupersName, Jianmin YaopersName, Tailiang GuopersName, Wanzhen LinpersName, Yuanhui ZhengpersName, Baogui YoupersName, Pai LiupersName, Yang LipersName, Lei QianpersName (2019date). “Inkjet-printed unclonable quantum dot fluorescent anti-counterfeiting labels with artificial intelligence authentication.” Nature Communicationstitle·j.

Object of study — inkjet-printed quantum-dot fluorescent security labelsterm·key with random flower-liketerm·key patterns

Research question — how to make an anti-counterfeitingterm·key label that meets all the practical requirements at once

Contribution — an unclonable, inkjet-printable, AI-decodable security labelterm·key from random substrate pinning pointsterm·dom

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

Inkjet-printed unclonable quantum dotterm·key fluorescent anti-counterfeitingterm·key labels with artificial intelligence authenticationterm·key Abstract An ideal anti-counterfeitingterm·key technique has to be inexpensive, mass-producible, nondestructive, unclonable and convenient for authenticationterm·key. Although many anti-counterfeitingterm·key technologies have been developed, very few of them fulfill all the above requirements. Here we report a non-destructive, inkjet-printable, artificial intelligence (AI)-decodable and unclonable security labelterm·key. The stochastic pinning pointsterm·dom at the three-phase contact lineterm·key of the ink droplets is crucial for the successful inkjet printingterm·key of the unclonable security labelsterm·key. Upon the solvent evaporation, the three-phase contact linesterm·key are pinned around the pinning pointsterm·dom, where the quantum dotsterm·key in the ink droplets deposited on, forming physically unclonable flower-liketerm·key patterns. By utilizing the RGB emission quantum dotsterm·key, full-color fluorescence security labelsterm·key can be produced. A convenient and reliable AI-based authenticationterm·key strategy is developed, allowing for the fast authenticationterm·key of the covert, unclonable flower-liketerm·key dot patterns with different sharpness, brightness, rotations, amplifications and the mixture of these parameters.

Introduction Counterfeitingterm·dom and forgery is a global problem that causes significant financial damageterm·dom and poses security threats to individuals, companies, and society as a whole 1 , 2 . Over the past decades, counterfeit products have spread from daily consumer goods, to medicines and high-tech products 1 . Although majority of the products are protected by an anti-counterfeitingterm·key technique, the global economic loss of counterfeitingterm·dom has been increasing annually and estimated to reach 1.7 trillion US dollars in 2015 3 . The reason for this is the currently used anti-counterfeitingterm·key technologies that rely on inkjet-printed security labelsterm·key can be readily duplicated by counterfeiters due to their uniform patterns and predicable, deterministic decoding mechanisms 4 – 6 . Despite this, the inkjet-printing technique itself has many distinguished advantages in low production cost, mass production, material-effective utility, unlimited pattern design ability, and excellent compatibility with various ink materials as well as supporting substrates 7 – 12 . Furthermore, the inkjet-printed macroscale security labelsterm·key allow fast, frequent authenticationterm·key by the naked eyes or using a smart phone.

Security labelsterm·key with physical unclonable functionsterm·mth (PUFsterm·mth) could offer a practical solution to the limitation of the widely used anti-counterfeitingterm·key technologies and seem to be the most viable path to combat the increasingly serious global forgery issue 13 . A PUFterm·def is a physical object with an intrinsic, unique, random physical feature generated in a non-deterministic process 1 , 13 – 20 . The randomness characteristic of the feature guarantees unreplicable code outputs. To date, a significant progress in PUFterm·def encryption has been made in the anti-counterfeitingterm·key field, mainly focusing on generating random features composed of either rough surfaces 21 , 22 or discrete nanoparticle arrays 23 – 25 within the predefined pattern areas. For example, Bae et al. has demonstrated the use of randomly wrinkling silica-coated polymeric particles as PUFterm·def codes to produce unique artificial fingerprints for anti-counterfeitingterm·key applications 22 . Our group recently have utilized advantage of electrostatic self-assembly strategy to generate randomly arranged plasmonic (metal) arrays as PUFterm·def codes using fluorescein-doped silver-silica core-shell nanoparticles as building blocks, giving rise to multi-optical signal encoded, unclonable security labelsterm·key 23 . However, the creation of well-defined two-dimensional (2D) graphic security labelsterm·key that carry PUFterm·def codes requires the aid of expensive lithography techniques. Moreover, the decryption of such PUF-based security labelsterm·key relies on a machine learning pattern recognition and comparison analysis that a real fingerprinting does 22 . The machine learning authenticationterm·key technique only focuses on algorithms for pattern recognition and enhancement to extract the PUFterm·def codes.

In this work, we develop an inkjet-printable, artificial intelligence (AI) decodable, unclonable, fluorescence security labelterm·key by combining the advantages of inkjet printingterm·key, portable smart-phone microscopes and AI techniqueterm·mth. The security labelsterm·key with diverse pattern designs are fabricated through inkjet printingterm·key using II-VI semiconductor core-shell quantum dotsterm·key as model ink. The surface decoration of print substrates, such as glass, plastic, or paper, with randomly distributed poly(methyl methacrylate)term·dom (PMMAterm·dom) nanoparticles is critical for the successful inkjet printingterm·key of unclonable security labelsterm·key. The polymer nanoparticles on the substrates are acted as stochastic pinning pointsterm·dom at the three-phase contact linesterm·key of the ink droplets. Upon the solvent evaporation, the three-phase contact linesterm·key are pinned around the pinning pointsterm·dom. The quantum dotsterm·key in the ink droplets subsequently deposit on the pinning pointsterm·dom, forming physically unclonable flower-liketerm·key dot patterns (i.e., primary units for any 2D logos). The designed surface modification makes the security labelsterm·key (i.e., 2D logos) unique and unclonable. By utilizing red, green and blue (RGB) emission semiconductor quantum dotsterm·key, full-color pictures can be generated, which are invisible in the ambient environment. The fabricated security labelsterm·key seem to be the same from batch to batch on the macroscopic level; however, at the microscopic level, they are quite different from each other. Compared with previous reports using the intrinsic surface topography of a material (e.g., scratch patterns, fiber weave, etc.) for PUFterm·def encoding 16 , the system presented here has many advantages: (1) the developed printing strategy for the security labelterm·key fabrication not only allows for various pattern design but also makes the mass-production at low cost possible; (2) the quantum dotterm·key ink is fluorescence active, guaranteeing the readout signals from suffering the interference by fingerprints and dusts; (3) the quantum dotterm·key security labelsterm·key are only visible upon UV excitation, which offers the first layer of security, while the PUFterm·def nature originated from the random flower-liketerm·key dot patterns provides the second more secure layer; (4) the first layer of security can be easily authenticated with naked eyes, while the second layer of security is able to be authenticated with AI techniqueterm·mth rather than time-consuming machine learning algorithmsterm·mth. Moreover, we introduce AI (specifically deep learningterm·key) security labelterm·key authenticationterm·key concept and succeed in accurately and robustly decoding the unrepeatable flower-liketerm·key dot patterns with different focusing degrees, brightness, rotation angles, amplification factors and the mixture of these parameters. The developed anti-counterfeitingterm·key technology meets all the requirements for commercial applications that are low cost, mass-producible, nondestructive, diverse full-color pattern design capability, unclonable, and convenient for authenticationterm·key.

Results Inkjet printingterm·key of unclonable quantum dotterm·key security labelsterm·key The fabrication process of unclonable security labelsterm·key by inkjet printingterm·key are illustrated in Fig. 1 . Three types of commercially available II–VI semiconductor core-shell quantum dotsterm·key (CdSe/CdS/CdZnS, ZnCdSe/CdZnS, and ZnCdS/CdZnS) that emitted red, green, and blue light were chosen as model inks owing to their high fluorescent quantum yield and outstanding stability (Fig. 1a ). These quantum dotsterm·key were synthesized using a well-established chemical route 26 , 27 , in which high boiling point oleic acid and 1-octadecen were used as a capping agent and solvent, respectively (see Methods for details). Typical low-magnification transmission electron microscopy (TEM) and high-resolution TEM micrographs present an extremely narrow particle size distribution and high crystallinity of the synthesized quantum dotsterm·key (Fig. 1a, b and Supplementary Fig. 1 ). The particle sizes of the red, green, and blue emission quantum dotsterm·key are 6.5 ± 2 nm, 10.5 ± 3 nm, and 11.0 ± 2.5 nm, respectively. Generally, for quantum dotsterm·key, the relationship between size and emissive wavelength accords with quantum confinement effect — that a small size corresponds to a larger band gap and emits short-wavelength fluorescence 28 . In this case, the emission wavelength of red, green and blue quantum dotsterm·key shown in Fig. 1a was determined by the composition rather than the size of quantum dotsterm·key (see Supplementary Fig. 1c ) 26 . The Fourier-transform infrared absorption peaks at 2923 and 2854 cm −1 , corresponding to C–C and C–H stretching modes of the -CH 2 - group, reveals the presence of oleic acid on the quantum dotterm·key surface (see Supplementary Fig. 2 ) 29 . The surface ligands are of importance for the preparation of highly dispersed and stable quantum dotterm·key inks as they can stabilize the quantum dotsterm·key in non-polar solvents for months at room temperature. In this work, n-octylcyclohexane with low evaporation rateterm·dom at ambient environment was adopted as model solvent for ink preparation. The low evaporation rateterm·dom of the jetted ink droplet leads to a slow movement of the three-phase contact lineterm·dom on the print substrate, which allows us to dynamically study the quantum dotterm·key deposition behavior. Photographs of twenty jetted single droplets show almost identical shape both before contacting a print substrate and at the early stage after the contact (see Supplementary Fig. 3 ). Fig. 1 Outline of formation processes of unclonable security labelsterm·key by inkjet printingterm·key. a Typical TEM/HRTEM images (left), particle size distributions (middle), and ink for inkjet printingterm·key (right) of red core-shell quantum dotsterm·key (QDs). b Surface modification of a substrate with partly dissolved PMMAterm·dom solution by spin coating. c Substrate with random-distributed pinning pointsterm·dom of PMMAterm·dom nanoparticles on PMMAterm·dom film after spin coating. d Inkjet printingterm·key of prepared QDs ink on the modified substrate. e Pinning process of QDs droplet contact line at random pinning pointsterm·dom. f Residual central ink teared in to several smaller droplets in the final evaporation process. g A resulting security labelterm·key composed of flower-liketerm·key dot patterns, which are protected with a thin, optically transparent sticky gel film Prior to inkjet printingterm·key, the print indium-tin-oxide (ITO)-coated glasses were cleaned by sonication in various solvents (see Methods for detailed sonication cleaning procedure), treated with oxygen plasma and then spin-coated with a partly-dissolved poly(methyl methacrylate)term·dom (PMMAterm·dom) 30 nanoparticle colloid, forming a randomly distributed PMMAterm·dom nanoparticle array on a PMMAterm·dom layer (Fig. 1b, c ). The surface cleaning and oxygen plasma treatment were utilized to remove the dusts and generate a super wettable surface for the PMMAterm·dom nanoparticle coating, while the surface decoration with PMMAterm·dom nanoparticles were used to create randomly arranged pinning pointsterm·dom for the quantum dotterm·key deposition during the subsequent inkjet printingterm·key process. Such randomly arranged pinning pointsterm·dom are critical for the successful inkjet printingterm·key of unclonable security labelsterm·key (see Supplementary Fig. 6 ). Representative dark-field and atomic force microscopy images revealed the presence of randomly distributed PMMAterm·dom nanoparticles with particle size in the range of a few nanometers to a few micrometers on the print substrates (see Supplementary Fig. 4 ). Following the surface modification of the print substrates, they were subjected to the inkjet printingterm·key to produce 2D graphic security labelsterm·key. The inkjet printingterm·key was conducted by continuously jetting the quantum dotterm·key ink droplets with size of ~130 μm on the print substrates (Fig. 1d ). The interval between two adjacent ink droplets is 200 μm. If each macroscopic security labelterm·key contains 1000 flower-liketerm·key points, we can achieve 1600 security labelsterm·key in 5 min (i.e., the drying time of each batch of security labelterm·key) with our single-nozzle printing machine. Each droplet represents a pixel of the inkjet-printed security labelsterm·key after being completely dried. The PMMAterm·dom nanoparticles on the poorly wettable PMMAterm·dom film were acted as stochastic pinning pointsterm·dom at the three-phase contact linesterm·key of the ink droplets. With the evaporation of solvent, the three-phase contact lineterm·key continually slides and shrinks at the smooth PMMAterm·dom areas owing to the poor wettability of the surface but is captured by some pinning pointsterm·dom. Consequently, the pinning pointsterm·dom stretch and pin the contact line, distorting the fluid convex and forming irregular quantum dotterm·key ink pattern (Fig. 1e ). During the solvent evaporation process, the quantum dotterm·key concentration increases gradually. When the quantum dotterm·key concentration reached their saturation point, they started to deposit at the stochastic PMMAterm·dom pinning pointsterm·dom on the three-phase contact lineterm·key, forming a unique flower-liketerm·key dot pattern (Fig. 1f ). With the shrinking of droplet, the volume-smaller droplet is more liable to be tortured by pinning pointsterm·dom, thus splitting into several smaller sub-droplets (see Supplementary Fig. 5 ). Due to the small space between these sub-droplets and some stochastic external factors (such as airflow), they experienced merging and splitting irregularly before eventual drying (Fig. 1f ). This non-deterministic, random pattern formation process was recorded by a camera to monitor the shape evolution of a droplet as a function of time (see Supplementary Fig. 5 ). Such a non-deterministic pattern formation process makes it impossible to reproduce the flower-liketerm·key dot patterns, which were used as PUFterm·def codes in our security labelsterm·key. The fabricated security labelsterm·key were then covered with a thin, optically transparent, sticky gel film to protect them being damaged during the real circulation (Fig. 1g ). If without the PMMAterm·dom nanoparticles on the print substrates, the accumulated quantum dotsterm·key induced by solvent evaporation directly deposited on the self-pinned three-phase contact lineterm·key, producing a coffee-ring-like pattern (see Supplementary Fig. 6a ). However, if the print substrates were coated with a continuous, smooth, poorly wettable PMMAterm·dom layer, the three-phase contact lineterm·key of the droplets are incline to shrink and slide towards to the center, rather than pinned on substrates, bringing the quantum dotsterm·key to center of the droplet. When the solvent is completely evaporated, a central bump pattern, i.e., a hilly accumulation with diameter much smaller than the initial wetted diameter of the ink droplet, is generated (see Supplementary Fig. 6b ). In the absence of the stochastic PMMAterm·dom pinning pointsterm·dom, either a uniform coffee ring or central bump pattern was formed, which was determined by the wettability property of the print substrates. These deterministic pattern formation processes ensure the reproducibility of the patterns under the same conditions.

The miniaturized dark-field and fluorescence microscopes integrated with smartphones have been developed for single nanoparticle imaging 31 , 32 . Such small, affordable, portable microscopes were utilized by consumers to authenticate the inkjet-printed security labelsterm·key. The magnification-adjustable objective lens of the portable microscope used here is covered with a cylindrical metal shell that creates a small dark imaging environment by blocking the ambient light (Supplementary Fig. 7 ). The portable microscope is able to quickly (typically within one second), accurately and nondestructively readout the produced flower-liketerm·key dot patterns, which show different geometries as expected (Supplementary Fig. 7 ).

Unlimited colorful pattern design capability Ideal fluorescent security labelsterm·key require the ink materials having strong fluorescence intensity, which guarantees the high fluorescence brightness. The prepared quantum dotterm·key inks exhibit a narrow fluorescence emission peak with maximum intensity at 459 nm, 532 nm, and 631 nm, respectively, when excited at 375 nm (Fig. 2a ). The full-width at half-maximum (FWHM) of these blue, green and red emission quantum dotsterm·key is 20 nm, 24 nm, and 28 nm, respectively. They also bear a high fluorescence quantum yield of 75, 90, and 80%, respectively. The time-resolved fluorescence decays of the blue, green, and red emission quantum dotsterm·key show their lifetime of 27.5 ns, 17.0 ns, and 7.6 ns (Fig. 2b ), respectively, in consistent with consistent with previous reports 26 , 33 , 34 . According to the Commission Internationale de I’éclairage (CIE) 1931 color coordinate triangle, the quantum dotsterm·key achieved a wide color gamut of 124% of the National Television System Committee (NTSC) standard (see Supplementary Fig. 8 ), better than phosphor (color gamut: 85.6%) 35 – 37 , implying outstanding full-color fluorescence properties for security labelterm·key applications. Fig. 2 Optical characterization. a Fluorescence spectra and b time-resolved fluorescence decay measurements of red, green, and blue emission quantum dotterm·key (QD) inks excited at 375 nm. c Transmittance spectra of ITO coated glass substrates before and after inkjet printingterm·key of red, green, and blue emission Fuzhou University logos (Inset: photographs of the four samples under ambient light). d – i Fluorescence images of red, green, and blue emission d – f Fuzhou University logos and g – i two-dimensional QR codes and bar codes. The Fuzhou University logos, QR codes and bar codes, composed of thousands of dot patterns (pixels), are adapted with permission of the Fuzhou University. The scale bar is 1 cm Complex Fuzhou University logos composed of red, green, or blue emission arrays were inkjet-printed on indium tin oxide (ITO) coated glass slides using the printing strategy described in Fig. 1 . The print substrates with Fuzhou University logos maintained transparent (Insert of Fig. 2c ). That is, the inkjet-printed logos are invisible by naked eyes under the ambient conditions. The transmittance curves of the printed ITO glass slides are almost identical to the pristine print substrate (Fig. 2c ), further confirming the covert characteristic of the fabricated security labelsterm·key. When irradiated with UV light, bright red, green and blue photoluminescence photographs of the university logos were observed by the naked eyes (Fig. 2d–f ). Any other 2D macroscale patterns with sharp corners, for example, QR codes and bar codes (Fig. 2g–i ), can also be produced using our inkjet printingterm·key technique. The property that these images are only seen upon UV excitation offers the first layer of security realized by naked eye authenticationterm·key of macroscopic patterns; and the PUFterm·def nature of the flower-liketerm·key patterns is the second more secure layer.

By tuning RGB ink component ratios, fine tonal variations from red to green were obtained (see Supplementary Fig. 9 ). In principle, the full palette of colors could be achieved with elaborate control of the RGB ink components. Alternatively, the full-color security labelsterm·key can be generated by creating individual color pixels including red, green, and blue sub-pixels, which has been widely used in display 26 , 38 , 39 and color images 40 – 42 . The inkjet printingterm·key strategy developed here offers us an opportunity to extend the ink to other fluorescent materials, e.g., lanthanide complexes 43 – 47 , carbon dots 48 , 49 , and up-conversion nanoparticles 50 – 52 , etc. and the print substrates to flexible ones, such as plastics (see Supplementary Fig. 10 ). Although the 2D macroscale security labelsterm·key shown in Fig. 2d-i contained thousands of discrete points (or pixels), the printing process is completed within a few minutes using our single-nozzle printing machine.

Figure 3 displays fluorescence micrographs of three batches of inkjet-printed macroscopic letters FZU (i.e., the acronym of Fuzhou University) that are composed of red, green, or blue emission arrays. At the macroscopic level, the printed FZU letters were seemingly identical (Fig. 3a–c ). However, under the microscopic level, they were completely different from each other, as every fluorescent pixel of the letters showed a unique, random and unrepeatable flower-liketerm·key geometry (Fig. 3d–i and Supplementary Fig. 11 ). Further scanning electron microscope and energy dispersive X-ray spectroscopy characterizations verify that the unique flower-liketerm·key micropattern obtained from optical imaging systems is consistent with quantum dotterm·key deposition zones (see Supplementary Figs. 1 2– 14 ). By carefully comparing all the pixels within the letters (i.e., the same sample), no identical flower-liketerm·key micropatterns were found. A copy of the red, green, and blue fluorescence counterparts of the FZU letters shown in Fig. 3 fabricated under the same conditions also shows entirely different geometries of the corresponding pixels (see Supplementary Fig. 15 ). The encoding capacity of the inkjet-printed security labelsterm·key that relies on the encoding capacity (defined as l) and the number (defined as m) of the PUFterm·def patterns (i.e., flower-liketerm·key pixels) within the security labelsterm·key can be described as l m 20 . Recently, Carro-Temboury et al. established a universal binary-bit model for the estimation of the encoding capacity of PUFterm·def pattern 1 . According to Carro-Temboury’s model, the encoding capacity of a red flower-liketerm·key PUFterm·def pattern, l, is calculated to be 4.7 × 10 202 (see Supplementary Fig. 1 6 and Note 1 for calculation details). Therefore, for a security labelterm·key composed of 1000 red flower-liketerm·key PUFterm·def patterns, its encoding capacity will be larger than 10 202,000 . This indicates that the inkjet-printed security labelsterm·key are unclonable even by the manufacturer. Moreover, the fabricated security labelsterm·key have been frequently exposed to UV light over the period of 2 months. No obvious decrease of the fluorescence brightness is observed, revealing that security labelsterm·key have excellent chemical and photo stabilities (see Supplementary Fig. 17 ). Fig. 3 Security labelsterm·key with physically unclonble flower-liketerm·key patterns. Typical fluorescence images of inkjet-printed macroscopic letters FZU composed of a red, b green, and c blue emission dot patterns (the scale bar is 4000 μm), d – f their local enlarged images (the scale bar is 500 μm), and g – i further enlarged images (the scale bar is 100 μm). All the samples are printed under the same conditions. The FZU patterns are identical on macroscale but completely different at the microscale (i.e., every dot pattern is unique) Deep learningterm·key decoding mechanism Conventional security labelsterm·key with macroscale features, for instance, QR codes or bar codes are convenient for authenticationterm·key with naked eyes, but easy to be counterfeited. Advanced security labelsterm·key that carry complex, random nano/microscale features (PUFterm·def keys) are formidable to be faked. The decoding of such security labelsterm·key relies on a similar character extraction, recognition, and comparison analysis that a real fingerprint does 22 , 23 . Such machine learning algorithmsterm·mth for pattern recognition, enhancement, and identification are widely employed for the authenticationterm·key of the PUF-based security labelsterm·key 22 , 23 . A general drawback of the machine learning approach is time consuming and having a high level of false positives of up to 20% 20 . Furthermore, conventional classification can only be used in the PUFterm·def that can be transformed to private keys; for flower-liketerm·key patterns, it does not work. To deal with this problem, deep learningterm·key, an artificial intelligence (AI) technique, was introduced to validate the fabricated unclonable security labelsterm·key, which pushes anti-counterfeitingterm·key technology to a higher dimension.

Although AI has been recently applied into organic chemistry synthesis and TEM image analysis 53 , 54 , to the best of our knowledge, this is the first time to use AI for security labelterm·key authenticationterm·key. Figure 4 demonstrates a typical authenticationterm·key procedure of the inkjet-printed security labelsterm·key through deep learningterm·key. First of all, each quantum dotterm·key security labelterm·key on a commercial product is imaged using an advanced fluorescence microscope (Fig. 4a , step 1). Only one image is taken from each security labelterm·key and represents one PUFterm·def code. Simply, randomly shifting and rotating the image creates a large number of images for an AI to learn the characteristic features of the security labelterm·key. Once the images are trained on AI, they are categorized in a very general manner (e.g., classes associated with geometry of the security labelsterm·key) and then stored as a database in a deep learningterm·key engine for the subsequent authenticationterm·key using (Fig. 4a , step 2). This is done by the manufacturer. When the consumers receive the products, they can simply use their portable mini-microscope-connected smartphones to readout the PUFterm·def codes by taking photos of the security labelsterm·key (Fig. 4a , step 3), which are automatically sent to the deep learningterm·key engine for validation (Fig. 4a , step 4). The deep learn engine immediately feeds back the authenticationterm·key results (real or fake) to the users (Fig. 4a , step 5). Fig. 4 Deep learningterm·key decoding mechanism. a Schematic illustrating the authenticationterm·key process: 1: image capture and database generation from the goods by the manufacturers, 2: image learning by AI, 3: Image capture from the goods by the consumers using their smart phones, 4: image recognition and comparing by AI, and 5: authenticationterm·key outcome feedback to the consumers. b – g A library of six single dot pattern security labelsterm·key, h – m six fluorescence images taken from b (referred as to genuine product) with different brightness, sharpness, rotation angles, magnifications, and the mixture of the above-mentioned factors, and n – s six fluorescence images from security labelsterm·key that are not in the database shown in panel b – g (referred as to fake products). t , u Recognition rates of labels (h1–h6) and (i1–i6) by the authenticating way of deep learningterm·key. Labels (g1–g6), (h1–h6), and (i1–i6) correspond to the labels in figure ( b – g ), ( h – m ), and ( n – s ) respectively. The color scales from blue to red stand for the matching score (ranging from 0 to 100%) of the captured pictures with the labels To experimentally demonstrate the above authenticationterm·key process, six quantum-dot security labelsterm·key (named as g n , n = 1, 2, ···, 6) were randomly chosen to establish a security labelterm·key database (Fig. 4b–g ). Five hundred fluorescence images of each security labelterm·key (e.g., g 1 ) obtained by randomly shifting and rotating a same image (i.e., g 1 ) are provided to AI for learning and classifying. The selected 72 out of the 500 images from g 1 show the exact same geometrical characteristics of the security labelterm·key (see Supplementary Fig. 18 ). The 500 images were divided into two parts: 80% for learning and 20% for validation. After every learning cycle (the parts for learning), the images for validation were sent to AI engine to test, providing a train accuracy plot. After about 1000 learning cycles, they can be recognized with accuracy fluctuating between 97 and 100% when being sent to AI for validation again (see Supplementary Fig. 19 ).

For decoding, a security labelterm·key that represented a genuine product (i.e., from the pre-established database) was imaged at various sample rotation angles, magnification, focusing degrees, and the mixture of the above-mentioned factors (Fig. 4h–m ). We try to cover all possible deviations from the imaging equipment, imaging conditions, personnel habits of users that may happen in a real authenticationterm·key scenario. None of the images shown in Fig. 4h–m (named as h n , n = 1, 2, ···, 6) has ever been previously learnt by AI. The image h n is fed into the trained AI for validation. The authenticationterm·key outputs show the accuracy of h n ( n = 1, 2, ···, 6) is 0.999, 0.758, 0.999, 0.909, 0.999, and 0.999, respectively (Fig. 4t ). The relatively low accuracy of h 2 and h 4 are attributed to their over-indistinct characteristics, for which parts of details are lost during imaging, implying sharpness exerting a much higher impact on authenticationterm·key accuracy than other variations, such as brightness, location, rotation angle, and amplification factor. The threshold of the accuracy at a value of 0.5 is then set to distinguish the real and fake security labelsterm·key. For comparison, six fake security labelsterm·key named as i n ( n = 1, 2, ···, 6) were sent to AI for authenticationterm·key in the same way (Fig. 4n–s ). The corresponding accuracy is almost zero for all the fake security labelsterm·key (Fig. 4u ). By simply comparing the accuracy on the test with the threshold, the deep learningterm·key machine can immediately provide the authenticationterm·key outcomes (real: accuracy ≥0.5, fake: accuracy <0.5) to the customers. Regarding to the rate of false positives, we achieved the false positives rate of 0 using the match score of 0.5 as the threshold when sampling 100 security labelsterm·key (see Supplementary Table 1 ). It only takes seconds or even less to finish the whole authenticationterm·key process.

Discussion In conclusion, we have demonstrated a non-destructive, inkjet-printable, smart-phone readable, AI decodable, unclonable, and fluorescence security labelterm·key. Such security labelsterm·key with various 2D patterns composed of red, green, or blue emission arrays were fabricated through inkjet printingterm·key using II–VI semiconductor core-shell quantum dotsterm·key as model ink. The surface modification of the print substrates with randomly arranged PMMAterm·dom nanoparticles is crucial for the successful inkjet printingterm·key of the unclonable security labelsterm·key. The polymer nanoparticles on the print substrates were acted as stochastic pinning pointsterm·dom at the three-phase contact linesterm·key of the ink droplets for the quantum dotterm·key deposition, forming physically unclonable flower-liketerm·key dot patterns. This non-deterministic pattern formation process guarantees the unclonability of the fabricated security labelsterm·key. By utilizing the RGB emission core-shell quantum dotsterm·key, full-color security labelsterm·key can be generated.

The fabricated security labelsterm·key are invisible in the ambient environment but can be visualized by naked eyes when irradiated with UV light, which offers an easy way for the preliminary verification. A more reliable authenticationterm·key strategy by using AI techniquesterm·mth has been developed. Covert and unclonable flower-liketerm·key dot patterns with different sharpness, brightness, rotations, amplifications, and the mixture of these parameters have been successfully decoded within seconds using the authenticationterm·key strategy developed here. The overall cost per security labelsterm·key has been estimated to be approximately US$ 0.011 (see Supplementary Note 2). The anti-counterfeitingterm·key technology described in this work is a good step closer to commercial applications that are low cost, mass-producible, nondestructive, diverse full-color pattern design capability, unclonable, and convenient for authenticationterm·key. Applications of this technology span the full range from established to emerging-technology industries, including pharmaceutics, food security, and nanotechnology. Compared with intrinsic surface topography of material itself (like scratch patterns, fiber weave, etc.), our current system shows advantages in fluorescent and multi-color information, multi-level security, convenient for authenticationterm·key, and well-designed patterns. The possibility of covert but easily detectable labeling through the use of a portable mini microscope and AI techniqueterm·mth (Supplementary Table 2 ) will ensure the security and trackability of sensitive substances and equipment, which will lead to a promising approach in sensitive industries such as the nuclear one.

Methods Materials Poly(methyl methacrylate)term·dom (PMMAterm·dom, average Mw ~996,000 g mol −1 , from Sigma-Aldrich), polyvinylpyrrolidone (PVP, average Mw ~40,000× g mol −1 ), dimethyl sulfoxide (DMSO), chlorobenzene (99.8% pure, from J&K Scientific), CdO (AR grade, from Aladdin), ZnO (99.7%, from Shijiazhuang hongda zinc industry co. LTD), zinc acetate dehydrate (AR grade, from Aladdin), oleic acid (OA, 90%, from Alfa aeser), 1-octadecene (ODE, 98%, from Toyata), Se powder (99.999%, from Alfa aeser), sulfur powder (99.95%, from Aladdin), 1-Dodecanethiol (DDT, ≥ 98%, from Chevron Phillips Chemical), toluene (for synthesizing quantum dotsterm·key, AR grade, from Guangdong Guanghua Sci-Tech Co., Ltd), and ethanol (AR grade, from Guangdong Guanghua Sci-Tech Co., Ltd) were received and used without further purification. Sticky-gel film with gel thickness of 1.5 mm and retention level of X4 was purchased from Gel-Pak company. Toluene (for dissolving PMMAterm·dom) was purchased from Sinopharm Chemical Reagents and further dried by distillation over sodium.

Synthesis of core-shell quantum dotsterm·key For a typical synthesis of CdSe/CdS/CdZnS red emission quantum dotsterm·key: 7 mmol of CdO, 10 mL of OA and 25 mL of ODE were mixed in a 250 mL round flask. The mixture was heated to 160 °C, degassed for 15 min, then filled with N 2 gas and further heated to 310 °C. Subsequently, Se precursor was injected swiftly into the flask. Then the reaction was cooled to 300 °C and remained at this temperature for 5 min. To grow the CdS shell, S source was injected dropwise to react with the remaining cadmium ions for 20 min. For the growth of CdZnS shell, keep the reaction temperature and dropwise injected the pre-prepared shelling materials, the shelling process last for ~30 mins. After the reaction, the temperature was naturally cooled down to room temperature. The synthesized quantum dotsterm·key were finally purified using toluene and ethanolfor several times and finally dispersed in n-octylcyclohexane. For ZnCdSe/CdZnS green emission quantum dotsterm·key: 35 mmol of ZnO, 25 mL of OA, and 20 mL of ODE were mixed in a 250 mL round flask. The mixture was heated to 160 °C, degassed for 15 min, then filled with N 2 gas and further heated to 300 g to get a clear solution. The solution was cooled to 200 °C, at which Se and Cd stock precursors were quickly injected into the flask sequentially. Then the temperature was elevated to 310 °C and remained at this temperature for 30 min to form alloyed CdZnSe core. For the growth of the CdZnS shell, S source and Cd source were dropwise injected into the flask repeatedly while keeping temperature at 270 °C. After the precursor injection, the temperature was kept unchanged for 30 min to 1 h, and then naturally cooled down to room temperature. For the synthesis ZnCdS/CdZnS blue emission quantum dotsterm·key with composition gradient, the procedure is the same as that for the green emission quantum dotsterm·key, except using S stock solution instead of Se solution as the anion precursor. At the elevated temperature, the S precursor was injected into the mixture of the Zn precursor and Cd precursor in a round bottom flask, forming the blue emitting ZnCdS cores. The shell materials (S source and Cd source) were added afterwards drop-wisely, and the reaction was allowed to proceed from 30 min to 1 h for the shell growth. The synthesized of green and blue quantum dotsterm·key were purified and prepared with similar process as above.

Fabrication of quantum-dot security labelsterm·key In a typical procedure, the precursors of surface modification layer were prepared by adding solutes into chlorobenzene, toluene, and DMSO for common PMMAterm·dom (4 mg ml −1 ), designed PMMAterm·dom (4 mg ml −1 ), and PVP (4 mg ml −1 ) layer respectively, and stirring vigorously for 12 h at 60 °C before use The quantum dotterm·key inks were dispersed in n-octylcyclohexane at a concentration of 20 mg ml −1 . Indium-tin-oxide (ITO)-coated glass substrates were cleaned with ultrasonication successively in deionized (DI) water, acetone, isopropanol, and DI water. Then, nitrogen stream was used to dry the substrates, followed by 10 min oxygen plasma treatment. The pre-prepared precursors of surface modification layer were deposited on the substrates by spin coating, and then were heated at 120 °C for 30 min. The quantum-dot inks were printed on the substrates with a Microfab JETLAB II equipped with a 30-μm diameter piezoelectric-driven inkjet nozzle and a motorized stage with the accuracy of 5 μm. Driving voltage waveforms to the inkjet printingterm·key nozzle and the flying single droplet are shown in Supplementary Figs. 20 and 21 . All the processes were operated in the ambient environment. If every single macroscopic pattern has 1000 unclonable flower-liketerm·key points, we can achieve 1600 macroscopic patterns in 5 min with our single-nozzle printing machine. The time of drying is ~5 min for each batch of security labelterm·key. The as-fabricated security labelsterm·key were covered with gel films by tearing of their polycarbonate coversheet and then stick their gel material on the labels for stability test (see Supplementary Fig. 17 ).

Characterization The surface morphology of the PMMAterm·dom films and quantum-dot patterns were characterized with atomic force microscopy (AFM, Bruker Multimode 8) and scanning electronic microscope (SEM, FEI, Nova Nano SEM 230). The UV−Vis absorption spectra were tested with a UV/Vis/NIR spectrophotometer (Shimadzu, UV-3600). The transmission electronic microscopy (TEM) image of quantum dotsterm·key was recorded using JEOL JEM-2100F microscope. The steady-state photoluminescence (PL) spectra were collected with a Hitachi F-4600 fluorescence spectrophotometer, by exciting the samples using a Xe lamp coupled with a monochromator. Time-resolved PL measurement was collected by using fluorescence lifetime measurement system (HORIBA scientific). Fourier transform infrared (FTIR) spectra were recorded with a Nicolet 50 FTIR spectrometer at room temperature. The PL microscopic images of quantum dotterm·key patterns morphology were characterized by using a fluorescent microscope (Olympus BX51M). For pattern readout with a smartphone microscope, the portable microscope was linked to a small WiFi box by a USB line, which allows for the smartphone to control the microscope for real-time imaging (see Supplementary Fig. 7 ).

Quantum yield of quantum dotsterm·key The results are obtained by comparing integrated PL intensities using the standard procedure 55 , 56 . The quantum yield (QYs) of blue, green, and red emission quantum dotsterm·key were measured relative to Coumarin 480 (QY 99% in ethanol) with excitation at 350 nm, Coumarin 480 (QY 99% in ethanol) with excitation at 370 nm and rhodamine 6 G (QY 95% in ethanol) with excitation at 450 nm, respectively. Solutions of quantum dotsterm·key in toluene were optically matched at the excitation wavelength. Fluorescence spectra of quantum dotsterm·key and dye were taken under identical spectrometer conditions in triplicate and averaged. The optical density was kept below 0.06 at the λ max , and the integrated intensities of the emission spectra, corrected for differences in index of refraction and concentration, were used to calculate the quantum yields using the expression. 1 QY of quantum dotsterm·key = QY R × I I R × A R A × n 2 n R 2 where QY is the quantum yield, I is the measured integrated PL emission intensity, n is refractive index ( n = 1.496 for toluene; n = 1.361 for ethanol) and A is the optical density at the excitation wavelength.

Deep learningterm·key All the deep learningterm·key networks employed in our paper are based on the TensorFlow backend. The code used are run in software Pycharm 2017.3. To generate a large enough training set from theoretical image(s) for Alexnet model, a data augmentation procedure to the original synthetic image(s) is applied. For a typical process, one lower-left dot representing a security labelterm·key is captured as an image. Such a clear image was rotated by a step of 0.72° for 360° using an algorithm, producing a set of 500 training images. The input images were resized to 512 × 384 using pixel area relation for training. Plots of accuracy on the training and validation data sets over training epochs (from http://host:6006) can be found in the Supplementary Fig. 19 . The training images do not need to be stored and are not stored in this case; The storage requirement is mainly determined by the neural network itself, about 200 M Bytes here. For the AI technology we used, the learning process takes 2 h. The computer used for deep CPU is equipped with the CPU (Intel(R) Core(TM) i7–6700 CPU @3040 GHz), the GPU (NVIDIA GTX 1080), the RAM (32.0 GB), and HDD Capability (1 TB). The computer rated power is 350 W/h.

Registration and validation methodology We created a file named as gn ( n = 1, 2, 3, …) per security labelterm·key to store the corresponding 500 training images prior to the training process. The training images stored in the file gn are named as gn_000, gn_001, …, gn_500. Many files from these security labelsterm·key composed a database. After the 500 training images of a security labelterm·key were learnt by AI, their structural information was remembered and linked to the file name gn (e.g., g1). Then the training images will be deleted. When consumers randomly take a picture of a real security labelterm·key and sent it to the AI, the AI can automatically recall the accurately corresponding relationship and output the indexing name with a detailed match score. According our results, if the captured image from the end-user is clear enough, the match score of the image of true security labelterm·key is more than 99% (Fig. 4t ). On the other hand, if the image (from a fake label) has never been learned, the engine will give a lower match score (Fig. 4u ). The authenticationterm·key process takes about 2 s. Deep learningterm·key, as a black box that nobody actually knows how it works in details up to now, is an advantage for unclonable anti-counterfeitingterm·key technique because it is a tamper proof.

Supplementary information Supplementary Information Peer Review File Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2