Profiling a Low Emissivity Glass Coating with ToF‐SIMS and Machine Learning

Characterization of multilayer coatings in 3D presents many challenges, as composition can change by area and by depth. Compositional characteristics of the interior of multilayer coatings emerge during analysis, so are frequently discovered only through exacting retrospective investigations. Time‐of‐flight secondary ion mass spectrometry (ToF‐SIMS) can be used to elucidate such complex systems; however, data analysis is a challenge. In this work a detail presentation is done of 3D chemical characterization of a low emissivity (low‐E) double silver coating on glass using ToF‐SIMS and machine learning. An unsupervised machine learning technique, the self‐organizing map with relational perspective mapping, is used to visualize the chemical similarity between different layers of the low‐E film. Repeating layers are easily identified at the single‐voxel level, based on their entire mass spectra, and are classified as chemically indistinguishable. All major film components are identified, including the use of SnO2 as a dielectric, ZnO seeding layers, TiOx blocking layers, a Zn base layer, and a TiOx topcoat. The thin optically active silver layers are examined in detail, demonstrating subtle chemical changes with depth. This technique provides excellent insight into manufacturing processes and production challenges and has excellent potential in forensic applications.


Introduction
Windows are generally considered the least energy efficient component of buildings, as they do not block the sun's heat entering a building, or prevent the warmth generated inside from being DOI: 10.1002/admi.202300645lost. [1]3] The emissivity of an object is a ratio that compares the heat it emits to that of a blackbody, which has an emissivity of 100%.A perfect reflector has an emissivity of 0%.Silver-based low-E coatings typically have emissivity values in the range 2-8%. [1]Low-E coatings are designed to have high infrared reflectivity, while allowing the transmission of visible wavelengths.][5][6] The total thickness of all silver layers must be less than 40 nm thick to allow for optical transmittance. [1,7,8]ow-E coatings are complex stacks of different materials, each serving a specific purpose, which aim to maximize both IR reflection and transmission of visible light.[4] Most coatings use double or triple silver layers to improve the selectivity between IR and visible wavelengths. [1]The electrical properties of the silver layers are highly dependent on their purity, crystalline structure, grain size, grain boundaries, and surface roughness. [1,9]][11] Blocker layers act to protect the silver from oxidation that leads to poorer performance in the IR spectrum. [10]These comprise of a metallic region at the silver interface, that becomes more oxidized toward the surface of the coating. [1,3,5]ielectric layers are non-adsorbing and act as anti-reflective coatings that enhance the energy transmitted in the visible region.[4] Common materials include bismuth oxide, tin oxide, zinc oxide, titanium dioxide, silicon dioxide, and silicon nitride. [1,5]Other layers include: an overcoat/topcoat that provides mechanical protection; [2,3] and the base layer that adheres to the glass surface and protects silver from alkaline ions (Na + ) that migrate from glass into the coating. [5,12]erforming a chemical characterization of a complex optical multilayer film stack is a challenging task.The position of materials and the interfacial properties need to be faithfully examined without altering them.One way to achieve this is by depth profiling through the low-E coating and collecting data at various depths.Techniques, such as X-ray photoelectron spectroscopy (XPS), [5,6,12] Auger electron spectroscopy, [4,6,10,13] and ToF-SIMS have been previously shown to be suitable for this task. [5,12,14,15]sing prior knowledge of the film, each of the main constituents of the film can be identified by a characteristic element or ion.While these techniques generate 4D datasets (three spatial dimensions and a single spectral dimension), the data are commonly displayed as the amount or intensity of a given material as a function of depth through the coating.That is, the x and y spatial dimensions are summed to reduce the data to 2D.While this can be effective, only a small selection (often to the analyst's discretion) of the collected data is shown, therefore creating experimenter bias.Reducing the dataset in this manner also makes it difficult to comprehend subtle or complicated relationships within the dataset as only a small section is displayed.Without prior knowledge of the sample, this can be a challenging and time consuming task.
ToF-SIMS depth profile data are particularly complex, as each voxel contains a complete mass spectrum.Manual analysis of ToF-SIMS depth profile data at the single-voxel level is a timeconsuming process due to the considerable number of peaks associated with each layer.It also generates bias due to subjective selection of peaks to be included or excluded.
Machine learning methods offer a powerful alternative to manual analysis, as they can analyze each voxel in the depth profile based on their complete mass spectra.We have previously demonstrated the value of an unsupervised machine learning method on ToF-SIMS data using self-organizing maps with relational perspective mapping (SOM-RPM).SOM-RPM has proven valuable for interpreting 3D data but has so far only been applied to layered polymer systems. [16,17]SOM-RPM models chemical similarities by tagging each pixel or voxel with a color.The pixels are then mapped back to their original spatial positions to visualize the model output as a similarity map, where their color reflects their chemical (mass spectral) similarity to all other pixels within the image.By utilizing SOM-RPM to visualize the large volume of data collected, 3D regions of the film can be uniquely distinguished at the single-voxel level.
In this work, we present a workflow for generation and interpretation of a complex ToF-SIMS depth profile of a low-E glass coating.By considering the entire 4D dataset and the accompanying spatial-spectral information, we can interpret critical functional and special purpose layers, such as seeding and blocking layers and obtain a detailed analysis of interfaces at various depths within the optical stack.

Results and Discussion
ToF-SIMS data were collected in both positive and negative polarity, to generate a complete picture of the sample.Secondary ion formation is strongly influenced by electron and/or cation exchange processes between departing species and the surface, which is known as the matrix effect. [18]Consequently, the yield of elemental ions can vary by several orders of magnitude according to their chemical environment.Monatomic ions of the electropositive metals produce a greater signal intensity in positive polarity, whereas their oxide ions yield greater intensity in negative polarity. [18]econdary ion yields are also affected by the use of reactive sputter ions during depth profiling, with O 2 + and Cs + enhancing those of positive and negative ions, respectively. [18]Sample mixing during depth profiling can be affected by the choice of sputter ion.As Cs + has a larger mass and is an atomic ion, it will induce more interlayer mixing that O 2 + ; a smaller molecular sputter source. [18]As each beam sputters at a different rate, the times taken to penetrate the coating in the two profiles differ; likewise, each beam sputters the different constituent layers of the film at different rates.The apparent thickness of the layers may not be an accurate representation of depth and instead reflects layer hardness.However, within a polarity, repeating layer thickness can be directly compared.
Figure 1 offers a traditional view of ToF-SIMS depth profiling data, as a 1D plot of the intensities of selected spectral peaks against sputter time.When comparing the two ion modes, the positive spectra, collected with a O 2 + sputter beam (Figure 1 top) indicates that the Ti + was close to coincident with the Ag + layer, while in the negative spectra, using a Cs + sputter beam (Figure 1 bottom), indicates the TiO − peak has shifted toward the center of the Ag peak and may be responsible for the shoulder on the right-hand side of the Ag − peak.The ZnO 2 − ion peak also overlaps much more with the Ag − peak than the Zn + peak with the Ag + peak.In the positive profile, the four Sn + bands have similar intensities and widths, whereas in the negative profile, the maximum in the first Sn − band is much greater than the other three.In contrast, the maximum in the first SnO 2 − band is much lower than for the three subsequent bands.The central structure (between the double silver regions) is very similar in both profiles; however, the relative intensities are different due to the matrix effect.These differences between profiles are strongly influenced by a sometimes complicated combination of sputtering characteristics and ion yield as the sputter ion is changed.3D renders of each of the ions shown in Figure 1 has been included in Figures S1 and S2 (Supporting Information) for comparison with the SOM-RPM models.
This data presentation illustrates the basic film structure, but it is difficult to distinguish whether repeated layers are identical, and where each layer ends and the following layer starts.While this view of the data can provide helpful insights if the structure of the sample in known, it relies on analyst expertise to identify which of the many peaks present in each spectrum are pertinent to characterizing the coating composition.This raises questions, such as 1.How many peaks should be considered?2. Which peaks should be considered?

Are the peaks selected faithful representations of the underlying composition?
There is no correct answer, and no metric with which to decide the best outcome.Furthermore, the complexity of ToF-SIMS data makes peak selection a time-consuming process and risks adding human bias into the dataset.This is particularly true when working with an unknown sample.The SOM-RPM approach, which examines all mass peaks within each voxel and provides an unsupervised classification based on similarity, removes user bias by eliminating the above questions.The entire 3D depth profile can be visualized in an intuitive way, using all the mass spectral peaks for classification.This approach can be used to summarize the global topology of the data and the key components for each region quickly.Upon deeper analysis, it can yield an extraordinary amount of chemical and spatial information for the sample.We propose this SOM-RPM technique be used to reduce the time and potential bias of selecting key ion peaks manually and to be used as a "first pass" to allow an expert to view the sample as a whole, before making interpretation.
Figure 2 illustrates the depth profile of a low-E double silver glass coating collected in positive ion mode using an O 2 + sputter source, analyzed using the SOM-RPM workflow.Figure 2A shows the toroidal SOM of the computational model, and the color distribution across all 64 neurons.It shows that approximately half of the neurons are devoted to defining both the substrate (burgundy) and layer +13 (emerald green).Other model sizes ranging for 16-100 neurons were explored with the results shown in Figure S3 (Supporting Information).And 64 neurons were chosen as the best compromise between computational time and mapping of subtle features.Figure 2B is the 3D similarity map, where each voxel is assigned to a neuron within the SOM and given its color.Regions of the similarity map can then be discriminated by voxels being assigned to different neurons within the SOM, highlighting the complex layer structure of the low-E film.The film is homogenous in x and y across each layer with no pinhole defects or transported contaminants.Repeating layers can be easily identified as repeating colors.Figure 2C shows the complexity of the structure in the silver region, with an accompanying mass spectrum for each layer presented in Figure 2D.In the positive ion model, there are no repeating layers in the upper silver region.From the surface there is a gradual increase in the intensity of peaks assigned to silver ions (Ag + , Ag 2 + , and isotopes) from layer +8 to a maximum at layer +11, then a decrease in layer +12, with negligible amounts present in the spectrum of layer +13.Peaks attributed to titanium (Ti + , TiO + , and isotopes), used as a capping layer to protect the silver coating are present in the spectra of layers +7 to +10.Zinc (Zn + , Zn 2 + , and isotopes) is present in layers +12 and +13, serving to seed the silver growth.
Figure 3 presents a depth profile collected nearby the profile illustrated in Figure 2. Data were collected using negative ion mode and Cs + sputtering and analyzed using the SOM-RPM workflow.Figure 3A shows the color distribution across the toroidal SOM, with a distribution similar to that in Figure 2A. Figure 3B highlights the layer structure in the negative ion similarity map, like that in Figure 2B, with some key differences: notably, the difference in the structure of the silver region is highlighted in Figure 3C S1 and S2 (Supporting Information).Presenting the ToF-SIMS data in this manner highlights repeated and, therefore, chemically identical layers.The position, thickness, and coloring of each layer provides key information and should be considered for each region as well as the chemical information.As the two sputter sources remove material at different rates, the positive and negative ion data each required their own SOM-RPM model, where the coloring of each model is unrelated.Likewise, the two models yield a different number of layers, the numbers of which do not necessary correspond, i.e., layer +14 = layer −15.
The surface layer (+SUR) contains a high proportion of contaminants, including Na + and K + .The peaks associated with these ions decrease in intensity with depth, being relegated from the ten most highly weighted peaks from layer +4 onwards.The titanium peaks (Ti + , TiO + , and isotopes) increase in intensity through the initial layers of the depth profile, reaching a maximum at layer +4 and not decreasing until layer +10.The similar coloring and adjacent position in the neural network of the first five layers are indicative of minute changes in composition.Layer +4 of the model is significantly thicker than the surrounding layers.Layers +3, +5 and +6 all exhibit some degree of interfacial mixing.Layer +5 introduces tin (Sn + and isotopes), which increases in intensity into layer +6, but is not present in layer +7.Zinc (Zn + and isotopes) appear in layers +7 and +8, but not in the silver regions, and so may be a protective capping layer.Layers +7 to +12 all comprise of two thin layers at similar depths, which are associated with the double silver regions and are shown in greater detail in Figure 2C,D).Layer +14 is very thin and only ever borders Layers +13 and +15.Its color is also between those of layers +13 and +15, again probably as a result of interfacial mixing.Layer +15 is a tin layer containing Sn + , SnO + , and isotope fragments.Layer +16 is very similar to layer +13, containing mostly zinc fragments, but without the inclusion of a Sn+ peak.Layer +17 is an interfacial fragment containing Zn + , Sn + , and ZnSnO + peaks.Impurities from the glass substrate (Na + , Ca + , and K + ), which were likely transported during annealing, are present within layers +18 and +19 together with Zn + .Zinc was probably used as a base layer to improve adhesion between the coating and the glass substrate, layer +SUB.
Figure 5 illustrates a central xz slice of the negative-ion 3D depth profile and divides the total coating thickness into 18 regions of interest.The 10 most intense peaks for each ROI, and the 10 most highly weighted peaks have been chemically assigned and are reported in Tables S3 and S4 (Supporting Information).
The surface layer (−SUR) contains a high contribution from peaks originating from hydrocarbon impurities, which are less mobile than the Na + and K + ions detected in positive ion mode and are only present at the surface and layer −2.Titanium and its oxides are present in layers −2 to −4, reaching a maximum intensity in layer −3.ZnO − or ZnO 2 − are amongst the top ten most intense peaks in every region, apart from the surface and substrate, this is likely due to surface mixing from the Cs + sputter beam.Examining the SOM weights reveals that, zinc oxide peaks are only key in distinguishing layers −5 to −8 and −12 to −15.Aside from layer −7, these are all interfacial layers.Tin and its oxides are present in layers −4 to −6 and −14 to −16, with intensity peaking in layer −5 and −16, respectively.This is consistent with the structure seen in positive ion mode.Layers −9 to −13 correspond to the double silver regions, as seen in Figure 3.While O − is a major peak in the silver layers, it is not found bound to silver; hence, the blocker layers appear to have maintained the purity of the silver.The O − may have originated from the titanium or zinc layers and is present due to interlayer mixing.Major constituents of layer −17, includes zinc, oxygen, and silicon, suggesting that (as in positive polarity) zinc is used as a base layer.The glass substrate layer (−SUB) is characterterized by high oxygen and silicon content, as expected.
Overall, the key structural features depicted in Figures 4 and 5 are similar, with one key difference: layers +13 and +16 are separated in the positive-ion model-with the former containing Sn + and O + , and the latter only zinc-peaks-whereas in the negative ion model these regions are combined as a single layer (layer −7).Looking at more subtle differences, positive ion mode has fewer distinct interfacial regions between the zinc and tin components, which is expected due the reduced sample mixing when using oxygen as a sputtering source. [18]Positive ion mode also shows the migration of electropositive metals into the coating from both the surface and substrate.
By consolidating all the information in Figures 1-5 and Tables S1-S4 (Supporting Information), we have created a schematic diagram showing each of the major components of the low-E film (Figure 6).This proposed structure is consistent with the patent literature for this type of low emissivity multilayer film. [7,8]We note, however, that specific coating structure is frequently optimized for specific optical performance.The surface of the film contains hydrocarbon, sodium, and potassium contaminants.The outermost layer of the coating comprises titanium and its oxides, which gives the film mechanical strength.Tin oxide is used as the dielectric layer, as it has low electrical resistance, high optical transparency in the visible spectrum, and high reflectance in the IR range. [6]The outermost tin layer is likely shown as chemically distinct in the SOM-RPM model due to the nature of the depth profiling technique, which propels titanium into the successive layers.All other tin layers are bordered by zinc and are therefore shown as chemically identical.Zinc oxide is a seeding layer, allowing the silver layers to have ideal crystal growth, as well as separating the dielectric layers.The base layer of metallic zinc improves adherence to the glass and reduces sodium migration from the glass into the film.

Experimental Section
The sample was a commercially manufactured double silver, low E glass coating comprising of 14 functional layers.A double silver optical stack, of this type, is reported to incorporate silver layers of thickness 12-16 nm with an overall thickness of the stack of the order of 200 nm. [7,8]The samples used in the study were recovered from the commercial supply chain and precise layer geometry was not disclosed.
ToF-SIMS depth profiles were collected using an IONTOF ToF5 ToF-SIMS instrument in positive and negative polarity by alternately rastering a 30 keV Bi + primary ion beam in spectrometry mode and a 1 keV sputtering beam.An O 2 + sputter source was used for profiling in positive polarity (232nA, with dose density of 6.80 × 10 17 cm −2 ), and a Cs + source in negative polarity (90nA with dose density of 3.59 × 10 17 cm −2 ).Data were collected at 128 × 128 pixels over a 100 × 100 μm region within a 400 × 400 μm sputter crater.A flood gun was used for charge compensation.Using a cycle time of 100 μs at one frame per patch and one shot per pixel, positive ion data were collected over 457 scans with a total sputter time of 745 s.Negative ion data were collected over 610 scans with a total sputter time of 999 s.Mass calibration was done using a list of inorganic fragments.Mg + , Si + , Ca + , 41 K + , 46 Ti + , TiO + , Cs + for positive and S − , Cl − SiO 2 − , and SiO 3 − for negative ion mode.
Spectral peaks were identified using the peak search function within the SurfaceLab 7 software, with a minimum count threshold of 100 (678 positive ion and 906 negative ion peaks).Data were exported from SurfaceLab 7 to the bif3D file format and then imported into MATLAB (utilizing an in-house importing script) as an unfolded n × m matrix, where n is the number of pixels and m is the number of features.Data were normalized to total ion count.The SOM-RPM workflow was used to create a machine learning model for each polarity, as previously reported by this group. [19,20]The workflow was modified to reduce computation time.Briefly, a flat RPM algorithm (as per the original work by Li [21] ) was used prior to the 3D toroidal RPM algorithm, as an initial approximation.The result from the flat RPM was then finetuned using the 3D toroidal algorithm.The SOM-RPM method is underpinned by the Kohonen, and CP-ANN toolbox developed for MATLAB (MATLAB R2019b, v9.7). [22,23]SOM models were created using a toroidal topology and square neurons and were initialized using an eigenvalue (principal component analysis) based approach. [19,20]Calculations were completed on a Dell Precision 3650 PC incorporating an Intel Xeon W-1390P CPU (8core) with 128Gb RAM.Calculation times ranged between 24 and 36 h.Models with 64 neurons were trained using a range of epochs/iterations (10-5000), where the 3000 epoch models indicated a minimum in quantization error and hence convergence.The RPM algorithm was applied to the models using in-house MATLAB scripts.Results of the SOM-RPM models were first visualized in MATLAB [24] and then by exporting all x-y slices in the 3D similarity map as TIF images.ImageJ v1.53a was used to produce the 3D renders by importing TIF images as an image sequence and viewing within the 3D volume viewer plugin using nearest neighbor smoothing. [25]

Conclusion
ToF-SIMS can produce datasets of significant size.It is challenging to create valuable and actionable insights from such data relating to the chemical and physical properties of complex materials, structures and devices.In this work, we have demonstrated that the SOM-RPM approach is a suitable paradigm for exploring and visualizing complex ToF-SIMS depth profiles.Traditional analysis of low-E coatings by manual examination of the many peaks associated with each layer of the film impart bias by presenting only a small fraction of the collected data.Prior methods make it difficult to acquire a complete understanding of the dataset, its major components, how they interact, and other more subtle features and interfacial properties.SOM-RPM can quickly and intuitively summarize and visualize vast amounts of data to reveal subtle chemical features in the dataset.Here, we chemically classified and identified all the major constituents of a double silver low-E film, identifying the use of SnO 2 as a dielectric, ZnO as a dielectric and as seeding layers, TiOx blocking layers, a Zn base layer and a TiO x topcoat.Repeating layers are easily identified and can be classified as chemically identical using the entire mass spectrum.Chemical changes in layers intended to be identical can be easily identified.This technique may be readily used to investigate manufacturing issues such as pinhole formation, contamination in one or many layers of a film and poor adhesion.Other applications include industrial forensic studies, characterization of manufacturing lines, and identifying points of origin.

Figure 1 .
Figure 1.Depth profiles of the double-silver low-emissivity coating highlighting key structural film features.Top: Positive ion profile (O 2+ sputter beam).Bottom: Negative ion profile (Cs + sputter beam).The two silver regions have been highlighted in both profiles.

Figure 2 .
Figure 2. Positive ion ToF-SIMS depth profile (O 2 + sputter beam) of a low-E double silver glass coating.A) SOM indicating colored clusters, containing 64 neurons.B) 3D visualization of depth profile indicating similarity using SOM-RPM model of the positive polarity data.C) Reconstruction of 3D structure of the upper silver region.D) Average ToF-SIMS spectra from each layer in C).

Figure 3 .
Figure 3. Negative ion ToF-SIMS depth profile (Cs + sputter beam) of low-E double silver glass coating.A) SOM indicating colored clusters, containing 64 neurons.B) 3D visualization of depth profile indicating similarity using SOM-RPM model of the negative polarity data.C) Reconstruction of 3D layer structure of the upper silver region.D) Average ToF-SIMS spectra from each layer in C).

Figures 2 and 3
Figures 2 and 3 illustrate the quantity and quality of information provided by the SOM-RPM models.The broad compositional structures are revealed quickly and, upon careful examination, remarkable fine structures, such as interfacial mixing and important functional layers are revealed.Repeating layers were easily identified, and the layer structure is intuitively visualized.The silver region structures are quite different between positive and negative ion mode data, likely due to the use of different sputtering sources.As Cs + sputter beam causes more interfacial mixing than the O 2 + , resulting in the silver ions exhibiting a more gradual decrease after reaching a maximum intensity.

Figure 4 .
Figure 4. 2D projection of positive ToF-SIMS depth profile, indicating the coloured toroidal SOM (top left), the total 2D projection of the depth profile (left) and the neural network and position in depth of 20 unique ROIs.

Figure 4
Figure4shows an xz slice through the positive-ion 3D depth profile and divides the total coating thickness into 20 regions of interest, where each region has an associated average mass spectrum.The 10 most intense peaks, and the 10 most highly weighted peaks used for SOM assignment have been chemically assigned and are available in TablesS1 and S2(Supporting Information).Presenting the ToF-SIMS data in this manner highlights repeated and, therefore, chemically identical layers.The position, thickness, and coloring of each layer provides key information and should be considered for each region as well as the chemical information.As the two sputter sources remove material at

Figure 5 .
Figure 5. 2D projection of negative ToF-SIMS depth profile, indicating the colored toroidal SOM (top left), the total 2D projection of the depth profile (left) and the neural network and position in depth of 18 unique ROIs.

Figure 6 .
Figure 6.Proposed structure of low-E coating, indicating the major components of the film structure.