Assessing the Impact of LED Lighting on the Stability of Selected Yellow Paint Formulations

2.1. Light exposure chambers

Accelerated aging was undertaken using purpose-built light boxes designed and manufactured by Complete Lighting Systems. Accelerated exposure provides insight into the long-term performance of different lighting regimes, within a short time period. Given the sector’s misgivings about LEDs, the aim of this project was to establish the extent to which paint stability differs on exposure to LED lighting relative to the tungsten lighting currently employed within English Heritage historic house environments. Current English Heritage guidelines restrict the use of LEDs with a 450 nm emission greater than 33% of the broad phosphoresce and with a full width at half height greater than 20 nm. Two LED lighting systems were therefore selected to interrogate the importance of these guidelines. These were compared against an accelerated tungsten lighting regime and a dark control environment. The color metrics for each light source (Table 1) and normalized spectral power distributions (Fig. 1) were measured using a GL Optic Spectis Touch 5.0 in situ within each chamber. Where applicable the UV filters were in place during measurement.

A further set of samples was exposed in real time to daylight to establish the relative color shift. Samples were exposed to UV-filtered daylight and periodically analyzed using colorimetry (Fig. 1 and Table 2). Total exposure levels were dependent on seasonal weather conditions, but exposures reaching 5.4 Mlux h were achieved over the duration of the experiment.

All samples were dried in the dark for 9 months at 20.1 ± 1.9°C and 42.8 ± 7.1% relative humidity (RH) prior to initial analysis and subsequent exposure. Separate sets of samples were exposed to the three artificial light sources, giving an illuminance at the sample surface of 47–50,000 lux. This was taken to be reciprocal because it is well below the level at which Del Hoyo-Melendez and Mecklenburg (Citation2011) demonstrated the relationship to break down. Lux levels, temperature, and relative humidity during the periods of exposure were monitored using Elsec 765C environmental loggers. Data were acquired every 10 min and the average values at each sampling interval for each chamber are included in Table 2. Diffusion tubes were used to monitor ozone and NO_x levels in each chamber throughout the period of exposure.

The two LED light boxes were cooled using both passive and active measures, namely, heat sinks and a fan system. The TH lamps were cooled using off the shelf heat sinks and fans. The halogen lamps were UV-filtered using a 4-mm polycarbonate sheet placed between the lamps and the samples. The control samples were masked from light exposure (dark) with metal foil and placed within the TH light chamber, exposing them to the ambient temperature and RH conditions.

2.2. Sample preparation

2.3. Colorimetry

Following 9 months of drying in the dark, initial colorimetry analysis was undertaken for each paint sample, with three replicate measurements on each sample taken before and after aging. Markers were used to ensure that the sampling window could be repositioned on the same sampling location.

2.4. Fourier transform infrared–attenuated total reflectance

To identify the chemical alterations associated with light-induced change, mid-infrared analysis was carried out using a Bruker Alpha Fourier transform infrared spectrometer (FTIR) fitted with a Bruker Platimum diamond attenuated total reflectance (ATR) accessory, using Bruker OPUS software version 7.5. A scan range of 400–4000 cm⁻¹ was employed, with a wavenumber resolution of 2 cm⁻¹ and scan accumulation of 32. Three replicates of each sample were acquired and averaged using Thermo-Scientific GRAMS AI software version 9.1. FTIR analysis was carried out on all oil-based samples by removing them from the Teflon substrate using a 3-mm Harris micropunch and analyzed with the exposed paint surface in contact with the analysis crystal.

FTIR analysis was carried out on all oil-based samples. The gum arabic paint samples were omitted from FTIR analysis because the intense signal from the paper substrate masked the relatively weak signal from the binder and pigments.

2.5. Multivariate analysis—Principal component analysis and partial least squares

Principal component analysis (PCA) and PLS were used to relate chemical changes to exposure type, duration, and color shift. Due to the large data set generated, this multivariate approach was aimed at mapping spectral variations displayed across the infrared spectral region and determining correlations across the sample set.

The initial PCA data matrix (X) was composed of rows of FTIR spectral data corresponding to the number samples in each data set, with each column of the matrix corresponding to the number of variables, in this case wavenumber. In this work, a 35 × 3 529 X matrix was constructed for each pigment/oil binder composition relating to 35 samples (seven sampling intervals for dark, daylight, TH, LED1, and LED2) and 3529 wavenumbers. Individual matrices were constructed for each pigment/oil binder composition to enable comparisons between the lighting systems. This aimed to highlight inherent clustering related to exposure type and duration and establish the key variables impacting on such differences in principal component space.

Partial least squares is a regression method based on an iterative algorithm between two matrices, X and Y. As before, the X matrix relates to the spectral responses. Accompanying each X matrix were two 35 × 1 Y matrices for PLS regression corresponding to the independent responses (1) level of exposure (Mlux h) and (2) color shift (∆E₀₀).

PCA and PLS were carried out using Camo Unscrambler X version 10.3. All FTIR spectra were preprocessed to reduce variability caused by differences in scattering and background noise. All data were smoothed using a Savitzky-Golay filter, fitting a second-order polynomial with an 11-point symmetric window (Savitzky and Golay Citation1964). The smoothed spectra were then subsequently normalized using standard normal variate (Fearn et al. Citation2009; Muehlethaler et al. Citation2011). Unless otherwise stated, PCA and PLS were carried out on spectra spanning the entire mid-infrared region from 400 to 4000 cm⁻¹, with a wavenumber resolution of 2 cm⁻¹, equating to a total of 3529 variables. Later iterations were performed on smaller, truncated regions of the spectrum determined by the PCA loadings and PLS regression coefficient plots of individual models. The data were mean centered and the models validated using the full cross-validation method (Esbensen Citation2002).

Overly optimistic calculations of principal components is as a common problem within MVA (De Maesschalck et al. Citation1999). Therefore, each model was calculated with a total of seven components; however, later principal components were often dominated by interference and therefore are not reported on.

3.1. Lead chromate sulfate (PbCrO₄∙xPbSO₄) in gum arabic

Color shift across all sample exposures appeared to be equivalent and followed similar patterns of change; the TH lamps performed somewhat better at exposures exceeding 4.5 Mlux h (Fig. 2). On initial exposure there is a relatively rapid change in ∆E₀₀, reaching levels of ∆E₀₀ ≈ 3.5 at exposures of 9.5 Mlux h. After this point there is an apparent plateau in ∆E₀₀ across all samples until the final sampling interval of 150 Mlux h, where LED1 displays a slight increase in ∆E₀₀. Across all sampling intervals the TH samples display higher standard deviations compared to the other sample sets. When considering that the reproducibility error is low for lead chromate sulfate in gum arabic, this indicates sample inhomogeneity.

The changes in color are due to a decrease in the L* and b* coordinates, both of which follow a pattern of behavior similar to that of ∆E₀₀: an initial decrease in L* and b* followed by a plateau. The samples display a rapid darkening, which is independent of exposure type, and a shift to a more blue hue (−b*; Fig. 3).

Up to 0.5 Mlux h the samples exposed to daylight exhibit a marginally higher color shift compared to those exposed to TH, LED1, and LED2, although ∆E₀₀ is borderline with levels of discernible change. Between 1 and 4.8 Mlux h all four light exposures are equivalent (Table 5).

3.2. Lead chromate sulfate (PbCrO₄∙xPbSO₄) in linseed oil

On exposure to tungsten halogen, the samples display a perceptual increase in ∆E₀₀ up to exposure levels of 37.5 Mlux h (Fig. 4). After this exposure, the color shift plateaus and remains constant (within error) up to 150 Mlux h. This color shift is due to a significant decrease in the L*, a*, and b* coordinates, namely, a darkening of the paint coupled with a shift in hue to a more green and blue paint layer (Fig. 5 and supplemental material 2). This behavior is mirrored by the samples exposed to both LED lamps; color shift is equivalent across all samples when accounting for repeatability error, standard deviation, and levels of discernible change. At 150 Mlux h there is a slight increase and divergence in ∆E₀₀ values for the LED1 samples, which is mirrored in the L*, a*, and b* coordinates.

The alterations in color displayed at low to moderate exposures are not significantly different between the three artificial aging regimes. At 150 Mlux h there is a divergence in the values of ∆E₀₀, with LED1 displaying marginally poorer performance and discernible difference compared to the LED2 and TH samples. It is worth noting that at exposures above 9.5 Mlux h the color shift in the gum arabic samples is approximately two-thirds lower than that in the equivalent linseed oil samples. Although it might be reasonable to expect greater change in watercolor layers due to the double exposure through reflection of the paper, the large contribution from the underlying paper in the translucent watercolor will have a significant impact on ∆E₀₀ compared to the opaque linseed oil paint. The dominance of the paper through the thin watercolor layers might also cause a less perceptible change.

The samples exposed to filtered daylight exhibit a significant increase in ∆E₀₀ compared to the TH, LED1, and LED2 samples. The deviations are higher but still significant after daylight exposure, with ∆E₀₀ values averaging 6.2 at 2.3 Mlux h, compared to ∆E₀₀ ≈ 2.4 for the other three exposures. At low exposure levels the daylight samples exhibit a marked increase in ∆E₀₀ relative to the artificially aged samples (Table 5).

The FTIR analysis shows the appearance of a spectral band centered at 470 cm⁻¹ with increasing exposure, which is attributed to the formation of PbO₂ (Fig. 6; Gautam et al. Citation2012). The exact route of this photochemical reaction is still not fully understood, but it has been proposed that lead chromate is reduced to chrome oxide and lead oxide (Erkens et al. Citation2001; Monico et al. Citation2011). Whatever the precise mechanism, it is clear from the spectra that the Pb-O band at 470 cm⁻¹ increases with exposure time across all artificial exposures, namely, TH, LED1, and LED2.

The distinct relationship between exposure and the increase across this particular spectral region was further interrogated using PLS to determine whether the measured levels of exposure (Mlux h) could be quantitatively correlated with the FTIR spectra and the formation of PbO₂. A PLS model was calculated regressing the variable region from 400 to 650 cm⁻¹ against Mlux h for the control, TH, LED1, and LED2 samples.

The corresponding regression coefficients for the first and third axes were shown to be influenced by 470 cm⁻¹, in addition to an inverse relationship between variables 595 cm⁻¹ and 610 cm⁻¹ in the case of the third axis (supplemental material 3). The bands at 595 cm⁻¹ and 610 cm⁻¹ are attributed to SO₄²⁻ bending modes (Lane Citation2007), and the intensity of these were also shown by Monico et al. (Citation2011) to invert during the aging of lead chromate sulfate.

The FTIR analysis and PLS regression demonstrate the increasing presence of degradation products in the lead chromate sulfate, namely, lead oxide (Fig. 7). Direct correlations were found between the spectral aging patterns, specifically the spectral band centered at 470 cm⁻¹, and the duration of exposure and color shift. This correlation (R² = 0.94) was particularly evident in the samples exposed to LED1 and is in keeping with the marginally higher ∆E₀₀ values shown at high exposures to LED1 (Fig. 4).

3.3. Arsenic sulfide (As₂S₃) in gum arabic

The repeatability error is low compared to the degree of color shift in the arsenic sulfide samples kept in the dark, suggesting that the fluctuations in ∆E₀₀ are not due to sampling errors (Table 4 and Fig. 8). However, the standard deviation is large. It should be borne in mind that orpiment is a ground mineral pigment and the particle size range is broad, giving a heterogeneous surface topography. Differences in color shift are likely to be influenced by alterations in scattering as a result of drying, which is physical change. On initial exposure the dark samples show an apparent increase in ∆E₀₀ followed by a temporary decrease. This initial increase is concomitant with a decrease in the a* coordinate, suggesting the formation of a greener paint layer on drying (Fig. 9).

The standard deviations at each sampling interval are high, making interpretation tentative. However, all samples exhibit a decrease in a*, indicating the formation of a greener pigment layer as exhibited in the control samples. Up to exposures approximating 4.5 Mlux h all three artificial lighting systems are equivalent and below the limits of visual detection (Fig. 8). Between 9.5 and 73 Mlux h LED1 appears to perform marginally better than the TH and LED2. All three light exposures exhibit a significant decrease in the b* coordinate, equating to a loss of yellow and a shift toward a blue hue (Fig. 9). At very high exposures, above 75 Mlux h this increase in blueness is retarded on exposure to the TH lamps. This is supported by the ∆E₀₀ values and indicates that at high exposures arsenic sulfide in gum arabic is more stable under TH lighting.

When accounting for the high standard deviations, the samples exposed to filtered daylight display equivalent levels of color shift.

3.4. Arsenic sulfide (As₂S₃) in linseed oil

The standard deviations and repeatability errors are low, indicating a homogenous surface and color shift across all samples. At exposures up to 9.5 Mlux h all three artificial lighting exposures show similar levels of change, ∆E₀₀ ≈ 1.0 (Fig. 10). At higher exposures there is a divergence, with the TH and LED2 samples exhibiting a greater color shift than the LED1 samples.

The LED2 samples demonstrate a clear, perceptual color shift above 18.5 Mlux h. The change in color is due to a decrease in L* and b* and an increase in a*. This latter change in hue is in opposition to the decrease in a* exhibited by the TH and LED1 samples (Fig. 11). The former lamp induces redder color shift, whereas the TH and LED1 lamps tend toward a greener shift in hue.

The discoloration of arsenic sulfide (orpiment) is well documented and known to form the more stable arsenic oxide compound, arsenolite (As₂O₃), particularly on exposure to wavelengths shorter than 428 nm (Allen et al. Citation2005). Whitmore and Cass (Citation1989) have also suggested that there may be a link between the presence of nitrogen dioxide and accelerated deterioration of both orpiment (As₂S₃) and realgar (As₄S₄). The concentrations of nitrogen dioxide were equivalent inside all three light chambers. Therefore, the measured color shift is likely due to the higher spectral power distribution between 380 and 429 nm in the LED2 lamps.

Figure 12 shows the appearance of an FTIR band at 790 cm⁻¹ that confirms the formation of arsenic oxide (As₂O₃; Paiuk et al. Citation2012; Vermeulen et al. Citation2016), accompanied by a band and As-O vibration at 1030 cm⁻¹. There were no bands associated with arsenic sulfide because its absorbance frequencies are outside of the mid-infrared region.

To interrogate the alterations in the infrared spectra following exposure, the color shift data were regressed against the primary FTIR variables using PLS. Figure 13 shows the predicted versus measured validation line for ∆E₀₀, indicating a high degree of correlation between the FTIR and colorimetry analysis (R² = 0.97). The associated regression coefficient plots for principal axes 1 to 4 are dominated by the band centered at 790 cm⁻¹, associated with the formation of arsenic oxide (As₂O₃). This is a clear indication that the color shift of the arsenic sulfide oil paint is primarily driven by the degradation of the orpiment pigment (Vermeulen et al. Citation2016) and exacerbated by exposure to LED2.

3.5. Cadmium sulfide (CdS) in gum arabic

Accounting for repeatability error and levels of perceptual change, the samples kept in the dark exhibited no significant color shift (Fig. 14). The same is true for the TH, LED1, and LED2 samples. Although there were slight increases in ∆E₀₀, the maximum value across all three samples sets was ∆E₀₀ = 0.9, which is below the limit of discernible change. The samples exposed to UV-filtered daylight followed behavior equivalent to that of the other exposed sample sets.

3.6. Cadmium sulfide (CdS) in linseed oil

The TH, LED1, and LED2 samples displayed slight increases in ∆E₀₀ when compared to the control samples, with the maximum value across all three samples sets reaching ∆E₀₀ = 1.4 (Fig. 15). Although this is potentially a detectable change in color, when considering the error of repeatability and the similarities across all of the exposed samples these differences are not significant. The minor changes detected are due to a positive shift in the L* and b* axes and a negative change in the a* coordinate, equating to a lightening of the paint surface and a slightly more green/yellow hue (Fig. 16).

Leone et al. (Citation2005) and Mass et al. (Citation2013) demonstrated that CdS was susceptible to chemical alterations in the pigment structure on exposure to light and high RH (85%), which resulted in the formation of brown, white, and colorless compounds. Eastaugh (Citation2008) also suggested that moisture may have a deleterious effect on the pigment if only leanly bound in the paint. However, Leone et al. (Citation2005) found that alterations were dependant on the physical conformation of the pigment and where there was a higher content of amorphous cadmium sulfide the pigment was more prone to deterioration. Although there were some changes to the color of the CdS samples, these were very low considering the degree of exposure of the samples. The reason for apparent stability across all lighting types is potentially twofold, namely, the low relative humidity within each aging chamber and quality of the manufactured pigment, which is likely to have a high degree of crystallinity. The samples exposed to filtered daylight followed the equivalent behavior.

The initial PCA across the entire spectral range displayed no inherent clusters or patterns relating to exposure type or duration, although there was evidence of the exposed samples clustering away from the control samples kept in the dark (data not shown). Following variable selection the overall description of the samples improved, with the explained variance at two components increasing from 87% to 93%. Primary variables influencing clustering were 1740 cm⁻¹, 2850 cm⁻¹, and 2922 cm⁻¹, the first exhibiting an inverse relationship to the latter variables. The band at 1740 cm⁻¹ is assigned to the C =O stretch associated with ester linkages formed during oxidative polymerization of fatty acid materials—that is, linseed oil—whereas the bands at 2850 cm⁻¹ and 2922 cm⁻¹ suggest the loss of hydrocarbon. A reduction in the intensity of the CH₂ bands 2850 cm⁻¹ and 2922 cm⁻¹ is in line with oxidative chain cleavage (Meilunas et al. Citation1990). There was no systematic clustering associated with exposure to particular light sources and no evidence of alterations in the chemical structure of cadmium sulfide (Mass et al. Citation2013) or the presence of pigment and oil binder interactions (Pouyet et al. Citation2015). All molecular changes on exposure were ascribed to the oxidation and polymerization reactions associated with the continued drying of linseed oil.