Preventing Glass Alteration in Museum Objects using ALD Deposited Amorphous Alumina Coatings


 The chemistry and kinetics of glass alteration have been topics of considerable study and debate over the past several decades. While work is still progressing in understanding the mechanisms by which glass degrades, what has not been as well studied are methods of slowing or preventing the alteration of glass surfaces and objects. This is of significant interest to the heritage science community, where the breakdown of historic glass or glass art objects is proceeding, with few viable options available to museum conservators to mitigate its effects. Experimental measurements using atomic layer deposited (ALD) aluminum oxide coatings on a model silicate glass show a significant reduction of the rate of glass alteration.


Introduction 30
Silicate glass is a thermodynamically metastable phase that over time undergoes chemical 31 and physical alterations to attain a more stable state [1,2,3]. Though the process of glass 32 alteration can be slow relative to a human timescale, it presents a complex and pervasive 33 problem in applications where the longevity of the glass structure is important [4,5]. These 34 applications range from the storage of nuclear waste [6,7], to maintaining the efficiency of solar 35 cell components [8], to the focus of this work: the preservation of historic glass objects in 36 cultural heritage collections. 37 Alteration of silica-based glass is a complex process, the kinetics and chemistry of which 38 are dependent on the composition of the glass as well as on the environmental conditions (e.g. 39 pH, temperature, altering solution composition) in which it has aged [4]. The initial alteration of 40 a glass most often takes place via ion-exchange between positively charged water species (H + 41 and H3O + ) and the cationic glass network modifiers (e.g. alkali and alkaline earths) [9]. This is 42 usually accompanied by the hydrolysis of the 43 glass network, which can result in the release of 44 other glass forming species into solution (e.g. 45 Si, Al, and Fe). As alteration proceeds other 46 processes may also occur. For example, a 47 change in solution chemistry may cause the 48 formation of crystalline secondary alteration 49 products, which may in turn change the rate of 50 alteration [7]. Hydrolysis and ion exchange can 51 result in the formation of a hydrated silicate gel layer on the surface of the glass, the 52 swelling/contraction of which produce stress across the interface between the alteration layer and 53 the glass driving crack nucleation [10,11]. Micro-crack networks in the gel layer and ultimately, 54 the formation of salt precipitates on the surface, may result in a hazed and crusted appearance 55 most often associated with degraded glass as shown in Figure 1 [7]. In museum studies and 56 historical materials conservation literature this phenomenon is often referred to as "glass disease" 57 [5]. 58 For the past two decades, one of the most widely accepted models for glass degradation 59 [12] has been based upon the formation of three distinct layers, with a reaction zone (layer 2) 60 forming between underlying bulk glass (layer 3) and a hydrated silicon-oxygen-rich gel altered 61 zone (layer 1) on the surface, illustrated schematically in Figure 2. Following this process, it is 62 often hypothesized that glass alteration is a diffusion-limited process in which the rate of 63 degradation is determined by the interdiffusion of ambient protons into, and cations out of, the 64 bulk glass to form a gel (alteration) layer and a reaction zone [7]. However, recent work 65 conducted by Hellman, et al. [13] suggests that the kinetics of glass degradation could also be 66 reaction-limited, involving coupled glass dissolution and selective re-precipitation of the 67 secondary, hydrated silicate 68 gel layer. 69 Regardless of which 70 kinetic mechanism limits 71 glass alteration rates, the 72 enabling factor is water or 73 other H-bearing molecular 74 species coming into contact 75 with the glass surface. 76 Therefore, the most effective way to slow glass alteration is to isolate the underlying glass from 77 the ambient reservoir of such species. Currently, very few options exist to markedly slow or 78 prevent glass alteration from occurring. These options include the application of sol-gel silica 79 coatings, which ultimately suffer from the same hydration driven swelling as the glass [14], the 80 application of graphene, which has shown to be effective but is not feasible for application over 81 large areas [15], and the use of dry cases filled with nitrogen gas in museums, which is effective 82 but expensive to purchase and maintain; this last approach is limited to relatively small, high 83 value objects [16]. The current paper examines the use of atomic layer deposited (ALD) 84 amorphous metal oxide coatings as diffusion barriers for reducing the rate at which water or 85 other H-bearing molecular species come into contact with the surface of glass objects, thereby 86 slowing alteration of these surfaces. 87 The scope of this paper is on the efficacy of ALD coatings for the mitigation of glass 88 alteration. However, in addition to efficacy, there are other factors that are important to 89 understand prior to the use of new methods as conservation treatments in museums which will 90 not be discussed here. These factors include the effect of the applied coatings on the appearance 91 of the object, and the reversibility of the treatment. Ongoing work to characterize degree of the 92 apparent color change imparted by the applied coatings as well as the relative rates of etching for 93 Al2O3 ALD from these glass surfaces will be reported elsewhere. 94

Methods/Experimental 95
Glass Samples: 96 Soda lime float glass, the glass type most typically used in modern windows, is silica 97 based with Na2O as the primary flux and Ca(OH)2 as a stabilizer. Samples of this type of glass 98 (from a window from a commercial manufacturer (Guardian Glass)) were chosen for this study. 99 The glass has a composition with mass fractions of approximately 0.75 kg/kg SiO2 and 0.10 100 kg/kg Na2O. Other major, minor and trace additives listed in Table 1 prior to the analysis of an unknown sample in order to verify that the compositional 124 measurements and calculations being used produce accurate results) and the percent deviations 125 from the expected values were found to be less than 7 % for the oxides of all major elements (SI 126 To mitigate potential effects of glass inhomogeneity, multiple samples taken from 128 random locations on a large glass pane were used in each experiment and the results were 129 averaged. Modern float glass manufacturing processes involve exposure to liquid tin on one side 130 of the window which allows a trace amount of tin to diffuse into that face of the glass [20]. To 131 account for this, the composition was measured on both faces of the glass window. 132 reflections at the surface and interface with the substrate can lead to visible color shifts [22]. In 150 spite of a nonuniform optical absorption across the visible spectrum, TiO2 ALD coatings, which 151 cause the glass to appear blue, were also examined in this work owing to a previous report of the 152 relative lack of durability of alumina ALD films in water [23]. The Al2O3 ALD coating process 153 used trimethylaluminum (TMA) and water as precursor gasses while TiO2 ALD used 154 tetrakis(dimethylamido)titanium (TDMAT) and water as precursors. Figure 3 depicts a 155 schematic of the idealized deposition process for Al2O3 ALD. 156 Exposure and purge times for the precursor gases were determined empirically based on 157 perceived color uniformity across a Si wafer that was included in the reactor chamber during 158 each deposition experiment; gradients in the local pressure within the reactor lead to thickness 159 variations if the individual reactions do not reach saturation The pulse/purge times suggested by 160 the ALD reactor manufacturer were found to be insufficient for coating the silicon wafers, and 161 presumably the silicate glass samples uniformly. Therefore, the suggested precursor exposure 162 and purge times were quadrupled, resulting in a uniform visual appearance; these resulting times 163 are listed in Table 2. All depositions discussed here were performed at 150 °C. 164  to the ALD precursor gases. Samples were flipped over halfway through each coating process to 171 expose the contact points that the sample made with the support and to promote complete sample 172 coating. 173

Coating Thickness 174
Subsequent to sample coating, the average thickness of the applied ALD coatings versus 175 the number of deposition cycles was determined by spectroscopic ellipsometry. The ellipsometry 176 was performed on a Woollam M2000 spectroscopic ellipsometer. Both a tungsten filament and a 177 deuterium bulb light source were used to cover a wavelength range of 190 nm -1800 nm. These 178 analyses were performed with the instrument in horizontal mode due to the weight of the 179 samples. The proprietary Woollam software package CompleteEase was used to build sample 180 models, including optical constants and film and substrate parameterization, and to fit the 181 resulting data. Results are reported in Table 3. 182 The Al or Ti signals obtained via SEM-based μXRF were measured for increasing 183 numbers of cycles of ALD deposition. The relationship between the XRF output and 184 ellipsometry thicknesses was found to be highly linearly correlated as the number of ALD 185 deposition cycles is increased beyond an initial transient (discussed below). Taking advantage of 186 this correlation, the XRF signal was used to estimate the ALD coating thicknesses. The μXRF 187 analysis was performed under high vacuum in the SEM using a Bruker XTrace μXRF with a 50 188 kV polychromatic Rh source and an X-ray optic that produced a 33 μm (measured at Cu K ) 189 elliptical spot with an approximate area between 900 m 2 and 1000 m 2 . X-ray spectra were 190 collected for 60 live seconds. Using Bruker Esprit v2.1, the resulting X-ray scatter was 191 background fit using a spline interpolation of non-peak channels and peaks were fitted using 192 Gaussians with overlapping X-ray lines deconvolved to yield the net counts of Al or Ti that 193 represent the sum signal resulting from the glass substrate plus the ALD thin film [24]. The 194 results for both ellipsometry and μXRF analysis are reported in Table 3.  Table 3. Analyses were run in sets of ten for each sample measured and 246 the relative standard deviations output by the instrument software (Optimass version 2.2) were 247 found to be on average 4.2 % for Na and 4.1 % for Si. 248 In addition to analyzing the aqueous leachate, analyses were also performed on three 249 samples of water that had not been in contact with a glass sample but that had been subjected to 250 identical aging and storage conditions to serve as aqueous "blanks". The values measured from 251 these blank samples were averaged and the average concentrations were subtracted from the 252 values obtained from the leachate that had contained the glass samples. The resulting 253 "background" corrected concentrations, determined for each of the multiple samples with a given 254 coating type, coating thickness, and aging time were averaged together and the resulting mean 255 values are reported in Figure 6 to one standard deviation of the population statistics. 256    Table 3, often result 287 in error bars that are too small to be seen in this figure.

288
A linear relationship between the ellipsometrically measured coating thicknesses and the 289 net counts of Ti and Al measured using μXRF can be seen in Figure 4, with regression 290 cefficients given in the figure. The non-zero y-intercepts obtained from the linear fits of the data 291 shown in this graph are consistent with the pressence of small amounts of both Ti and Al in the 292 base glass composition. The uncertainties in the linear fits were calculated for both ALD 293 coatings. For the Al2O3 ALD measurements, the resulting linear fit was found to have a slope of 294 406 ± 16 net counts/nm, and a y-intercept of 2154 ± 1316 net counts which is indistinguishable 295 from the measured value for net counts of Al in the base glass (1833 ± 543 net counts) within the 296 statistical uncertainty. Similarly for the measurements made on the TiO2 ALD coatings, the slope 297 was found to be 863 ± 68 net counts/nm and the y-intercept to be -737 ± 2470 net counts, while 298 the measured value of Ti net counts in the base glass was determined to be 1444 ± 38. 299 The measured value of Al and Ti in the base glass were found to be within one standard 300 deviation of the y-intercepts obtained from the fits for each coating. However, it is interesting to 301 note that there is a systematic deviation apparent between data and the linera regression fits, 302 particularly for Al2O3 ALD coatings for small numbers of coating cycles, which can be 303 attributed in a deviation from layer-by-layer growth. It is apparent for these thinner coatings of 304 Al2O3 ALD, the initial thickness shows a sublinear dependence on the number of deposition 305 cycles. Following standard practice, during each of these coating experiments, a control Si wafer 306 was placed inside the ALD chamber along with the glass samples to allow for an estimation of 307 the applied coating thickness on an ideal substrate. A large discrepancy between the thicknesses 308 measured on the control wafer and the glass samples was noted for the thinnest Al2O3 ALD 309 coatings (Table 3), with the latter values initially being consistently lower. This suggests that the 310 precursors are reacting less readily with the glass surface than with the Si wafer, which may be 311 due to locally unfavorable surface terminations on the glass. We hypothesize that island 312 nucleation and growth occur in the early stages of the formation of these coatings [ consistent with island nucleation and growth leading to some degree of coalescence ( Figure 5). 321 In addition to the topographic maps, Figure 5 displays plots of the average RMS roughness 322 measured as a function of coating thickness for both Al2O3 and TiO2 ALD coatings. 323 The peak value of the RMS surface roughness, measured for an Al2O3 ALD dose of 50 324 cycles (≈ 2.3 nm), was found to be approximately 3 times larger than that of the uncoated glass 325 and on the asymptotic value; this is qualitatively consistent with island nucleation, followed by 326 subsequent lateral, as well as vertical growth [28]. The onset of roughening and island formation 327 seemingly requires higher doses for the TiO2 ALD coatings, which in turn suggests a lower 328 reactivity of TDMAT, as compared to TMA, with the silicate samples. For both the Al2O3 and 329 TiO2 ALD coatings the intial increase in roughness was followed by a gradual drop with 330 increased coating thickness. This behavior is consistent with growth and at least partial 331 coalescence of multilayer islands at longer deposition times [30]. where the field of view is approximately 10 m × 10 m area (green is shallow while purple represents peaks). It 339 has previously been shown that for AFM measurements where the features of interest are at least double the radius 340 of the AFM probe tip, the instrumental uncertainty is estimated to be less than 15 % of the measurement [31].

341
Uncertainties are reported here to 30% of the measured value for n = 1.

Uncoated Glass 345
It was first necessary to characterize the alteration of float glass without any applied ALD 346 coating under the accelerated alteration method described previously in order to determine the 347 effect of an ALD coating on the glass alteration rates. The graph shown in Figure 6   The results obtained from analysis of the water used in aging Al2O3 ALD coated samples 373 contrast sharply with the results obtained from analysis of the water used in aging TiO2 ALD 374 coated samples as shown in Figure 6. No statistically significant decreases were observed in the 375 either the Si or Na concentrations, measured from TiO2 ALD coated samples ( Figure 6A The ICP-MS data indicates that while there is a significant decrease in Si and Na loss for 381 Al2O3 ALD coated glass samples ( Figure 6B), the loss values remain measurable even for the 382 Figure 7: Si and Na ion concentration data obtained from the water from the 14 days +/-4 hours of accelerated aging of samples of float glass that had been coated with increasing numbers of deposition cycles of (A) TiO2 ALD and (B) Al2O3 ALD. Points plotted at zero cycles show data obtained from glass samples without any applied coatings (also shown in Figure 6). Multiple samples were aged (n = 6) for each coating and relative thickness reported. Uncertainties reported to one standard devation of the population statistics. thickest films investigated. However, even defect-free films allow interdiffusion of Na and Si 383 under ambient conditions giving rise to finite loss but this should fall off with a characteristic 384 thickness dependence not observed here [34]. Indeed, the bulk of the loss of Na and Si measured 385 for the thickest films is likely attributable to the presence of uncoated areas on the glass as shown 386 in Figure 8 and is discussed in the next section. 387 388 ALD Coatings After Aging: 389 390 Appearance 391 After accelerated aging for 14 days, the coated glass samples were removed from the aqueous 392 solutions and visually inspected. It was immediately apparent that significant coating loss had 393 occurred during immersion in hot water. The surface of the samples appeared rough and pitted, 394 and with greater scrutiny using various microscopic methods (see Figure 8), it was apparent that 395 large areas of localized coating were lost during alteration. We postulate that these bare areas 396 may result from the expansion of smaller defects in the ALD films (discussed below), which 397 allow local alteration of the underlying glass; if so, expansion of the altered regions due to 398 hydration could cause the coating on the surrounding surface to flake off. Such bare areas were 399 observed to various degrees on all samples after accelerated aging treatment. Similar behavior 400 has been observed previously in other thin film coatings exposed to ambient humidity [35,36].  Figure 9: A. Tiled-visible light micrograph-image of a surface of a float glass sample that had been coated with 90 nm of Al2O3 ALD and subsequently exposed to 14 days of accelerated aging and B. the same image after contrast thresholding. The binary image, B, is comprised of white pixels representing remaining ALD coating, and black pixels that represent exposed glass substrate.
The results of these measurements for several of the ALD coatings examined are 418 presented in Table 5. The measured spatial resolution for this optical microscope and these 419 samples was found to be ≈ 2 m. As a result, closely spaced (≤ 2 m) fissures and pinholes were 420 not detectable by this imaging method. Therefore, the values given in Table 4 represent an upper 421 limit on the fraction of the surface still coated. Additionally, this resolution constraint also 422 prevents the detection of any closely spaced defects or pinholes that are present on the surface of 423 the ALD coating prior to aging. 424 Table 4 The AFM results ( Figure 5) indicate that a degree of island coalescence occurs in both of 435 these coatings; however the finite radius of the probe does not allow fissures smaller than ≈ 10 436 nm to be detected, and the presence of such fissures could be the root cause behind this coating 437 loss [31]. In principle the loss could also be due to thermal fracture [37].

475
Another important mode of defect formation in ALD coatings can be thermal fracture. 476 ALD is typically performed at elevated temperatures (≈ 150°C) to allow for practical reaction 477 and purge saturation times [27]. Differences in the coefficients of thermal expansion between the 478 ALD coating and the substrate results in stress as one of these (substrate or coating) contracts 479 more than the other during sample cooling subsequent to deposition. The coefficients of thermal 480 expansion are: 4.2 x 10 -6 K -1 for Al2O3 ALD [42], 13.9 x 10 -6 K -1 for TiO2 ALD [43], and 8.0 x 481 10 -6 K -1 for commercial float glass [44,45]. These differences could result in the blistering 482 (Al2O3) or fracturing (TiO2) of the ALD coatings during cooling to room temperature, provided 483 the thermal stress exceeds a critical value [37]. However, systematic experiments by Jen, et al. 484 [46] showed that the critical compressive stress for crack formation for Al2O3 films on Teflon 485 FEP substrates saturates with film thickness at a lower limiting value of approximately 1.8 GPa. 486 This exceeds the thermal stress calculated for lowering the temperature from 150˚C to 25˚C, at 487 113 MPa, by more than an order of magnitude, and seemingly makes thermal fracture an 488 unlikely source of failure of the ALD films studied here. 489 490

Conclusions: 491
The application of optically transparent Al2O3 films via atomic layer deposition was 492 shown to dramatically decrease the release of select major elements (Si and Na) from float glass 493 when immersed in water at elevated temperatures, and it is thus potentially a conservation 494 treatment for glass in museum environments. 495 While the uptake of the elements by other solid phases formed during alteration cannot be 496 excluded, and therefore artificial suppression of those elements in the altering solution, as 497 another possible explanation for these results, such an explanation could be ruled out if the 498 increasing concentration of Na and Si in the leachate scales with their depletion in the glass 499 matrix. Experiments to test this are in progress. 500 In contrast, application of TiO2 films did not significantly decrease the release of these 501 elements. There are a number of possible reasons for the difference in performance between the 502 two oxide coatings. It could be due to a difference in the rate, or saturation extent of coalescence 503 of islands for the two types of oxides studied here, possibly because of the relative reactivity or 504 sizes of the organometallic precursors used, respectively. 505 Our observations suggest that the limit to the efficacy of ALD Al2O3 coatings in 506 mitigating glass alteration is set by delamination of the films when immersed in water; these may 507 be nucleated at existing pinholes and/or fissure defects in applied ALD Al2O3 coatings on silicate 508 glass. Further study is needed to test for the presence and role of such defects and what steps 509 might prevent their formation. Disclaimer: Trade names and commercial products are identified in this paper to specify the 527 experimental procedures in adequate detail. This identification does not imply recommendation 528 or endorsement by the authors or by the National Institute of Standards and Technology, nor 529 does it imply that the products identified are necessarily the best available for the purpose. 530

Contributions of the National Institute of Standards and Technology and Museum Conservation 531
Institute (a member of the Smithsonian Institution) are not subject to copyright. 532 Availability of data and materials: The datasets used and/or analyzed during the current study are 533 available from the corresponding author on reasonable request. 534 Competing Interests: The authors declare that they have no competing interests. 535