A cross-reactive plasmonic sensing array for drinking water assessment

The continuous monitoring of remote drinking water purification systems is a global challenge with direct consequences for human and environmental health. Here, we utilise a “nano-tastebud” sensor comprised of eight chemically-tailored plasmonic metasurfaces, for testing the composition of drinking water. Through undertaking a full chemometric analysis of the water samples and likely contaminants we were able to optimise the sensor specification to create an array of suitable tastebuds. By generating a unique set of optical responses for each water sample, we show that the array-based sensor can differentiate between untreated influent and treated effluent water with over 95% accuracy in flow and can detect compositional changes in distributed modified tap water. Once fully developed, this system could be integrated into water treatment facilities and distribution systems to monitor for changes in water composition.

Electronic Supplementary Material (ESI) for Environmental Science: Nano.This journal is © The Royal Society of Chemistry 2023

Sample collection
Scottish Water provided 10 samples of influent (INF, I01-I10), 12 samples of effluent (EFF, E01-E12), and 3 sample of tap (TAP, T01-T03) water from various drinking water treatment works sites and randomly selected consumer taps across Scotland.The exact location of each sample collected is not provided due to an NDA with Scottish Water.Samples were frozen upon collection.Prior to analytical chemistry analysis and transmission microscopy, samples were defrosted overnight at 4°C and then filtered using a 0.22 μm filter.

Analytical chemistry
Figure S1 ICP-OES and DOC analysis results of all 25 samples collected from Scottish Water.From top left to bottom left, concentration in ppm for Ca, Na, S, K, Fe, Mn, and Mg (ICP-OES).In bottom right, DOC in ppm.
Inductively coupled plasma optical emission spectroscopy (ICP-OES, Agilent Technologies 5800) was used to assess the levels of the following ions in the filtered samples: Na, Mg, Mn, K, Fe, S, and Ca.Water samples (10 mL) were acidified to 2% with HNO3.Four calibration standards containing a known concentration of analytes were prepared with the 2% HNO3 acidified, ultra-pure water.Calibration standards concentration ranged from 0.05 to 10 ppm (0.05, 0.1, 1, 10).A blank containing no analytes was prepared under the same condition.The ICP-OES connected to Agilent SPS 4 autosampler was used to quantify the standards and the samples.Each sample was injected 6 times for 17 sec and read for 10 sec.A radial viewing mode was used.Analyzed element wavelength for Na, Ca, K, Mg, Mn, Fe, S can be found in SI table 1. Weighted linear regression was used to achieve calibration fit.Fluorescence excitation-emission spectroscopy (FEEM) of each sample spectroscopy was performed on a Horiba Duetta Bio and data exported form the EZSpec software package.instrument was configured with settings as described here: The excitation range was set from 250 to 470 nm with 10 nm increments, while the emission range was set from 280 to 550 nm.To account for the Inner Filter Effect (IFE), an automatic correction was applied.
The excitation and emission band pass were both set at 5 nm, and the integration time for measurements was set to 1 second with a single detector accumulation.During the measurement, each sample was prepared by loading 3 mL into the cuvette, which was then inserted into the spectrofluorometer.Absorbance and fluorescence measurements were taken sequentially, with automatic correction for IFE.Prior to each sample measurement, a blank (pure water) was measured and its signal was subtracted from the sample signal.
Additionally, the samples were Raman calibrated using an excitation wavelength of 350 nm, following the previously described method.

Synthesis of custom nitrilotriacetic acid thiol
Reagents were used as supplied from Sigma Aldrich and Fisher Scientific without further purification.Water used was deionised (>15 MΩ) and non-aqueous solvents used were of analytical grade.Thin layer chromatography was performed on Merck aluminium backed silica gel 60 F254 plates.Visualisation was achieved with UV light, iodine on silica, or KMnO4 in basic aqueous solution.Devices were fabricated using a standard top-down electron-beam lithography process.As shown in Figure S5a, a poly(methyl methacrylate) (PMMA) bilayer (Layer 1: AR-P 642.04 (200k, 4%), anisole, 100 nm thick; Layer 2: AR-P 679.02 (950k, 2%), ethyl lactate, 70 nm thick) was spun on borosilicate glass (PI-KEM Ltd.) followed by the evaporation of a 20 nm aluminum (Plassys MEB 550S) charge conduction layer (CCL).A Raith EBPG 5200 electron-beam lithography tool was used to pattern multiple 500 um diameter circular extents of nanostructure (130 nm x 130 nm squares with 390 nm periodicity in XY) and reticule alignment markers in an array with periodicity of 1.5 mm in Y and 4.5 mm in X. Post-patterning, the CCL was removed using Microposit MF CD-26 (Shipley).Samples were then developed in 2.5:1 isopropyl alcohol (Scientific Laboratory Supplies, Ltd.) to methyl isobutyl ketone (Merck Chemicals) followed by evaporation of 2 nm titanium adhesion layer and 50 nm gold (Plassys MEB 550S) and PMMA/metal lift-off in acetone (Scientific Laboratory Supplies, Ltd.).Commercial microfluidic chambers (Microfluidic Chip Shop) were aligned then attached to the device.After thiolation, each channel was then interconnected to form a single channel.Figure S5b shows the fully assembled device.The nanopatterned surfaces were chemically modified in the microfluidic chambers, with self-assembled monolayers of 10 mM concentrations of functional thiol molecules (R-SH, Table S2, Figure S6) for 1 hour, followed by flushing of 1 mL of pure solvent.A balance between thiolation reaction time and minimizing the exposure of the microfluidic channel adhesive degradation caused by ethanolic solution was necessary.

Figure S7.
Transmission resonance pre (dashed, blue) and post (solid, black) thiolation for each nano-tastebud.The table in the bottom right provided the identified resonance peak pre-and post-modification and the resulting resonance shift.To make the comparison easy, each spectra is smoothed (30 point moving average) and normalized.
The R-group of the thiols were chosen to provide a wide of a range of surface types, which promotes high cross-reactivity between nano-tastebuds.In addition to looking at transmission spectra shift, we also looked at Raman spectroscopy signals once the NTBs were thiolated (Figure S8(a)).Raman spectra maps (4x4, 25 μm step in X and Y) were recorded using a NT-MDT NTEGRA Raman microscope with a 633 nm laser excitation (35 mW power) with 10 second accumulation time.Excitation and collection of Raman scattered light was done using a 20x objective with an estimated spot size of 100 μm.The averaged Raman spectra of the NTBs with the four phenol-containing thiols (ATP, MBA, MPBA, and NTP) are in good agreement with the literature [8][9][10] .The thiols present on the other four NTBs (DDT, NTA, GLU and PFDT) did not exhibit a Raman signal.
We attribute this to the physical structure of the molecule itself combined with the lowmagnification objective used in the setup.Also shown in Figure S8(b) are spectra from NTBs modified with NTP and MBA, before and after storage in water for 30 days.These results show that there is no measurable desorption of the monolayer over this period.Transmission resonance minima was calculated using the second derivative high order polynomial fit of the data and the shift was calculated based on the change in resonance from deionized (DI) Water.

Transmission microscopy
To ensure easy visual comparisons for all figures showing transmission shifts, the plotted transmission spectra were smoothed (30-point moving-average) and normalized.

Linear discriminant analysis (LDA)
LDA with k-fold cross validation (k=5) was used on the NTB sensor dataset to estimate accuracy generate the ROC curves (JMP17 Software) using the same matrix layout as shown in Table S4.

Figure S2
Figure S2 FEEM analysis of results of all 25 samples collected from Scottish Water.

Figure S3
Figure S3 Summary figure of FEEM results grouped by INF, EFF, and TAP.

Figure
Figure S5 (a) Electron-beam lithography fabrication flow diagram.(b) Photo of the fully assembled sensor.(inset) Brightfield microscopy (5x) of one element in the sensing array with alignment reticule and label to the left and circular extent of nanostructure array to the right.

7 Figure S8 .
Figure S8.(a) Averaged Raman spectroscopy signal from 4x4 mapping of the four NTBs with phenol-containing thiols.Clock-wise from top-left: ATP, MBA, NTP, and MPBA.The red line is the average signal and the green outline represents the standard deviation from all 16 measurements.(b) Raman spectra showing the stability of the molecular modifications on sensors that have been stored in water for 30 days.Top row: A sensor identical to those

Figure S9 .
Figure S9.Custom-built microscope (transmission-mode) with programmable XY-stage, syringe pump microfluidics, and spectrophotometer.A custom-built microscope with a programmable XY-translational stage (ThorLabs) was used to measure transmission spectra across each nano-tastebud (FigureS9).Prior to measurements, reticules in the design were used to identify position and correct for angular rotation of the sensor.Light from a broadband LED (10dB, 470-850 nm range, MBB1F1, ThorLabs) was used to probe each nano-tastebud.The transmitted light was collected by a 10x objective and coupled to a StellarNet Microspectrophotometer (StellarNet Blue Wave).Each nano-tastebud was measured five times in a cross-like pattern (50 μm step, see FigureS10).

Figure S10 .
Figure S10.Cross-like pattern used for measuring five spots around a nano-tastebud.Each 'x' represents a 50 μm step.

7 .Figure S11 .
Figure S11.Average resonance shift from DI Water for each nano-tastebud when exposed to the 10, 12, and 3 samples of INF, EFF, and TAP water, respectively.

Figure S12 .
Figure S12.Eigenvectors plots for the first three components of the PCA of (a) the chemometric data and (b) the NTB sensor.PCA scatterplots of the first three PCs of the PCA of (c) the chemometric data and (d) the NTB sensor.

Table S1 .
Wavelength (nm) of analyzed element.Dissolved organic carbon (DOC) concentration (non-purgeable organic carbon) in filtered samples was assessed by TOC-LCPH analyzer with an ASI-L autosampler (Shimadzu, Japan).Control of 2.5 ppm and ultra-pure water was used to assess the quality of the measurement.A full breakdown of the concentrations for each sample is shown in Figure S1.Comparing the INF to EFF 1S-NTA was synthesised in a method adapted from Du Roure et al.1Nα,Nα-Bis (carboxymethyl)-L-lysine hydrate (0.1 g, 0.38 mmol) , NaHCO3 (0.1 g, 1.19 mmol) and 4butryothiolactone (51 μL, 0.59 mmol) were dissolved in 5 mL water and heated to 72 °C for 3 days.The resultant solution was cooled to RT and acidified to pH 3 with 100 μL of glacial acetic acid.The solvent was removed under reduced pressure.The resultant oil was crystalized in EtOH and the crystals were washed with cold 10 mL EtOH and 10 mL hexane.
1H NMR and 13 C NMR were collected on Bunker AVI 400 spectrometers in deuterated solvents as noted.Chemical shifts are recorded in ppm, relative to the residual protonated solvent.Coupling constants are recorded in Hz.Mass spectra data for each compound were recorded on a Bruker microTOFq system using positive mode electrospray ionisation (ESI).Molecular ions or other major ion peaks are reported as m/z. Figure S4.Thiol-Nα,Nα-Bis(carboxymethyl)-L-lysine (HS-NTA)

Table S2 .
List of surface chemistries used to modify the nanopatterned regions.

Table S6a
corresponds to the averaged classification matrix with bootstrapping (2500) and no cross-validation.To validate the sensor, we used k-fold cross-validation (k=5) on the second LDA.This results in 4/5 of the data being used to train the model and 1/5 being used to validate the model, iterated 25 times.TableS6bcorresponds to the accuracy of classifying the data used to train the model (the 4/5), in the model, itself, and TableS6ccorresponds to the accuracy of classifying the validation set (the 1/5), in the model.

Table S6 .
Classification matrices for the LDA grouped by INF and EFF.(a) Average classification from bootstrapping 2500 reps, seed = 456, no cross-validation.(b,c) Average of 25 iterations using k-fold, cross-validation (k=5) for the training set and validation set, respectively.