Wafer‐Scale 200 mm Metal Oxide Infrared Metasurface with Tailored Differential Emissivity Response in the Atmospheric Windows

Metasurfaces with sub‐wavelength nanoscale features have emerged as a platform to achieve desirable electromagnetic responses. However, it remains technically challenging to fabricate metasurfaces in large size and at low cost for mass production. This work demonstrates a 200 mm wafer‐scale Al:ZnO metasurface coating based on deep‐UV lithography. The metasurfaces are targeted to achieve infrared (IR) reflectivity and emissivity characteristics at bandwidths across the two atmospheric windows in the IR spectrum. The wafers demonstrate a high uniformity of optical response with tailored reflectivity of around 50% at the 3–5 µm mid‐wave IR band and less than 10% at the 8–13 µm long‐wave IR band. This article furthermore shows that the design principle allows achieving a wide range of dual‐band reflectivity values using a single underlying materials stack, offering a versatile platform. The proposed approach is compatible with CMOS‐compatible mass‐production manufacturing and brings IR metasurface coatings closer to commercially relevant and scalable technology.

Here, we demonstrate a wafer-scale Al:ZnO metasurface fulfilling a dual-band optical response at MWIR and LWIR. Our design approach combines in-plane metasurface and outof-plane multilayer degrees of freedom, as well as the strong material dispersion of several selected materials, in order to achieve independent control in the two technologically important atmospheric windows with a spectrally flat response within Metasurfaces with sub-wavelength nanoscale features have emerged as a platform to achieve desirable electromagnetic responses. However, it remains technically challenging to fabricate metasurfaces in large size and at low cost for mass production. This work demonstrates a 200 mm wafer-scale Al:ZnO metasurface coating based on deep-UV lithography. The metasurfaces are targeted to achieve infrared (IR) reflectivity and emissivity characteristics at bandwidths across the two atmospheric windows in the IR spectrum. The wafers demonstrate a high uniformity of optical response with tailored reflectivity of around 50% at the 3-5 µm mid-wave IR band and less than 10% at the 8-13 µm long-wave IR band. This article furthermore shows that the design principle allows achieving a wide range of dual-band reflectivity values using a single underlying materials stack, offering a versatile platform. The proposed approach is compatible with CMOS-compatible mass-production manufacturing and brings IR metasurface coatings closer to commercially relevant and scalable technology. the bands. This approach allows covering a large range of parameter space using only a single basic device stack, which enables designer functionality in a single device and materials platform. Our approach is distinct from previous studies involving either multiple narrowband resonators [32,34,42,53] or intricate multi-resonator antenna designs [31,37,41,54] . We present the experimental realization of this device structure using stateof-the-art cleanroom nanofabrication on 200 mm silicon substrates. In our experimental studies, a homogeneous 200 mm wafer-scale metasurface has been fabricated with a reflectance of around 50% in MWIR and 10% in LWIR. Achieving such uniform functional optical coatings with unique features with optical metasurfaces requires large area fabrication with high fidelity. Fabrication methods capable of producing patterned metasurfaces that cover a large area and in high volume are still lacking. Most reported metasurfaces are fabricated through e-beam lithography (EBL), [31,34,41,42,[55][56][57] which is unsuitable for mass production and large area size, or standard photolithography, which is suitable for minimum feature sizes of micrometers. [46,50,51,54] Some recent works have adopted deep-ultraviolet (DUV) lithography for metasurface fabrication. [32,45,[58][59][60][61] . To achieve large-area device patterning, we make use of a DUV lithography scanner. In this method, a KrF UV laser at 248 nm wavelength is used to project a mask from a reticle onto the fabrication area. The wafer is subsequently scanned under the fixed projection system to stitch together many projections over the full size of the wafer. We show results demonstrating control over the design space using small area test patterns with variations in the metasurface geometry. We also present the full-wafer large area demonstrator showing uniform performance within the design specifications.
A particular feature of the design is the use of transparent conducting oxide Al-doped ZnO (AZO), which forms a promising material class for tailoring the response in the mid-IR through its strong dispersion of optical properties from dielectric to metallic regime across the bulk plasmon frequency. These are large-bandgap semiconductors with free electrons in the range of 10 19 -10 21 cm −3 . [62,63] Infrared metasurfaces based on conducting oxides have so far been reported to operate in the broadband range [18,64] and it is a challenge to form a multiband infrared metasurface using AZO. We report here a design using a novel 5-layer stack incorporating a thin Ti layer and the AZO metasurface as characteristic features. We explore the potential for extending the design space using the particular sets of materials and design stack of this work, which is beneficial to achieve a wide range of parameters with a single underlying design principle and opens up a platform for tailoring infrared responses that can be achieved using a single process workflow and technology qualification.

Design and Numerical Results
The main attribute of the proposed meta-surface is that it controls the infrared emissivity or equivalently, reflectivity, of an incident electromagnetic wave in two different spectral bands corresponding to the response of typical infrared imaging equipment covering the two atmospheric windows (3-5 µm labeled R MW , and 8-13 µm labeled R LW reflectivities). The proposed metasurface design process is based on the principles of operation of metallo-dielectric periodic structures but also exploits to a large extent the dispersion of material response with the wavelength in order to effectively control electro-optic (EO) emissivity with a broadband response within each of the two spectral windows of interest.
Achieving an arbitrary difference in the reflectance or emissivity response between the two spectral windows is a non-trivial task that is challenging to achieve using conventional materials and designs. Our design principle involves a multilayer Salisbury screen stack with a thin Ti-film, sensitively controlling the coupling between the top and bottom dielectric spacer. Strong dispersion properties of Ti between the MWIR and LWIR range provide a convenient spectrally selective mechanism. Strong material dispersion is also offered by the low-conductivity AZO material, which is intentionally tailored to achieve spectrally broad, resonant plasmonic features in the LWIR while having dielectric characteristics in the MWIR band.
The IR metasurface design is schematically shown in Figure 1a,b, showing a 3D view and corresponding cross-section of the stack. From bottom to top, the stack consists of an 80 nm aluminum (Al) back-reflector, a 1 µm thick silica (SiO 2 ) spacer, the thin titanium (Ti) layer, a 350 nm thick alumina (Al 2 O 3 ) layer and the 150 nm thick metasurface made of AZO periodically arranged square features. The critical layer in the overall design is the AZO metasurface. The metasurface fulfills two essential roles: first, the metasurface is used to create an artificial impedance that is controlled by changing the dimensions of the square features and their gap, and second, we take advantage of the partial absorption of radiation due to the low conductivity of the AZO.
The percentage of reflection of each spectral band determines the sequence in which the remainder of the layers are stacked (see Section S2, Supporting Information, for the loss contribution from each individual layer). Aiming to absorb the incident electromagnetic wave by more than 90% in the spectral band 8-13 µm (R LW < 10%), we must match the upper layer's impedance to the impedance of free space (377 Ω). This matching allows the wave to propagate within the structure. The complex refractive index of SiO 2 presents an abrupt fluctuation at a wavelength of around 9.5 µm, which results in the partial absorption of the radiation and affects the impedance matching between the layers. Thus, another part of the radiation is reflected. To mitigate the dispersion of SiO 2 , we place it directly above the ground plane. Then, we place a very thin layer of Ti where in the 3-5 µm spectral band, the values of the complex refractive index n are about 4.5 and k assumes values ranging from 5.9-9.2 while in the spectral band 8-13 µm, n covers values from 5.8 to 9.1 and k gets values from 14.8 to 21.3. Τhe Ti thickness, therefore, critically affects the EO response in the two bands (see Section S3, Supporting Information, for more details). Above the Ti-film, we place a second, thinner spacer for which Al 2 O 3 was selected as the material most compatible with the fabrication process. Other choices of dielectrics are discussed further below.
For the metasurface, we choose AZO with a carrier density of 6 × 10 20 cm −3 , consistent with our previous works and the highest carrier density achieved in our experimental system. [63] The metasurface (Figure 1b) is designed as a square array of AZO features with varying sizes and gap spacing between adjacent structures. Figure 1c shows the simulated optical response of the complete design for various feature sizes from 0.94 to 1.74 µm and a fixed gap size of 0.2 µm. Here, a Ti thickness of 5 nm was used. For the 1.34 µm structure with a 0.2 µm gap, a reflection of around 50% is achieved in the MWIR range (3-5 µm), whilst a lower reflection of ≈10% is achieved in the LWIR range (8-13 µm). Τhe surface impedance in the MWIR range (3-5 µm) is around 1600 Ω while in the LWIR range (8-13 µm), where the reflection is around 10%, the impedance is around 390 Ω. If there is a mismatch between the impedance of free space and the surface impedance, the absorption/reflection depends on the upper layer of the structure. The wave penetrates to a greater depth when there is a match between the aforementioned impedances-the absorption peak in the LWIR range shifts along with the increase in feature size. We observe a similar effect on optical response in Figure 1d when tuning the gap from 0.20 to 0.30 µm, with a fixed feature size of 1.34 µm.
Further investigation of the dependence of the reflectivity on the thickness of both SiO 2 and Al 2 O 3 dielectric layers is presented in Figure S3, Supporting Information. Both dielectrics are important in setting up maximum reflectivity in the MWIR range. The strong absorption feature in the LWIR band (i.e., low reflectivity in Figure 1c,d) can be attributed to the plasmonic resonance effects of the AZO metasurface. Figure 1e,f shows the electric near-field enhancements |E|/|E 0 | around the nanostructured surface for selected wavelengths of 4 and 10.5 µm, both in isometric 3D plots, as well as in cross-sectional maps through the center of the AZO feature (logarithmic scale).
The combined metal-insulator-metal (MIM) stack of the AZO metasurface/Al 2 O 3 /Ti structure forms a metamaterial-perfect absorber [8] configuration with a pronounced optical resonance in the LWIR. The near-field map at 10.5 µm wavelength indeed shows one order local field enhancement around the structure characteristic for the dipolar (half-wavelength) resonance mode of the AZO feature. In comparison, at 4 µm wavelength, the AZO metasurface does not show any significant resonance effects, furthermore, fields are seen to propagate through the semitransparent Ti-film into the underlying SiO 2 layer.

Fabrication and Optical Characterization of Metasurfaces
Our experimental studies report results for two full-size 200 mm wafer demonstrators. In the first wafer fabrication run, various DUV doses from 14 to 28 mJ cm −2 were tested to achieve the designed feature size. In the second wafer fabrication run, the optimized dose of 20 mJ cm −2 has been applied uniformly to the entire wafer. The Ti thickness has been increased slightly from 5 to 10 nm for this run to protect against variability due to the lack of optimum control of thickness. Wafers were prepared as detailed in the Experimental Section.
The variable dose wafer is presented in Figure 2a  The dose of 20 mJ cm −2 was subsequently used for the whole wafer metasurface with homogeneous coverage. Figure 2c shows experimental spectra of metasurfaces with L from 0.94 to 1.74 µm and d = 0.2 µm. The spectra of an unpatterned AZO film with an identical stack structure are also presented for comparison. All arrays show a consistent MWIR reflectivity of around 60% but varying reflectivity in the LWIR range. With the increase in size L, the peak around 9 µm becomes narrower and lower and thus the smaller feature size is preferred for lower overall reflectivity in the LWIR range. Compared with the unpatterned AZO film, reflections of metasurfaces in the LWIR are significantly reduced owing to their resonantly enhanced absorptions around the resonance wavelength of the structured metasurface. In agreement with simulated results (Figure 1c), experimental spectra show a maximum reflectivity in the MWIR band centered around 3.5 µm wavelength, which is attributed to the variation in skin depth of the 5 nm Titanium film.
The variation of response with the metasurface gap size is presented in Figure 2d for a fixed L = 1.34 µm. The dependence on gap size in this limited range is quite small but confirms that the smaller gap sizes are the most appropriate for achieving the favored response, that is, a high reflectivity in the MWIR and low reflectivity in LWIR. In all experimental curves, we note that the reflectivity at longer wavelengths from 13-20 µm does not reach as high as in the simulations of Figure 1c,d for the patterned metasurfaces, while the thin-film stack only shows a strong reduction above 18 µm. We attribute this difference to a metasurface-induced additional absorption in experiments, which is not fully captured by our model. As the region lies outside of the two bands of interest, we do not further consider this discrepancy in our analysis.
The dual-band reflectivities in MWIR and LWIR, R MW and R LW , are calculated by averaging the reflection over the respective spectral bands. For direct comparison with the results presented further below, the individual wavelengths are weighted against the response function of the reflectometer used in the directional hemispherical reflective measurements, as discussed in detail in the Experimental Section. This response function is very similar to the response function of commonly used infrared sensors and imagers used in realworld applications. Extracted values are R MW = 55.7% ± 0.3% independent of gap size d or feature size L. The long wave response is significantly influenced by the feature size though and R LW decreases from 18.1% to 5.6% with decreasing feature size.
To demonstrate a wafer-scale large metasurface, a design of L = 1.34 µm and d = 0.2 µm and a fixed exposure dose of 20 mJ cm −2 was chosen to process a 200 mm substrate with a uniform metasurface covering its entire surface.  SEM image of the fabricated metasurface. The feature sizes of the metasurfaces at different regions over the substrate have a very narrow distribution of L = 1.30 ± 0.05 µm, which is attributed to the high uniformity process tools used, for example, the DUV scanner and ion beam etcher. The cross-section SEM image is provided in Section S5, Supporting Information. The reflection spectra at different regions are presented in Figure 3c. High consistency is seen in the 8-13 µm region, whilst some variation is noticeable in the range between 4 to 8 µm. Reflectivities R MW and R LW were again extracted. Variations are low with R MW at 51.3 ± 0.9% and R LW at around 11.0 ± 0.4%, indicating an excellent performance uniformity from a uniformly fabricated large-scale metasurface. Both values are slightly inferior to the maximum values achieved on the variable dose wafer, due to trade-offs in manufacturing risk against performance, specifically the slightly thicker Ti-film of 10 nm instead of 5 nm. We note that the slightly thicker film gives a flatter response over the bands, which could be a performance metric, as discussed in the Discussion section.

Directional Hemispherical Reflectance Analysis
To further characterize the response of the devices, a number of characterizations were done using accepted standardized methods. Directional hemispherical reflectance (DHR) measurements were taken for both wafers at 20° and 60° using a commercial multi-band reflectometer (SOC-410). Figure 4a shows the LWIR thermal image of the first metasurface wafer with variable exposure doses and for the 5 nm Ti-film. The exposure dose is increasing from the right side to the left side as columns in the presented orientation of the wafer. In the thermal image, the substrate is homogeneously bright in most areas except the four columns from the right in dark contrast. As presented in Section S4, Supporting Information, the right four columns are underexposed, and thus, the metasurface is not well-defined in those regions. Beyond the right four columns, column borders are barely seen, indicating an excellent overlapping control to almost eliminate stitching artifacts in the tile-like DUV exposure regions. DHR measurements were performed in seven positions marked by colored spots and numbered from 1-7, with DHR values presented in Figure 4b. The DHR values of positions 4-7 in the 3-5 µm band (R MW ) consistently lie around 55% and 45% for angles of 20° and 60°, respectively, while DHR values of underexposed positions 1-3 are only slightly higher. In the 8-13 µm band (R LW ), DHR values of sufficiently exposed positions 4-7 are around 15% and 21% for angles of 20° and 60°, respectively, while underexposed positions give much higher values at both angles. This higher reflectance is attributed to the effect of the improperly formed metasurface, which resembles the response of the continuous thin film. The DHR values of 0° can be calculated from FTIR measurements with calculations detailed in the Experimental Section and Section S6, Supporting Information, under the assumption that scattering outside of the collection angle (numerical aperture 0.58) of our FTIR microscope is small. Figure 4c compares the FTIR and multi-band reflectometer measurements (red, blue symbols) and simulated values (lines see simulated spectra in Section S7, Supporting Information). The FTIR and reflectometer measurements show an excellent agreement with the simulated results. Figure 4d shows the thermal image of the second wafer-scale metasurface, with a fixed exposure dose of 20 mJ cm −2 and 10 nm Ti thickness, with corresponding DHR values shown in Figure 4e,f. A homogeneous brightness is seen over the whole substrate with no tile borders of the DUV wafer scanner exposure identified, which indicates that a highly uniform optical metasurface is achieved over the entire diameter of 200 mm. DHR measurements were performed at four different positions, which are marked by the colored dots in Figure 4d and presented in Figure 4e. The measured DHR values consistently lie around 46% and 42% for angles of 20° and 60°, respectively, in the 3-5 µm (R MW ) band and 13% and 20% for angles of 20° and 60°, respectively, in the 8-13 µm (R LW ) band. Figure 4f again compares DHR values measured from multi-band reflectometer, FTIR, and simulated values (with simulated spectra in Section S7, Supporting Information). These results are consistent with our FTIR measurements ( Figure 3) which indicate R MW around 50% and R LW around 10% for this wafer-scale metasurface. Like those for the 5 nm Ti substrate, the R MW and R LW for the 10 nm Ti substrate are also in good agreement with simulations.

Exploration of Design Space: Random Designs
In the previous sections, we demonstrated a specific design aimed at achieving a low reflection or high emissivity in the thermal LWIR band combined with a high reflectivity contrast in the MWIR band. The 5-layer stack was conceived from  rational design principles inspired by concepts taken from the microwave domain. It is particularly useful to assess whether the same basic design allows to achieve a variation in the dualband EO response by simply altering layer thicknesses and in-plane metasurface parameters. Such a flexible platform for designing EO responses using a set of common materials and fabrication methods would hold significant advantages with respect to manufacturability and technology qualification for specific application domains such as space or defense.
The design space for the 5-layer metasurface stack has six freely tunable parameters (four layer thicknesses and two metasurface geometric parameters). To explore this space, we first evaluated 3000 random designs with parameters set by the physical constraints in layer thickness and patterning capabilities of the tools used in this work. We evaluated thicknesses of the bottom and top dielectric of 500-1500 nm and 200-1000 nm respectively, a Ti-film thickness ranging from 4-150 nm, and AZO thickness from 0-400 nm. For the in-plane metasurface, we chose square features with individual sizes of 300-1300 nm and gaps ranging from 200-500 nm. Results are given by the red open circles in Figure 5a, which show the joint distribution of mid-wave and long-wave IR reflectivity values R MW and R LW . The basic 5-layer perfect metamaterial absorber (PMA) design allows for structures covering a considerable range of the parameter space within the physically realistic parameter range. We furthermore note that the experimental demonstrator metasurface (PMA) device with R MW = 60%, R LW < 10% corresponds to the bottom right corner of the parameter space, as also shown in panel (i) of Figure 5c, indicating that this is indeed a nontrivial example that pushes the boundaries of the conceptual design.
The benefit of metasurface patterning is illustrated by comparing it with a set of equivalent unpatterned thin-film stacks without metasurface patterning. The results for unpatterned devices are given by the green open circles in Figure 5a and show a very limited coverage of the design space limited to the top right of the range with both high R MW and R LW , which is attributed to the much-reduced tunability of the surface impedance in this case. The high LWIR reflectivity of the thin-film design was already seen experimentally in Figure 2. Obtaining high reflectivities in both spectral bands may indeed be considered a trivial challenge that can be fulfilled with many conventional thin-film or metallic reflector designs.
Next, to demonstrate the benefit of the Ti reflection layer, we compare it with a set of calculations without Ti-film, as indicated by the yellow open circles in Figure 5a. Without the Ti-film, the design space is essentially reduced to that of a conventional Salisbury screen with a single dielectric spacer consisting partially of SiO 2 and Al 2 O 3 . For this type of design, higher values of R MW are achieved; however, at the cost of a reduced parameter space coverage.
A performance figure of importance in some applications is the flatness of the response over the spectral window, as often a uniform response within the selected band is desired. In Figure 5b, we plot the standard deviation of the reflectance (σ MW , σ LW ) for the same datasets. The standard deviation is defined from the absolute reflectivity in the bands, that is, without any spectral weighting. The solutions cover a range of  standard deviations from 0 up to 50%, where the higher values indicate poor flatness. Typically, flatness values below 10% would correspond to a reasonably smooth spectral response within each band. Generally, the distribution of flatness in the 8-13 µm (LWIR) is better than in the 3-5 µm (MWIR) range, and the devices without Ti-film show a less flat behavior than those with Ti-film. The thin-film stacks are very flat in the LWIR response because they act as a simple mirror in this range.

Deep Learning Neural Network Inverse Design
To achieve a more controlled inverse design of the dual-band reflectors, we made use of an artificial neural network (ANN) approach. We follow here the general approach presented in our previous work [65] using a tandem encoder-decoder scheme based on the ResNet ANN to test the inverse design of a range of devices covering the (R MW , R LW ) space. As the tandem ANN is trained on the inverse problem, it directly outputs a design structure for any combination of target objectives. Flat spectral bands with minimized standard deviations (σ MW , σ LW ) are achieved by using the full spectrum within the two bands as input for the target response. Details of the model are presented in Section S8, Supporting Information.
Results for the ANN generated designs are shown in Figure 5a,b as the solid blue dots. These solutions were generated using 20 equally spaced target values each for R LW and R MW (i.e. 2% steps in both) as input parameters, thus resulting in a grid of 400 target combinations. It is seen that the inferred designs do not project onto this grid, but rather the solutions trace out a particular pattern in parameter space, which is characterized by a band of closely lying solutions with larger gaps in between. The characteristic contour of inferred solutions is a combination of two main features, namely i) that physical solutions with good spectral flatness can be found primarily on this contour, and ii) the tandem ANN has a tendency to collapse onto a single set of solutions in parameter space. The latter is a well-known feature of this type of ANN, which has been critically discussed elsewhere. [66] Importantly, the solutions inferred by the ANN inverse design show a significantly improved flatness as compared to the randomly generated dataset, as shown in Figure 5b. Figure 5c shows selected spectra from the solutions inferred by the ANN over four different ranges of values for R LW , as shown in panels (i-iv), and up to four for R MW , as indicated by the different colored spectra in each panel, where each spectrum represents a design with the flattest spectral response within this range. Overall, the solutions appear smooth over the two ranges of interest, where we point out that the bending of the spectral response above 12 µm is a consequence of the dispersion of the Al 2 O 3 . Solutions with larger bandwidth in the LWIR spectral range can be obtained by substituting a material with  lower dispersion in this range, such as Ge, as discussed in Section S10, Supporting Information.

Comparison with State-of-the-Art
Our experimental and computational study has focused on achieving a platform for tailoring the infrared response in the two atmospheric bands at 3-5 µm and 8-13 µm. As briefly discussed in the Introduction, other works have investigated metasurface-based approaches targeting specific functionalities. A more general overview of single-frequency and broad-band metasurface perfect absorbers has been the subject of several high-quality reviews. [8,21] Table 1 shows the state-of-the-art literature on dual-band infrared functional coatings, exploiting metasurface design degrees of freedom. Many of these focus on the use of narrowband antenna structures made from metals Al, Ag, and Au, where either multiple antennas are combined or a single antenna is shaped to provide multiple resonances. Notable is also the use of complementary structures, such as hole arrays, to achieve a frequency selective surface with resonance absorption in the visible and near-infrared combined with high reflectivity over a large spectral band. [51] In order to achieve additional radio-frequency (RF) or microwave functionalities, multi-scale patterning is employed, where micrometerscale and millimeter-scale structural features are sufficiently far apart in size to allow their separate definition without cross-talk. [50] Our generalized approach is distinct from specific applications that focus on achieving high reflectivity in the two atmospheric windows combined with one or more resonant absorption features and/or RF functionality, [15,46,50,51] as well as from those aiming to synthesize multiple high optical absorption/emissivity through embedding multiple resonators [32,34,42,53] or multi-resonance antennas. [31,37,41,54] Our approach specifically addresses the problem of achieving arbitrary average reflectivities with a flat spectral response, which is not easily obtained using multi-resonance synthesis.
The overview in Table 1 also includes the minimum dimension size and device area that is explicitly demonstrated in the respective works. A wide variety in feature sizes from 100 nm to several micrometers requires techniques ranging from EBL to photolithography, the latter of which can be more readily implemented at large surface areas. Our experimental demonstration involves minimum feature sizes of 200 nm and DUV scanner lithography, which fills a critical gap in fabrication capabilities between EBL and photolithography.

Conclusion
In conclusion, we have demonstrated a 200 mm wafer-scale infrared metasurface platform for the infrared using a CMOScompatible fabrication method involving DUV lithography. By forming a 5-layer stack with the AZO metasurface on top, the device realizes distinctive optical responses in the 3-5 µm MWIR and 8-13 µm LWIR atmospheric windows, with a flat optical response within the bands. A homogeneous waferscale metasurface has been demonstrated with a reflectance of around 50% in MWIR and 10% in LWIR. Numerical simulations and deep learning inverse design show that the general design principle of the device allows for a wide range of reflectivity within the atmospheric windows to be realized using a common platform of materials and fabrication workflow. There is a significant emerging research trend for utilizing atmospheric windows in new ways requiring different combinations of reflectivity and emissivity functionalities. We show here that metasurfaces can reach areas of this parameter space that are inaccessible through conventional materials. This study demonstrates a more generalized design and fabrication capability that is needed to adapt it to any application in the MWIR and LWIR atmospheric transparent windows. A general platform for achieving a wide range of reflectivity values in the infrared offers significant advantages in manufacturability, performance testing, and device qualification for a range of applications.

Experimental Section
Fabrication of Metal Oxide IR Metasurfaces: Metal oxide IR metasurface was fabricated onto 8-in. (200 mm) silicon substrates. The structure consisted of a sputtered aluminum film of 80 nm thickness as an optical back-reflector and then a 1000 nm SiO 2 spacer using the Bühler Helios sputtering system, followed by a thin titanium film of 5 or 10 nm using the Angstrom EVOVAC sputtering system. A 300 nm Al 2 O 3 was deposited and followed by 150 nm Al:ZnO using Oxford Instruments FlexAl ALD system. The Al 2 O 3 was grown at 175 °C using trimethylaluminum (TMA) and water. The Al:ZnO was grown at 250 °C using TMA and diethyl zinc (DEZ) and water with a TMA:DEZ cycle ratio of 4% (detailed in previous works) [18,63] and carrier concentration was measured to be 6 × 10 20 cm −3 . The AZO was subsequently patterned using Nikon NSR-S204B 248 nm DUV system with KrF M91Y photoresist (680 nm thickness after development). For the 5 nm Ti wafer, various exposure doses were varied from 14 to 24 mJ cm −2 with a step of 2 mJ cm −2 . For the 10 nm Ti wafer, the exposure dose of 20 mJ cm −2 was used over the whole substrate. The AZO was subsequently etched using Oxford Instrument IonFab 300 plus ion beam etch (IBE) system with a pure Ar physical etch. [67] The etching rates were measured to be 6 nm min −1 for AZO due to the micro-loading effect (small opening). The resist was subsequently stripped by an ultrasonic bath in NMP at 60 °C for about 11 h to ensure the removal of AZO redeposition during the IBE etch. The resist features and AZO features were planarly checked using Hitachi S8840 critical dimension SEM with cross-section checked by JEOL JSM-7500F.
Optical Instrumentation: The IR reflectance was measured in the range of 2 to 20 µm using an FTIR setup consisting of a Thermo-Nicolet Nexus 670 with a Continuum microscope, far-IR light source, KBr beam splitter, and nitrogen-cooled DTGS detector with a spot size of 100 × 100 µm 2 . For all measurements, an 80 nm aluminum-coated SiO 2 /Si substrate was used as a reference mirror.
In diffuse hemispherical reflectance experiments, the spectrally averaged total reflectivity R was measured, which was weighted with the SOC-410 DHR reflectometer response for each detection band. For consistency, all the reported R MW and R LW were chosen to weigh with these response functions, in order to allow direct comparison between simulations and experiments. The reflectivities of the two spectral bands were defined by R MW and R LW for the 3-5 and 8-13 µm ranges, respectively, using: where R(λ) is the normalized reflectance, A(λ) is the absorption response of SOC-410 DHR reflectometer, [68] and λ is the wavelength. DHR was measured at Malvern Optical Ltd. DHR was taken using a SOC-410 instrument. The SOC-410 collected data in six wavebands across the IR with a primary focus on 3-5 and 8-13 µm (corresponding to MWIR and LWIR, respectively). DHR data were collected at two polar angles of 20° and 60° relative to the surface normal.
Numerical Simulation: For numerical simulations of metal oxide IR metasurface in the initial design, CST Microwave Studio was used, based on the finite integration technique. Although CST includes various electromagnetic solvers, the finite-element frequency-domain (FD) solver was chosen, as it could accurately calculate the electromagnetic response of a periodic structure for an ultra-broad spectral band. Furthermore, it was the only solver that an oblique angle of incidence could be considered. An additional feature of the FD solver includes periodic boundary conditions that can be applied to arbitrarily shaped unit cells, assuming infinite size structures and introducing the phase shift between elements.
Parameter Studies and Neural Network Design: For numerical simulations in the parameter studies, Ansys Lumerical FDTD Solutions was used. A wide range of spectra were generated for different device parameters. The deep learning neural network was a tandem network consisting of a pre-trained forward network and an inverse network. Each of the networks was implemented using an encoder-decoder layout. Details of the network architecture and hyperparameters are given in Section S8, Supporting Information.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.