NANOSTRUCTURED SURFACE ENHANCED RAMAN SPECTROSCOPY SENSOR FOR MARINE POLLUTANTS

Excess concentrations of nitrate and phosphate in seawater can lead to harmful algae blooms that damage coastal ecosystems, pose health risks and adversely impact commercial activity. Early in situ detection of over-nutrification is necessary for rapid response and mitigation plans. Commercial nitrate and phosphate sensors utilize UV-Vis spectroscopy methods. Those sensors show interference with ions present in seawater and are prone to biofouling, necessitating new approaches for in situ monitoring. Surface enhanced Raman spectroscopy (SERS) is a technique theoretically capable of single molecule detection, and therefore may be a promising approach for nitrate and phosphate detection. However, there are clear challenges as SERS sensing is negatively affected by interference in complex media and in situ sensing in a solution phase reduces accuracy and resolution. It is because of these challenges, in part, why much of the data reported in the literature are taken for purified samples that are then dried on a SERS substrate. Our goal is to address the engineering challenges for a SERS in situ seawater nutrient concentration measurement system. Batch and flow-through devices have been designed to incorporate commercially available, nanostructured gold SERS substrates. By benchmarking against 4-nitrobenzenethiol/ethanol solutions and ultrapure water spiked with nitrate and phosphate, our results show that our SERS devices can be used as a development platform for a seawater nutrient sensor, showing a route for a commercializable product that will greatly benefit scientific operations.

UV-Vis spectroscopy methods. Those sensors show interference with ions present in seawater and are prone to biofouling, necessitating new approaches for in situ monitoring. Surface enhanced Raman spectroscopy (SERS) is a technique theoretically capable of single molecule detection, and therefore may be a promising approach for nitrate and phosphate detection. However, there are clear challenges as SERS sensing is negatively affected by interference in complex media and in situ sensing in a solution phase reduces accuracy and resolution. It is because of these challenges, in part, why much of the data reported in the literature are taken for purified samples that are then dried on a SERS substrate. Our goal is to address the engineering challenges for a SERS in situ seawater nutrient concentration measurement system. Batch and flow-through devices have been designed to incorporate commercially available, nanostructured gold SERS substrates. By benchmarking against 4-nitrobenzenethiol/ethanol solutions and ultrapure water spiked with nitrate and phosphate, our results show that our SERS devices can be used as a development platform for a seawater nutrient sensor, showing a route for a commercializable product that will greatly benefit scientific operations.
iii ACKNOWLEDGEMENTS First, I would like to thank Dr. Bothun for his mentoring, support and trust in me. I would also like to thank my lab mates and friends. You made feel welcome at URI from the first day and thanks to you I never regretted my decision to come to Rhode Island. I would especially like to thank Dr. Akram Abbasi, Robert Chevalier and Dr.
Buddini Karawdeniya for many fruitful discussions about SERS. I must also acknowledge the work my former SURF student Andrew White did on the design and testing of 3D-printed flow channels. Also, thank you to the EPSCoR office team -Bj, Sally and Shaun for always having my back and adding a little laughter to my days.
Lastly, I would like to thank my parents for supporting me in everything I do, no matter how far away I am. I wouldn't be where I am today, if it wasn't for you. Thank  Table 1: Characteristic 4-NBT peaks and their assigned modes as well as the molecular structure of 4-NBT. Printed bold is the main peak. Spectral data and mode assignments adapted from (Kim et al. 2003 (Daniel R. Lombardi et al. 1994;Rousseau et al. 1968;Waterland et al. 2001) and phosphate data was adapted from (S. K. Sharma et al. 2006;Toupry-Krauzman et al. 1979). Crystalline sodium nitrate mode assignments adapted from (Rousseau et al. 1968), 1 M sodium nitrate from (Waterland et al. 2001) and sodium phosphate assignments from (Toupry-Krauzman et al. 1979 or higher frequency hνaS (anti-Stokes scattering). Image adapted from (Kneipp et al. 2002

INTRODUCTION
The ocean plays an essential role for life as we know it on our planet. It is habitat for millions of species, acts as a key part in climate control by storing greenhouse gasses and functions as a heat sink. It has been utilized by humans for centuries to provide food, transportation and economic success by enabling trade and tourism (Glöckner et al. 2012;Paytan and McLaughlin 2007).
However, the ocean is a complicated ecosystem. Nitrogen and phosphorous are naturally occurring in ocean water and an essential part of the nutrient cycle. The presence of these nutrients promotes the growth of marine life (Paytan and McLaughlin 2007;Baturin 2003;Vitousek et al. 1997;Zehr and Ward 2002).
Human activity, such as the use of fertilizers, discharge from wastewater treatment facilities as well as burning of combustion fuel can change the nutrient balance and lead to elevated levels of nitrogen and phosphorous in the ocean.
Eutrophication is an effect describing increased plant growth in water bodies, caused by excess concentrations of phosphorous and nitrogen in ocean water. Cultural eutrophication describes eutrophication caused by human activity. If not addressed with countermeasures, especially nutrient reduction in affected water bodies, the composition of the plant life might become dominated by algae, which not only poses risk of toxin release, but can also result in oxygen depletion. Hypoxic or anoxic conditions mitigate fish populations and alter the water composition making it undesirable for utilization by humans (Smith and Schindler 2009;Conley et al. 2009).
Because of the challenges associated with elevated nutrient concentrations in ocean water, affordable, frequent, accurate and in situ detection of nitrogen and phosphorous is of great importance to provide data to improve computer models and enable early prediction and warning mechanisms for algae blooms, so that countermeasures can be taken and negative effects associated with eutrophication can be reduced.
Current measurement methods include detection via chemical reactions, which either necessitate the collection and transport of samples to a research laboratory, adding a lag in data availability and a potential change of the sample composition due to biological activity, or require the availability of laboratory space, bulky instruments and potentially toxic materials on marine vessels (Patey et al. 2008). UV-Vis spectroscopy is a commercially available technique used for the detection of nitrate. Unfortunately, UV-vis is negatively affected by interference of the combined signal of bromide and nitrate in seawater, necessitating skilled operators to interpret and advanced equipment to collect the data (Johnson et al. 2013).
A promising technique to overcome some of the challenges associated with the previously mentioned measurement techniques is surface enhanced Raman spectroscopy (SERS). SERS is described as a molecular fingerprinting technique and is capable of single molecular detection when ordered metallic nanostructured substrates are employed (Nie 1997 The goal of this work is to demonstrate in situ SERS detection of nitrate and phosphate in aqueous solutions using commercial and custom fabricated nanostructured gold substrates. In situ detection is a critical step to assessing the suitability of SERS as a platform for continuous, field-deployed in situ nitrate and phosphate measurements.
The goal was pursued through the following specific aims.
• Aim 1. Identify, for in-house designs fabricate, and characterize SERS active nanostructured substrates. SERS substrates were characterized by electron microscopy and benchmarked for SERS detection using 4-nitrobenzenethiol.
• Aim 2. Design and fabricate measurement devices for stationary and continuous in situ SERS detection. Devices were 3D-printed to physically secure SERS substrates and aid in the reproducibility of SERS measurements.
• Aim 3. Evaluate the detection performance of the SERS substrates for nitrate and phosphate in batch and continuous mode.  Decay processes consume oxygen, leading to hypoxic or even anoxic zones.
Furthermore, metabolism byproducts of certain algae and bacteria can be toxic for higher life forms such as fish or humans (Smith and Schindler 2009), lowering the habitability of affected zones and rendering them less attractive for human activity.
Taken to the extreme, affected water bodies might become inhabitable for higher life forms, leaving behind so called dead zones (Smith and Schindler 2009). Healthy water bodies act as reservoirs for greenhouse gases, with the formation of dead zones the natural balance is disturbed and the greenhouse gases are released into the atmosphere, potentially elevating climate change and leading to self-amplification effects by raising the water temperature and increasing the growth rate of algae (Glöckner et al. 2012;Paytan and McLaughlin 2007).

Currently applied detection techniques
For the reasons described in the previous chapter, monitoring of nitrate and phosphate concentrations in marine water bodies is desired. Patey et al. (Patey et al. 2008) give an overview of commonly used techniques for nanomolar detection techniques of nitrate, nitrite and phosphate in marine water.

Nitrate and nitrite detection methods
The most widely used method for nitrate detection in sea water is the reduction of nitrate to nitrite, followed by spectral analysis of the products of the Griess reaction: the formation of a highly colored dye through diazotization with sulfanilamide (SA) and coupling with N-(1-naphthyl)-ethylenediamine dihydrochloride (NED) in the presence of nitrite (Patey et al. 2008;Correa-Duarte et al. 2015). A schematic of this reaction is given in Figure 2. This method yields the total amount of nitrate and nitrite. Because the Griess reaction and its variations are specific to nitrite, the nitrate concentration can only be assessed through conducting the analysis twice: before and after the nitrate reduction step (Patey et al. 2008). Many variations of the Griess reaction have been developed over the years, with reported limits of detection ( = 3 • , being the standard deviation of the blank) as low as 1.5 nM for nitrate through a combination of the Griess reaction with segmented continuous flow analysis (SCFA) and a liquid waveguide capillary cell (LWCC) system (Patey et al. 2008). UV absorbance spectroscopy, as well as fluorescence spectroscopy and fluorescence quenching are additional methods used for nitrate detection. The reported limits of detection are between 6.9 (fluorescence spectroscopy) and 40 nM nitrate (UV absorbance spectroscopy) (Patey et al. 2008). UV absorbance spectroscopy has been successfully tested with 27 spatial profiling float systems deployed over 3 years at several ocean locations ranging from the subtropical ocean, over the Southern Ocean to the Arctic Ocean (Johnson et al. 2013). Johnson et al. reported a limit of detection of 0.4 µM for their system. Sensor drift and initial correction of the data was necessary to account for sensor variation (Johnson et al. 2013).
A potentiometric method with an ion-selective membrane permeable for nitrate was tested for monitoring of a river over two months. A limit of detection of 0.007 mg nitrate reported as nitrogen (~0.5 µM nitrate) was reported. The stability of the electrode was reported to be 5 months under laboratory conditions (Le Goff et al. 2003). as low as 0.8 nM for phosphate through a combination of the Molybdenum blue reaction with segmented continuous flow analysis (SCFA) and a liquid waveguide capillary cell (LWCC) system (Patey et al. 2008).

Phosphate detection methods
An extension to the molybdenum blue method is the nowadays widely used magnesium induced coprecipitation "MAGIC" method, developed by Karl and Tien (Karl and Tien 1992). Phosphate dissolved in the sample is pre-concentrated by sodium hydroxide induced precipitation of brucite (Mg(OH)2) from the solution and adsorption of the phosphate to the precipitate. The precipitate can be removed from the solution through centrifugation and is dissolved in acid and tested with the regular molybdenum blue method. Limits of detection as low as 0.2 nM can be achieved with this method (Patey et al. 2008). The nature of MAGIC requires large sample volumes of up to 250 mL and consists of several steps, possibly introducing contamination and making it challenging to automate (Patey et al. 2008).
Other methods utilize chemiluminescenceluminol (3-aminophtalhydrazide) emits blue light when oxidized. The method yields a strong signal and limits of detection comparable to the MAGIC method. It is not specific to phosphate and requires a preconcentration step (Patey et al. 2008).

Summary of findings
Our review of the available measurement methods shows that the detection of nitrate and phosphate at nanomolar concentrations is possible. Many of the described methods require either manual handling, extensive know-how for the data analysis, availability of reagents or large volumes of sample. Automatization of these methods is often difficult and requires additional instruments that might exceed the confined space on a marine vessel (Patey et al. 2008).
It is desirable to overcome these challenges to enable scientists to analyze water samples on research vessels without the need to store hazardous reagents that pose a risk of harming the environment. The availability of an easy to use, highly specific and cost efficient measurement method would enable researchers worldwide to collect more accurate data and would possibly allow for the early detection of excess nutrient concentrations to enable counter measures and to prevent or mitigate harm to humans and the environment. It is also obvious that the currently available measurement techniques are not sufficient to fulfill the requirements of such a sensor. A promising technique to overcome some of the most pressing challenges is Surface Enhanced Raman spectroscopy (SERS), which will be discussed in the following section.

Surface Enhanced Raman Spectroscopy
Raman scattering is an effect described by C.V. Raman in 1928 (Raman andKrishnan 1928). It is based on the inelastic scattering of incident light on molecules and results in a spectrum unique to each molecule and is therefore considered a molecular fingerprinting technique (Kneipp et al. 2002).  (Kneipp et al. 2002).

Figure 3: Schematic representation of Raman scattering. Incident photons hνL are inelastically scattered from molecules. The energy of the characteristic molecular vibrations hνM, results in scattered photons of lower frequency hνS (Stokes scattering) or higher frequency hνaS (anti-Stokes scattering). Image adapted from
Compared to effects like fluorescence, Raman scattering is a very weak effect with Raman cross sections being 12-14 orders of magnitude lower than fluorescence cross sections (Kneipp et al. 2002). Partly because of this it was neglected as a scientific tool for several decades (Li et al. 2015). An advantage of Raman spectroscopy compared to fluorescence spectroscopy is the higher resolution of the Raman peaks compared to the broad adsorption/emission bands observed in fluorescence spectroscopy (Mosier-Boss 2017).
It was the discovery of Surface Enhanced Raman Spectroscopy (SERS) in the 1970s with reported Raman signal enhancements of up to 10 6 (Fleischmann et al. 1974; Jeanmaire and van Duyne 1977; Albrecht and Creighton 1977) that made Raman spectroscopy more appealing to the scientific community. Nowadays signal enhancements as high as 10 15 can be achieved, allowing for single molecular detection (Nie 1997;Stiles et al. 2008).
With the discovery of SERS, utilization of Raman spectroscopy was investigated for a wide range of applications, such as in vivo detection of glucose levels in animal models (Stuart et al. 2006), as well as for explosive (Dasary et al. 2009) and drug abuse screening methods (Andreou et al. 2013). Monitoring of single molecular electrochemical processes was investigated (Cortés et al. 2010). SERS can be used for the monitoring of chemical reactions (Kundu et al. 2004) as well as for process and quality control in the food and pharmaceutical industry McNay et al. 2011) and the detection of pesticides used in agriculture (Pang et al. 2016). Product developments such as fiber and handheld analyzers allow for easy on-site application of the previously mentioned techniques (Lucotti and Zerbi 2007;. SERS is considered a non-destructive technique (Du et al. 2013).
The surface enhancement effect is most likely to occur in the presence of nanostructured noble metal surfaces ranging from 10 to 100 nm in size (Moskovits 2005).
SERS is a near-field effect that is strongest for analyte molecules adsorbed to the metal surface and scales with a factor of r -12 , with r being the distance between analyte molecule and surface (Stiles et al. 2008). The distance dependency was shown experimentally by measuring SER spectra of pyridine adsorbed on silver film over nanosphere substrates covered with aluminum oxide (Al2O3) multilayers of varying thickness. A decrease of the SERS signal intensity by a factor of ten for an increase of the distance r by 2.8 nm was observed in these experiments (Stiles et al. 2008).
Even though controversially discussed in the past (Moskovits 2005), two different mechanisms are believed to be responsible for the 10 6 signal enhancement that was historically observed for SERS -electromagnetic enhancement with a contribution of ~10 4 and chemical enhancement with a contribution of ~10 2 to the total enhancement (Stiles et al. 2008;Moskovits 2005).
The electromagnetic enhancement is caused by localized surface plasmon resonance (LSPR) on the surface of the involved nanostructures, an effect that is visualized in Figure 4. Localized surface plasmons can be described as collective oscillations of conducting electrons of ionic metal cores (labeled "Electron cloud" and "Metal sphere" in Figure 4) caused by interaction with an electromagnetic field, in this case an incident light beam. Even though other excitations are possible, dipolar plasmon resonance is dominantly observed for small structures between 10 to 100 nm in size (Moskovits 2005). When the incident light source resonates with localized surface plasmons a local dipolar radiation field is emitted from the surface of the nano-structure, which can then excite the electromagnetic field of the analyte molecule (Moskovits 2005;Stiles et al. 2008 with ISERS and I0 being the intensities of the incident and enhanced fields. At low-wavenumber bands g and g' become nearly identical, which allows for the simplification of above equation and yields the finding that | | 4 = | | 4 (Moskovits 2005).
The dipole radiation field can be generated on single structures, but is stronger in clusters of particles, with the inter-particle gaps allowing for the induction of enhanced electromagnetic fields as it is shown in Figure 5    Chemical enhancement is a much more general term and summarizes enhancement effects caused by processes such as charge transfer, that alter the electromagnetic field of the complex of nano-structured surface and analyte molecule in direct contact with each other (Moskovits 2005;Kneipp et al. 2002).
From the previously described underlying principles of SERS a series of attributes can be extracted that make a "good" SERS substrategood meaning in this context a high signal enhancement, a high reproducibility, robustness and a long lifetime. To achieve this the surface requires the presence of homogenously distributed hot spots over the surface and must show effective analyte adsorption. The substrate must show a high resistance to photodegradation. Furthermore the availability of a standard to monitor for an eventual time dependence of the measurements is desired (Mosier-Boss 2017).
Materials commonly used for chemical detection include noble metal nanoparticles in suspension or deposited on a surface, that allow for tuning of the SERS enhancement by changing the size and shape of the particles. The SERS signal increases with increasing particle size until the size approaches the scale of the wavelength. The signal increase can be explained by the higher number of available electrons, while the decrease in signal after reaching a critical particle size is caused by a shift of the particle excitation to non-radiative modes. The lower intensity at smaller particle sizes can be explained by a lowered conductivity of the particles as well as diminishing of the light scattering. The signal enhancement as a function of particle shape can be attributed to the increased availability of intrinsic SERS hotspots for particle shapes such as triangular or star-shaped structures (Mosier-Boss 2017).
Deposition of particles on a surface is possible through various chemical linking methods, application to filters, as well as embedding in paper matrices (Mosier-Boss 2017). It was shown that the effect is strongest for interparticle distances of 1 nm or less Another method to immobilize nanostructures on surfaces is the fabrication on the surface itself. Fabrication of highly ordered nanostructures can be achieved through nanolithography techniques such as nanosphere lithography (NSL) and electron-beam lithography (EBL). NSL describes a technique in which nanoparticles are brought into contact with a surface and are used as a template to form metal films. EBL utilizes the solubility change of photoresist on noble metal surfaces when exposed to an electron beam. The electron beam is manipulated to draw into the photoresists to generate nanostructures on the surface that can then be used for SERS sensing (Mosier-Boss 2017).
A variety of SERS substrates is commercially available. Since the description of all of them would exceed the scope of this report we refer to the discussion by Mosier-Boss (Mosier-Boss 2017). For this work two types of commercial substrates were tested. Gold particles incorporated in a paper matrix, distributed by Ocean Optics as "RAM-SERS-Au" (Ocean Optics) as well as gold coated silicon nanorods grown on a silicon wafer distributed as "SERStrate" by Silmeco (Silmeco ApS; Mosier-Boss 2017).
Common challenges with the application of SERS are the sensitivity of the system to inhomogeneities in the structure of the SERS substrate as well as of the inhomogeneities in the concentration of analyte over the surface. When drying after drop-casting without taking special measures, convective forces in the droplet lead to the accumulation of particles or analyte molecules on the edge of the droplet and leave behind stains after the drying process is completed. This process is described as "coffee stain effect" (Deegan et al. 1997) Since SERS is extremely sensitive, these differences in local concentration heavily influence the signal. Several approaches exist to overcome these challengesincreasing the homogeneity of the surface as well as the distribution of the analyte above the surface and increasing the measured area by moving the laser or the substrate are common strategies to achieve this (Moskovits 2005; Mosier-Boss 2017). Another possibility is to conduct measurements in situ. Measuring in situ leads to a more homogenous distribution of molecules in the system but reduces the number of molecules in SERS active proximity to the surface. It can also add a diffusion limitation to the system. This results in a lower overall signal strength, but drastically increases the handling and automatization capabilities of SERS (White et al. 2012 Ultrapure water was prepared with a MilliPore Milli Q3 UV system. Solutions were stored in vials sealed with Parafilm® at room temperature. The resistivity of the water was >18.2 MΩ.
Paper based gold nanoparticle SERS substrates marketed as "RAM-SERS-Au" were purchased from Ocean Optics and used as received. Gold sputtered silicon nanorods grown on a silicon wafer, marketed as "SERStrate Au" were purchased from SILMECO® and used as received. Whatman® Anodisc™ 13 -nano porous aluminum oxide discs with pore diameters of 20 and 200 nm and a disc diameter of 13 mm were purchased through Fisher Scientific and sputtered with gold to fabricate SERS substrates.

Equipment being used
Most Raman spectra shown in this paper were collected with one of the two following instruments. Initial Measurements were taken with a Raman Systems

a) b)
The experiments were started with varying concentrations of 4-NBT in ethanol as a benchmark. 4-NBT was chosen because the thiol group binds to the gold surface and therefore a prominent SERS signal can be observed. After confirming the measurement principle measurements with nitrate and phosphate solutions were conducted.

Normal Raman measurements
The

In situ batch measurements
All measurements in in situ batch mode were conducted in the bottom to top measurement mode. Commercially available Silmeco® "SERStrate" substrates were tested and used as received. were carefully assembled and placed under the laser, using the grey 3D printed spacing system shown in Figure 9 to ensure that the same spot on the substrate was hit for every measurement. After assembly the laser was focused on the substrate and the background of the dry substrate was measured in triplicate and averaged. The system was covered with an ambient light blocking laser protection housing. Background spectra were collected and automatically subtracted with the instrument software for every spectrum. Spectra for each integration time were collected in triplicate and averaged. The system was then filled with 0.75 mL of pure solvent (ethanol or ultrapure water) and the laser was refocused to account for the focal change caused by the introduced media.
Measurements were started as soon as the laser was focused. Each concentration was measured either every two or 5 minutes for up to one hour. Due to evaporation of ethanol the 4-NBT measurements had to be stopped after approximately 20 minutes.
Concentrations were changed by pouring the solution out of the beaker and rinsing it three times with 0.75 mL of solvent. The beaker was then filled a fourth time with solvent and spectra of the cleaned substrate in solution were taken in triplicate to investigate the cleaning capabilities. After measuring, the fourth solution was poured out, and the beaker was re-filled with 0.75 mL of solution of the next higher concentration. The measurement procedure was repeated until the highest concentrated solution was reached. After each experiment the substrates and substrate holders were dried, moved to a petri dish and sealed with parafilm. The so prepared systems were stored at room temperature in the dark for future analysis.

In situ continuous flow measurements
A flow channel was designed and manufactured using 3D-printing technology.
The system is shown in Figure 10 and consists of a 525 µL flow chamber with a fit for the laser lens of the Raman spectrometer. Silmeco® SERStrate substrates can be mounted inside the channel. Lens and fluid are separated through a microscope cover glass slip. The device has connectors for 1/16" tubing.
A picture of the assembled system without the Raman spectrometer is shown in Figure 10. The opening that can be seen on top of the system fits the lens of the Snowy Range SIERRA 2.0 instrument and ensures that the distance between lens and substrate is kept constant as well as that ambient light is blocked.
Triplicates of dry spectra were collected before each run as before with the batch system. To test the measurement capabilities under continuous flow, liquid was pumped through the system at a rate of 0.18 mL/min, using a peristaltic pump, resulting in a Each measurement series was started by filling the device with solvent (ethanol or ultrapure water). The concentration was increased by switching the flow channel inlet to the next higher concentration. After testing the highest concentration, solvent was pumped through the device for 60 minutes, to clean the system. Spectra were continuously collected with the same parameters as before while flushing the device.
After the flushing step the devices were emptied, dried and the inlets and outlets were sealed with parafilm and stored at room temperature for future analysis.

Baselining
Raman spectra were baselined using the TBB Baseline method, a polynomial fit method implemented in the "PEAK" software distributed with the Snowy Range instrument. The sensitivity of the method was left at the standard parameter of 115 out of 1000. Baselined spectra were exported to Microsoft® Excel for further analysis. Figure 11 shows the comparison of a raw spectrum of 1 mM 4-NBT collected in the in situ batch measurement setup. The baselined data was manually shifted by a value of 50,000 a.u. to allow for a better comparison of the spectral features. The figure shows the same signature peaks for both data sets, but the background is straightened, meaning that the method produced reliable baselining results. Throughout our data analysis this finding was consistent.

Standard normal variate
Baselined data was normalized as necessary by standard normal variate method, described in (Gautam et al. 2015), according to the following equation: Where ( ) describes the standard normal variate modified peak intensity at a given wavenumber x. ̅ represents the average intensity of the entire spectrum and σ stands for the standard deviation. The method can produce negative values for the baselined spectra. In case of the presence of negative normalized intensities, the entire spectrum was manually shifted to zero, to allow for the comparison of normalized intensities by a common starting point.

Time dependency
The individual spectra of a concentration series were investigated for the presence of the characteristic peaks of the analyte of interest. If characteristic peaks were present the peak intensity was plotted as a function of time. For data sets showing no time dependency the individual spectra were averaged over the entire time series, resulting in a single spectrum of averaged intensities. For data sets showing a time dependency only intensities after reaching a steady state were used for the averaging procedure. Peak intensities for the characteristic peak of interest were extracted and plotted as a function of concentration. Correlation functions were applied to allow for the quantification of analyte concentrations in unknown solutions. This was done for spectra before and after normalization.

Discussion of 4-NBT measurements
The following discussion of 4-NBT measurement results follows the individual experiments of the experimental procedure and will guide the reader through the various process steps of the method evaluation. The discussion will begin in section 4.1.1 with transmittance cuvette Raman measurements of 4-NBT to confirm the characteristic peaks shown in Table 1 (Kim et al. 2003).

Normal Raman measurements
The confirmation of literature values was conducted by transmittance cuvette measurements as described in section 3.3.1, p. 22. The results are shown in Figure 12.
The expected peaks of 4-NBT are highlighted with arrows. The characteristic 4-NBT peaks are in good agreement with the literature (Kim et al. 2003), but very small in comparison to the ethanol peaks. This means that 4-NBT can be detected with our instrument and is therefore a suitable benchmarking molecule for our purposes.
However, the circled peak at 879 cm -1 is the main ethanol peak and is cut off due to the limitation of the photodetector to 60,000 counts, meaning that the detectability in the absence of a SERS substrate is limited with our instrument. The dominance of the ethanol signal limits the use of longer integration times in solution, because of overcompensation of the 4-NBT peaks. As the thiol group of 4-NBT binds to gold the detectability is suspected to be higher using SERS. This will be shown in the following section 4.1.2.

In situ batch measurements
After confirming that the detection of 4-NBT using Raman spectroscopy is possible, testing of in situ batch measurements was conducted with the 3D printed device, discussed earlier and shown in Figure 8.  Table 2.

In situ continuous flow measurements
We believe the use of flow channels can support the diffusion of analyte molecules to the laminar boundary layer on the surface of a SERS substrate, leading to a higher quality signal. To test the influence of flow on the measurement system the experiments discussed here were conducted. The general functionality of the system was assessed with 4-NBT in a similar way as before for the in situ batch system. The 4-NBT concentration in the inlet was kept constant and only changed after the chosen measurement time for a given concentration. Due to binding of 4-NBT molecules to the surface the actual 4-NBT concentration that was measured was different from the concentration in the inlet stream. As the substrate was not changed in-between concentration measurements, 4-NBT from the previous experiment was already attached to the substrate surface, further altering the measured concentration.
Nonetheless an increase of inlet concentration was expected to show an increase in signal strength.
The flow channel was filled with pure ethanol, while SERS spectra were collected. An overview of representative averaged spectra at various 4-NBT concentrations is given in Figure 15.  To gain an understanding of the system response to increasing concentrations, an event that will be observed in an on-site application such as the deployment of a measurement device on a buoy, a plot of the intensity of the main peak at 1331 cm -1 over the entire experimental time is shown in Figure 16. The ideal residence time of 2.9 minutes of the device is represented by the labeled bar on the right side of the graph. on the type of molecule, the medium as well as the temperature. As all these parameters are close to constant in our system the diffusion coefficient must be close to constant, too. As the concentration gradient is increased in the system by increasing the concentration at the system inlet, the diffusion rate from the stream into the boundary layer increases too. The concentration versus time curve for 100 µM 4-NBT in ethanol shows a decrease of the growth rate with increasing measurement time, indicating that the system was approaching an equilibrium state for the given parameters. This observation indicates that the concentration gradient at an inlet concentration of 100 µM 4-NBT was high enough to force diffusion of 4-NBT molecules in measurable concentrations through the boundary layer. Increasing the concentration further leads to a signal increase over time. The sudden decrease in signal that can be observed around 180 minutes is related to air bubbles that had been introduced to the system while the concentrations were changed. By removing the air bubble, the signal could be regained.
Switching the inlet to pure ethanol to flush the system at the end of the measurement series leads to a lowered signal increase speed.
Plotting the normalized averaged peak intensity after 11 minutes of the most dominant 4-NBT peak at 1331 cm -1 against the concentration at the inlet, leads to the plot shown in Figure 17. The standard deviation for all concentration series is low and is overlapped by the plot markers. The data shows a power law correlation spanning the concentration range from 10 nM to 1 mM, with an R 2 -value of 0.98. As the system was not in equilibrium the power law correlation most likely results from the diffusion and chemisorption processes between incoming solution and the SERS substrate. A limit of detection calculation is not reasonable, as the system was not in equilibrium.   Table 2, p. 35 and the averaged normalized intensity data shown in Figure 17 were used to calculate the surface concentrations in the flow channel. The results of these calculations are presented in Figure 18 and show the relation between the inlet concentration and the concentration at the surface of the substrate. The dotted 45° line represents the idealized case that the inlet concentration equals the surface concentration. Examining Figure 18 shows that for low 4-NBT concentrations of 10 and 100 nM the surface concentration is close to the ideal case. As expected, the surface concentration increases with increasing inlet concentration. It can also be seen that the surface concentration increases much slower than the inlet concentration, which indicates that a transport limitation of 4-NBT molecules to the surface must be present.

Limit of detection
The 1331 cm -1 peak intensity versus 4-NBT concentration plot of in situ batch measurements presented in section 4.1.2, Figure 14, showed a clear correlation and can therefore be used to carry out a limit of detection calculation according to the procedure described in section 3.4.4. The data follows a linear trend of the form 0.0002 • + 0.0014 and can therefore be used for the limit of detection calculation without further modification. Using the Excel line estimation function "linest" yields the standard deviation of the y-intercept b to be = 0.0085 and therefore the limit of detection for the in situ batch measurement device is LOD = 111 nM 4-NBT.
The continuous in situ measurement data presented in Figure 17 was not in equilibrium and would therefore not yield a reasonable result for a limit of detection calculation.

Detection of nitrate and phosphate
After showing in the previous section 4.1 that the developed measurement system works with 4-NBT solutions, the evaluation of aqueous nitrate and phosphate solutions using SERS will be evaluated following the same procedure as before for 4-NBT in ethanol.   (Daniel R. Lombardi et al. 1994;Rousseau et al. 1968;Waterland et al. 2001) and phosphate data was adapted from (S. K. Sharma et al. 2006;Toupry-Krauzman et al. 1979). Crystalline sodium nitrate mode assignments adapted from (Rousseau et al. 1968), 1 M sodium nitrate from (Waterland et al. 2001) and sodium phosphate assignments from (Toupry-Krauzman et al. 1979).

Normal Raman measurements
In order to quantify the detectability of nitrate and phosphate with our SERS measurement setup, the peak locations of both analytes must be known. Figure 19 shows

In situ batch measurements
The hydrophobic nature and the low weight of the SERStrate substrates, that were used for the investigation of the in situ SERS detectability of nitrate and phosphate made it challenging to submerge them in aqueous solutions, as it is illustrated in Figure   20 a). Because of this, the previously described and in Figure 20   to the more dominant features of the spectra, therefore a close-up of the data range of interest is shown. The inset shows the full spectra for comparison. A peak shift from 1047 cm -1 to 1079 cm -1 was observed. This behavior is in good agreement with the literature (Daniel R. Lombardi et al. 1994). The peak intensity at 1079 cm -1 increases with increasing nitrate concentration.          that the sensor became saturated in between these two concentrations.
We were unable to detect phosphate with the Raman Systems R3000QE, therefore no comparison can be made. In summary it can be stated that nitrate as well as phosphate were detected using in situ SERS measurements. Nitrate was detected with both instruments. Phosphate was detected with the SIERRA 2.0 but not with the R3000QE.
One concern with in situ SERS measurements is the overall higher distance of analyte molecules from the SERS substrate, as compared to measurements of analyte dried out on SERS substrates. 4-NBT shows a chemical affinity to gold surfaces through its thiol group, a mechanism that is not present for nitrate and phosphate, meaning that there is no specific affinity of these molecules to the surface. Figure 28  SIERRA 2.0. The data is shown normalized to account for slight variations in substrate composition as well as instrument focus and to increase the comparability. Comparing the three datasets, the normalized intensity of the characteristic 4-NBT peak at 1331 cm -1 is higher than the characteristic peak intensities of nitrate and phosphate for concentrations of 100 nM and above. The chosen standard normal variate method is influenced by the noise within the data series, with higher values meaning less noise.
This means that the signal of 4-NBT is less noisy than nitrate and phosphate which can be explained by 4-NBT molecules binding to the surface, while nitrate and phosphate do not bind to the surface and therefore result in a noisier signal.
These results indicate that it is reasonable to attempt continuous in situ measuring of aqueous nitrate and phosphate solutions in a 3D-printed flow channel.

In situ continuous flow measurements
Experiments with in situ SERS measurements of nitrate and phosphate in 3D-printed flow channels of the form previously shown in Figure 10, were conducted.
The functionality of the system was shown before by benchmarking with 4-NBT in    The experiment was conducted in the same way with solutions of phosphate in ultrapure water. Representative spectra of the experimental findings are shown in Figure   31. As the intensity of the characteristic peak, previously identified to be around 989 cm -1 , was very small a closeup of the area of interest is shown in the bottom of the figure. It can be observed that the water filling step, labeled as "water (fill)" shows a lower signal strength than the other presented spectra. This becomes particularly clear by examination of the peak associated with the optical filter of the system between 200 and 400 cm -1 . The spectrum collected while filling with water yielded an average filter peak intensity of ~2000, while the other spectra showed intensities of this peak much closer to ~15,500. It is possible that the flow channel still contained small amounts of air, changing the laser focus and therefore influencing the signal strength. Figure 31 also shows a small peak at a wavenumber of 997 cm -1 . This finding is in good agreement with the previous in situ batch mode phosphate measurements that showed a peak at 3D-printed in situ batch system measured with the SIERRA 2.0 instrument, presented in Figure 23, followed a linear trend for concentrations from 22 to 332 nM nitrate. In order to show the entire span of concentrations, the data was presented as a semi-log plot. This means that the signal intensity for water had to be excluded from Figure 23.  the Excel line estimation function "linest" yields the standard deviation of the yintercept b to be = 0.029 and therefore the limit of detection for the in situ batch measurement device is LOD = 30 nM nitrate in ultrapure water. Because of the lack of signal in the continuous in situ SERS measurements no limit of detection for these measurements could be obtained.
The limits of detection for nitrate calculated for both instruments in the in situ batch setup are very similar, indicating that the calculated limits of detection are independent of the tested instruments and can be attributed to the tested devices. A calculation of the limit of in situ detection for phosphate was conducted as before for nitrate. It was discussed earlier that only in situ phosphate measurements conducted with the Sierra 2.0 showed a signal. Figure   Building on the results presented in this report, additional work will be conducted in the future with ocean water samples to answer the question if in situ measurements can be achieved in the field. It is expected that the overall signal quality will be impaired by the complex composition of sea water. The possible presence of other contaminants in high concentrations might overshadow the weak nitrate and phosphate signals. At the same time the presence of naturally occurring ions such as bromide will interfere with the signal too.
Microfluidic flow channels are currently investigated in our lab to reduce the influence of external factors and to improve the nitrate and phosphate signal by forcing analyte molecules on the SERS substrate surface. The use of flow channels will also allow to test for the re-usability of substrates and their long term stability. In the meantime, the substrates that are being used need optimization. The affinity for the analytes nitrate and phosphate needs to be increased and more work is needed to increase the reproducibility of the measurements. This could be achieved through the deposition of reporter molecules on the surface. One possibility that is currently investigated is the deposition of modified azo dyes on substrate surfaces.

APPENDIX A
A-1 Initial substrate evaluation The initially conducted evaluation of substrate performance is described in the following subsections.

A-1.1 Experimental procedure
Four different types of substrates were initially tested. In house designs consisted of concentrated spiky gold nanoparticles drop casted on glass substrates and gold sputtered nanoporous aluminum oxide filters (Whatman Anodisc). Commercially available substrates tested were gold sputtered silicon oxide nanorods grown on a silicon wafer, provided by Silmeco and sold as SERStrate, as well as gold particles embedded in a paper matrix, provided by Ocean Optics as "RAM SERS Au".
Drop deposition measurements were conducted as following with the Raman Systems R3000QE. Measurement parameters were a laser power of 250 mW for all tested substrates except the paper-based Ocean Optics substrates, which were tested at 57 mW to minimize substrate burning. The integration time for all measurements in this series was 30 s. The instrument was focused by measuring multiple times until the highest overall signal strength was achieved. After focusing, three dry measurements were taken. After that 5 µL of analyte were deposited on the substrate and measured every two minutes until dried. These experiments were conducted with increasing 4-NBT concentrations ranging from 5 nM 4-NBT to 25 µM 4-NBT. For the Ocean Optics substrates 10 µL instead of 5 µL were used to account for volume losses through soaking of the cellulose matrix.

A-1.2 Results
The graphs presented in Figure 35 to Figure   Additional measurements were taken with gold nanoparticles embedded in a paper matrix, provided by Ocean Optics. The results of these measurements are shown in Figure 37. For these substrates only a weak signal intensity of 1200 was observed.
Peaks can be observed at 1573 cm -1 and 1331 cm -1 , but no peak is present at 1079 cm -1 .
In addition to the lack of signal the paper-based substrate got damaged through laser burning, even though the laser intensity had been reduced to account for this. These observations lead to the decision not to proceed with the investigation of Ocean Optics paper-based substrates. In comparison the SILMECO SERStrate shows the highest 4-NBT signal in this study, with an intensity as high as 27,000 counts for a 4-NBT concentration of 25 µM for the peak at a wavenumber of 1331 cm -1 .
The SEM image of a SERStrate shown in Figure 39 reveals a uniform distribution of surface features. The observed structure was consistent over the surface and with all investigated SERStrate substrates.
Because of these observations and the suspected higher stability for in situ applications, due to the nanorods being fixed to the surface and unlike deposited