Zeolites represent a cornerstone of today’s industry as the most-used catalyst by weight in the world. Zeolites are nanoporous, crystalline, alumino-silicate materials with regular arrays of molecule-sized nanopores, giving rise to shape- and size-selective adsorption, diffusion, and reaction of adsorbed guest molecules. Synthetic zeolites can be fabricated as pure silica (SiO

) polymorphs, alumino-silicates, alumino-phosphates, or include other elements such as boron, germanium, tin, titanium, zinc, or zirconium. Because of this fabrication, a host of new, energy-relevant applications, such as biomass conversion to biofuels, carbon capture, hydrogen purification, and energy-efficient separations, can be realized with targeted zeolite architectures. Zeolite structure is often thought of in terms of a collection of rings that make up zeolite channels and cavities. Achieving the synthesis of tailor-made zeolites, with targeted rings and cavities, is greatly hampered by a poor understanding of how zeolite crystals actually form in solution. Solving this problem by identifying critical structures that lead to zeolite crystals has presented a difficult analytical challenge. Recently, we spoke to Scott M. Auerbach a professor of chemistry and chemical engineering at the University of Massachusetts Amherst, about his work in the use of Raman spectroscopy for zeolite structure characterization, focusing on a collection of nine all-silica zeolites.

From both theory and experience, it is understood that the precursors for zeolites are not sufficiently well characterized using infrared (IR), nuclear magnetic resonance (NMR) or X-ray diffraction (XRD). In researching the analysis of zeolites, you have recently recognized Raman spectroscopy as an analytical tool capable of understanding the formation of zeolites, especially their individual ring structures (1). Is it possible to assign Raman spectral features to individual ring structures, or should assigned Raman spectral features be associated with other zeolite building units?

Since the pioneering work in the 1980s by Prabir Dutta and others, there is a history of assigning Raman spectra of zeolites to individual rings, especially for Raman bands in the 200–700 cm

 range. This is perhaps not surprising, given that Raman spectroscopy can be sensitive to collective, nonpolar vibrations, such as those of silica-based rings. This zeolite Raman assignment approach arises from observed correlations between known rings in zeolites and the locations of Raman bands. For example, zeolites with so-called “four-membered rings” (four alternating pairs of ...-Si-O-... atoms in a ring) are known to exhibit a strong Raman band around 450–550 cm

, while zeolites with six membered rings show Raman bands around 300–350cm

, and those with eight membered rings show Raman bands at even lower wavenumbers, around 200–250 cm

The apparent anticorrelation between ring size and Raman wavenumber also makes sense, because frequency is expected to decrease with increasing wavelength, and ring size should be related to vibrational wavelength. So, in the end, there is logic to assigning Raman bands to rings. However, the problem is that these rings exist in environments of other rings, raising the question, “How do nearby rings influence Raman excitations?” We thought to ask this question because of zeolites like Linde Type A (LTA) and Chabazite framework (CHA), which have 4-rings, 6-rings, and 8-rings (in different arrangements), but exhibit different Raman spectra. We wondered why. Answering that question was the original reason for our study.

How is your Raman work on zeolite structural characterization different from what others have done previously to study these structures?

The innovative aspect of our approach is that we engaged in an integrated synthesis, spectroscopy, and simulation study to address two big problems that exist in the previous literature. The first problem is the difficulty in comparing properties for different zeolites, which can exhibit distinct framework structures, and contain different compositions, such as different silicon:aluminum ratios in alumino-silicate frameworks. As such, if Raman spectra for different zeolites were distinct, it’s not obvious why, because we typically have to change two variables at a time, and, in science, we’re not supposed to do that. The second problem is the difficulty in assigning Raman spectra, especially by first principles calculations, because of the large sizes of typical zeolite unit cells. We tackled both problems by identifying nine distinct zeolites all sharing the same properties: 1) they could be synthesized as all-silica materials, which removes ambiguities associated with locations of aluminum atoms in the zeolite framework, and locations of charge-compensating cations in zeolite pores; 2) they have unit cells small enough to be treated accurately with periodic density functional theory, including the calculation of Raman intensities; and 3) they all share a common structural feature—the zeolite four-membered ring to determine if Raman bands associated with four-membered ring vibrations are sensitive to different local environments or not. Studying this disparate collection of zeolites was made possible by our zeolite synthesis capabilities, allowing us to consider relatively novel zeolite materials that are not commercially available. In the end, our approach allowed us to make systematic comparisons among Raman spectra of different zeolite frameworks in ways that others have not.

You and your coworkers have approached the question of zeolite structure using such methods as integrated synthesis, Raman spectroscopy, and periodic density functional theory (DFT) modeling. How do these methods complement one another in understanding zeolite structures? 

Insights in materials chemistry often follow the central dogma that structure explains properties. In the case of our study, structure means a given zeolite’s framework structure with its collections of interlocking rings and cages, which after synthesis is tested and confirmed by comparing mea- sured XRD patterns to known standards. Then, property means the corresponding Raman spectra, which were collected with our colleagues at Worcester Polytechnic Institute (WPI). As mentioned above, using zeolite structure (rings) to explain properties (Raman band locations) has, in the past, required assumptions, and lacked accurate, first principles theory. We have bridged that gap in our study by applying periodic density functional theory on zeolite unit cells in a three-step process: 1) optimizing atomic positions and lattice parameters; 2) computing normal mode frequencies and vibrational coordinates; and 3) computing dielectric susceptibilities that underlie the Raman intensities. As such, the theory provides both accurate Raman spectra and illustrative pictures of key vibrations—the normal modes, thus answering whether bands can indeed by assigned to individual rings.

What are the larger questions associated with the accuracy of the assignments of Raman spectra related to zeolite structure? What are the aspects of these spectral assignments that are not well understood?

Zeolite structures consist of “secondary building units,” such as four-membered rings, and “composite building units,” such as small and large cages. The big question is whether Raman bands can be assigned to rings, cages, or both, depending on the given zeolite. This is a burning question, because, after all these years in zeolite science, we still don’t really understand how zeolites form small crystals that grow into larger ones. Discovering this could help us synthesize new zeolites for advanced applications in, for example, creating carbon-neutral biofuels and capturing carbon dioxide. That’s why we’re funded by the Department of Energy.

Two diametrically opposed theories for how zeolites form are 1) random monomer addition, like a sand castle forming from random addition of grains of sand, and 2) hierarchical growth from populations of pre-formed rings and cages in solution, like a sand castle forming from pre-formed buckets of sand. The former seems rather unlikely, but we don’t really have evidence for the latter yet. Being able to assign Raman spectra to rings and cages could answer this grand challenge question in materials chemistry, which is important for the reasons stated above.

Using normal-mode analysis, you discovered that Raman bands can be assigned to tricyclic bridges (which are structures comprised of three zeolite rings that share common Si−O−Si

bridges). You also found that the vibrational frequency of a given Raman band can be correlated to the smallest ring of its tricyclic bridge, and not to the ring that is actually vibrating. Lastly, you have discovered a precise anticorrelation between Raman frequency and Si−O−Si angle. How do these discoveries reveal new ways to investigate structures like those found in zeolites? 

The new structural concept of the tricyclic bridge offers a new organizing principle for understanding Raman bands from different, though related, materials. Tricyclic bridges are both building blocks for understanding zeolite structure, and objects that collectively vibrate during Raman scattering. Grouping normal modes by the smallest ring of the corresponding tricyclic bridge allows us to categorize and structurally correlate vibrations with similar but slightly different band locations, and also explain the reasons for outliers that don’t obey a given correlation. The angular correlation you mention allows us to identify specific molecular coordinates (the Si-O-Si angle) from Raman band locations. In the end, it’s like finding the key to unlock the treasure of zeolite structure, elucidated by Raman spectroscopy.

How do you properly sample materials like zeolites to understand their general structural properties?

The key to accurate calculations is optimizing both geometrical and electronic structures. The key to a good geometry is a good first guess, and that is furnished by previous experimental XRD studies on zeolite crystals. When we optimize atomic locations and unit cell parameters, we start from experimental data, and make small tweaks consistent with quantum density functional theory (DFT). The Nobel Prize in Chemistry was given to Walter Kohn in 1998 for DFT, which has revolutionized computational materials chemistry. DFT is tractable for zeolites because of the use of “plane waves,” which are great mathematical functions for modeling valence electrons in crystals such as zeolites. The only problem with zeolites is their typically large unit cells, meaning many atoms, many valence electrons, and thus, many functions in the basis set. That’s why we had to focus on materials with smaller unit cells to make calculations tractable.

What were the key challenges you encountered during this research work? Each component of the project—synthesis, spectroscopy, and modeling—had its unique challenges. Synthesizing all-silica materials often requires an esoteric organic molecule to act as a sacrificial “structure directing agent” (SDA). Before our team could actually synthesize a given zeolite, we had to do the multistep organic synthesis of a certain SDA, and that could take weeks. The spectroscopy was challenging to get good Raman signals, we needed to “hunt” for zeolite crystallites with an orientation that happens to give good back-scattering (our WPI collaborator, Dr. Geoff Tompsett, has Jedi skills when it comes to backscattering), and good signal-to-noise (S/N) requires plenty of time to collect and average over spectra. Finally, the quantum calculations of Raman intensities were, by far, the most computationally intensive aspect of our work. Our computations were performed on fast computers at the Massachusetts Green High Performance Computing Center in Holyoke, Massachusetts. In the end, the graduate students on this project, Tongkun Wang and Song Luo, showed heroic resilience and resourceful- ness, leading to our breakthrough findings.

What do you consider to be the most useful contribution of your work? 

Actually I see two: one about science, and the other about scientists. The science contribution is about the discovery of a new “Platonic form.” In other words, a new fundamental shape, the tricylic bridge, that helps us understand zeolite structure. The trycyclic bridge is an interesting shape, involving three rings that share one edge. I can’t help but wonder where else in nature we might find the tricyclic bridge. The contribution related to scientists is about how our integrated science approach made a real difference. By crossing disciplinary boundaries in our team, we taught one another about our respective fields, creating fertile ground for new ideas. I am the Executive Director of an integrated science program for science, technology, engineering, and math (STEM) undergraduates at UMass Amherst called the UMass iCons Program (icons.cns.umass.edu), that gave me a head start in launching and running a research program aimed at creating breakthroughs through “radical synergies.”

We’d like to connect the dots between Raman spectra of perfect zeolites (our recent 

 paper [1]), to Raman spectra during zeolite formation, to shed light on how zeolites form. To do this, we have to consider what’s in the soup that makes zeolites, the signature vibrations of each of these ingredients, and how interactions among these ingredients can be probed by Raman spectroscopy. Step by step, we will cross the chasm between zeolite structure to zeolite formation with the nexus among zeolite synthesis–Raman spectroscopy–quantum modeling.

1. T. Wang, S. Luo, G.A. Tompsett, M.T. Timko, W. Fan, and S.M. Auerbach, 

 is a professor of physical chemistry at the University of Massachusetts at Amherst (UMass Amherst). Professor Auerbach’s research focuses on modeling nano-structured materials such as zeolites, which are of importance to renewable energy technologies including biofuels and fuel cells, leading to 2 books and 120 articles.

Articles

Spectroscopy Spotlight

Zeolites are the most-used catalyst in industry. Synthesizing tailor-made zeolites is hampered by a poor understanding of how zeolite crystals actually form in solution. Scott M. Auerbach of the University of Massachusetts at Amherst is addressing this challenge with Raman spectroscopy.

Using Raman Spectroscopy for the Characterization of Zeolite Crystals

To enhance the process analysis robustness of an ethylene analyzer based on tunable diode laser absorption spectroscopy (TDLAS), an extreme learning machine (ELM) model is tailored through genetic algorithm (GA) parameter searching, activation function selection, and sample dataset rotation. Particularly, two robustness issues are the inaccurate gas concentration measurement caused by the variation in gas background components, and the inaccurate spectra shift calculation caused by spectral interference. Using our applied algorithm, the concentration calculation error was reduced by one order of magnitude. Additionally, it was shown that, with our algorithm, the wavelength shift calculation accuracy was improved to less than 1 count over 1000 counts spectra length with methane spectral interference. To test the viability, the algorithm was implemented in trace ethylene (C

) measurement with the coexistence of interfering methane, and the experimental measurements supported the robustness enhancement effect.

Being a highly-sensitive, selective, fast, non-destructive, and in situ method, tunable diode laser absorption spectroscopy (TDLAS) has gained popular acceptance for trace gas measurement in various industries. Its uses include, but are not limited to, coal mine safety (1), combustion measurement (2), petrochemical process analysis (3), and gas emission monitoring (4).

There are challenges, however, in maintaining analytical robustness of TDLAS methods for onsite application. The major common obstacles are spectral fringe (caused by the etalon effect of multiple re- flections from optical surfaces) and spectral interference (caused by background gas and spectra deformation due to system degradation). When these factors are combined in a complicated process analysis, they can cause a TDLAS analyzer to deviate from the original calibration.

Strategies to reduce the effects of spectral baseline structures include anti-reflection coatings, wedging or tilting the optical surfaces, dithering one surface and integrating the signal to average out its influence, and selecting a modulation frequency (5). Li introduced a discrete wavelet transform (DWT) method with optimized wavelet pairs to analyze spectra with different noise levels and baseline drift (6). Chemometric methods, such as the classical least squares (CLS) fitting method (7) and the partial least square (PLS) fitting method (8), have been commonly adopted in TDLAS to suppress spectral interferences.

When spectral interference or deformation becomes strong, it is hard to achieve satisfactory performance with linear regression algorithms. This changed with the introduction of machine learning algorithms such as artificial neural networks (ANN) (7), and support vector machines (SVM) (8) and extreme learning machine (ELM) models (9,10). ELM, a novel single-layer forward neuronetworking (SLFN) algorithm developed by Huang (11,12), aims to reach not only the smallest training error, but also the smallest norm of output weights. In addition, ELM offers better training speed and generalization ability than backprop (BP) neuronetworking and SVM (13).

ELM is capable of regression and classification (12,14), and has been used in micro-NIR, Fourier-transform infrared (FT-IR), and laser-induced breakdown spectroscopy (LIBS) (9,10). Application in TDLAS has not been carried out so far. Therefore, we introduce the ELM algorithm in a TDLAS process analyzer to improve its performance.

 is the number of the sample, L is the number of hidden layer nodes, 

 is the input layer neuron number, and the output layer neuron number is 1.

The goal of ELM model training is to obtain the estimated 

 by minimizing the difference between real value 

The first step in spectral analysis is preprocessing, which involves multi-scan averaging, Savitzky–Golay smoothing, and optical power normalization.

An ELM model can be optimized through several factors. The activation function is critical in terms of model convergence speed and predication accuracy (12). Recent research showed that the stability and accuracy of an ELM model were further improved by optimization of 

GA is useful for searching for the global optimized parameter and avoiding local optima, through many generations of selection, crossover, and mutation operation. GA is also ideal for 

The purpose of ELM application is to enhance the robustness of the spectroscopic analyzer in process analysis. The accuracy of key measurements is expected to improve after spectra restoration with ELM.

As the basis, CLS is introduced to derive the concentration from spectra:

 = 1...n) is the regression coefficient which is used to derive the gas concentration. Relative error (RE) is used as model performance indicator:

 are the prediction and real values of the 

 is the full range. In contrast to 3σ criterion, the 

) measurement in petrochemical process analysis was chosen as an example. There were strong spectral interferences from coexisting background components. The spectral region from 1626.25 nm to 1626.70 nm was chosen for C

The analyzer’s micro controller (MCU) controls a laser module PCI-FPGA-1A(Port City Instruments) to drive the laser with scan and modulation functions. The tuning range of the distributed feedback (DFB) laser is around 0.45 nm, which covers two C

 absorption peaks at 1626.35 and 1626.53 nm. The laser temperature is stabilized by a TEC module WTC3293. The cascade integrator comb filter–based demodulation output DC and 

The process of TDLAS analyzer gas concentration measurement is shown in Figure 1. The spectra generation process is illustrated in Figure 1a. The procedure of calibration model training and verification is illustrated in Figures 1b and 1c, where the impact on accuracy by laser frequency drift and spectral interference are demonstrated.

We will address two spectral analysis issues. The first is the gas concentration prediction accuracy caused by background and gas pressure variation during process measurement. The second is the inaccurate spectral shift calculation in the presence of strong spectral interference. With ELM, we aimed to accurately calculate both gas concentration and real spectra shift. Two experiments were designed correspondingly.

This experiment was to simulate the industrial process in a controlled scenario. Fifty mixing gas streams were generated with C

 up to 2000 ppm and the pressure was in 900 to 1100 mbar range. They were randomly varied within the range. Part of collected spectra are plotted in Figure 2a. Because of CH

 interference, accurate concentration cannot be obtained by direct counting of the C

Spectral drift causes an adverse impact on the accuracy. In the experiment, spectral peak shift was varied up to 16 counts over 1000 counts range by randomly shifting current up to ±0.2 mA without changing the range amplitude. The concentration of C

 was randomly set. Examples of collected spectra are shown in Figure 2b. For isolated spectral features, the spectral shift can be calculated by peak tracking of peak position, but it is unlikely to get correct spectral shift by tracking for this scenario shown in Figure 2b.

The 50 process test spectra are used to calculate the C

 concentration in dynamic backgrounds with ELM. First, we evaluate the effect of activation functions, with subgroups 1 to 4 as a training set and subgroup 5 as a verification set. The RE for the different activation functions, sigmoid, sin, hardlim, tribas, radbas and rectified linear unit (ReLU), are calculated. The 

 for ReLU is 0.65%, which is substantially better than others, because it has linear expression and broad linear response input range. Therefore, ReLU is chosen for future ELM modeling.

 are identified through GA global searching, 

Finally, the impact of training set and verification selection on model performance is evaluated through a fivefold cross-validation (CV). The model was trained by the four subgroups, and verified by one remaining subgroup. The subgroups are rotated in training and verification selection. The 

 of each subgroup with the corresponding ELM model is listed in Table I. The ELM model trained by subgroup 2–5 has the lowest RE of 0.27%.

The concentration error with the final optimized ELM modes for 50 spectra is plotted in Figure 3 (circle). The training set errors are closer to zero, because of deep suppression by the nonlinear activation function. The errors of 10 verification set spectra are larger, which are accountable for ELM model performance.

 concentration is obtained by CLS equation 4. The RE is listed in Table I, with max value of 6.30%. Errors for each sample are plotted in Figure 3 (triangle). The error is reduced more by ELM than by CLS.

The real spectral shift in the dynamic process is the key to obtaining correct concentration through the calibration model. With real shift, spectra can be restored to calibration status with high fidelity. The ELM algorithm is adopted for this purpose. During model training, spectral shift is treated as the output in the model. The 

 is defined as the current prediction error divided by the current shift range (

). The peak shift RE for subgroups is listed in Table II. The best ELM model, verified by subgroup 3, has an 

 of 4%, equivalent to a maximum spectral shift error of 0.64 counts. The right C

 peak is selected for peak tracking with the RE listed in Table II. Due to concentration variation and CH

 interference, its accuracy is much less than ELM. The max error is 16 counts (-100%).

Concentration Error After Backshift Correction

After the shift calculation, the spectra are backshifted to match with the calibration. The C

 concentration prediction error with and without backshift correction is calculated based on corresponding calibration model.

 concentration for each spectrum is plotted against current shift in Figure 4. Before correction, the error pattern demonstrates a V shape. As spectra shift more, the error increases more. After ELM correction, the error pattern demonstrates a flat belt shape, and the widths are greatly shrunk. The error is improved from 6.92% to 1.50% (Figure 4a). For peak tracking and CLS method, the error has no obvious improvement (Figure 4b), due to inaccurate peak shift calculation.

 TDLAS analysis through two experiments. In the process test, RE of C

 concentration is obtained at 0.27% for the optimized ELM model. For the traditional CLS method, it is 6.30%. In the spectral shift test, the spectral shift prediction errors are less than 1 count for ELM, and less than 16 counts for peak tracking. The concentration relative error was reduced from 6.92% to 1.50% after ELM backshift correction. For the peak tracking and CLS methods, however, the error remains in the 20% region. ELM has demonstrated a dramatic robustness enhancement in trace C

 measurement. It can be extended to more complicated nonlinear spectral deformation scenarios.

This work is supported by Shandong Provincial Key RD projects (No. 2019GGX104103, 2019GHY112084).

(1) Y. Wei, J. Chang, J. Lian, and T. Liu, 

(2) D. Shi, W. Song, J. Ye, B. Tao, and Y. Wang, 

(3) Y. Wang, Y. Wei, J. Chang, T. Zhang, T. Liu, T. Sun, and K.T.V Grattan, 

(4) Z.G. Yuan, X.T. Yang, W.Q. Xie, and X.M. Li, 

(7) H. Li, Y. Zhu, F. Dong, X. Li, and L. Zhang, 

(8) Y. Wang, Y. Wei, T. Liu, T. Sun, and K.T. Grattan, 

(9) B.T. Le, D. Xiao, Y.C. Mao, L. Song, D.K. He, and S.J. Liu, 

(11) G.B. Huang, Q.Y. Zhu, and C.K. Siew, 

(14) C. Malegori, E.J.N. Marques, S.T. de Freitas, M.F. Pimentel, C. Pasquini, and E. Casiraghi, 

(15) X.X. Yin, S. Hadjiloucas, J. He, Y. Zhang, Y. Wang, and D. Zhang, 

(16) K. Javed, R. Gouriveau, and N. Zerhouni, 

 are with the College of Control Science and Engineering at the China University of Petroleum, in Qingdao, People’s Republic of China. Direct correspondence to: liguolin@upc.edu.cn

Featured Article

Tunable diode laser absorption spectroscopy (TDLAS) is combined with an extreme learning machine (ELM) model, tailored by genetic algorithm (GA) parameter searching, to produce a more robust analytical method for trace gas analysis of ethylene.

A Genetic Algorithm–Optimized Extreme Learning Machine Model for Process Ethylene Analysis Robustness Enhancement

This article outlines the progress made at the United Kingdom’s Na- tional Physical Laboratory (NPL) in using inductively coupled plasma mass spectrometry (ICP-MS) for nanomaterial research using three case studies. First, we examine the measurement of metal impurities in graphene as part of the development of international standards. Then, we assess the uranium uptake of graphene oxide (GO) to optimize functionalization of the material for decontamination of aqueous nuclear waste. Finally, we establish the uptake mechanism of different cell and nanoparticle (NP) types to optimize radiation dosimetry assessments. Further advancements made in ICP–MS analysis to characterize single-particles (SP-ICP–MS) are also discussed with recommendations for future work.

The widespread use of nanomaterials has shown significant benefits in the areas of sensors, nuclear waste treatment, and nuclear medicine. This is evident because of the increasing number of publications citing the use of graphene, functionalized graphene oxide (GO) and high-Z nanoparticles (NPs) for these applications on account of their intrinsic material properties at the nanometer level (1,4,6). These properties include high specific surface area, increased chemical surface functionality, and more effective radiotherapy treatment, respectively. As a result, the growing use of these nanomaterials has driven a need to produce robust characterization protocols. Inductively coupled plasma–mass spectrometry (ICP-MS) is one of the characterization techniques available, and is being increasingly applied in nanomaterial research. The technique allows rapid multielemental information to be attained, as well as a thorough assessment of the materials performance. This paper discusses the use of ICP-MS across several groups at the National Physical Laboratory (NPL) for nanomaterial measurement to contribute towards material standardization, nuclear waste treatment, and nuclear medicine.

An Agilent 8800 Triple-Quadrupole instrument (ICP-QQQ-MS) was used for the analysis of graphene, functionalized GO, and commercial high-Z NPs. The instrument was equipped with two quadrupole mass filters separated by a collision-reaction cell. The sample introduction system was equipped with a glass concentric nebulizer, a quartz torch, a quartz spray chamber, standard nickel sampling, and skimmer cones. The main instrument operating parameters used for each study are listed in Table I. For all case studies, the instrument was tuned using a 1 μg/L multielement (beryllium, yttrium, cerium, or thallium) standard solution to ensure the instrument sensitivity, uncertainty, doubly charged and oxide formation rates, and peak axes were within the accepted ranges.

Each analyte of interest was measured using ICP-MS in single-quad (SQ) mode (that is, with only the second mass filter operating). For iron analysis, helium gas was introduced to the collision-reaction cell at a flow rate of 4 mL/min to remove the 

 polyatomic interference formed in the sample introduction system that otherwise overlaps with the 

Chemicals, Standards and Sample Preparation 

Graphene-based materials were obtained in powder form from Chinese industrial collaborators, and were sourced for the VAMAS TWA 41 Project 2 study (3,10). Carboxyl-functionalized graphene oxide (COOH-f GO) was prepared following the method outlined in the literature (9). The magnetic graphene oxide nanocomposite (MGO) was prepared by coupling (3-aminopropyl) triethoxysilane-coated (APTES)-functionalized iron oxide (Fe

) NPs to COOH-f GO by following the EDC.HCl/s-NHS crosslinking protocol (5). The interaction between the target radionuclide species and functionalized graphene oxide material was determined by suspending each prepared material in a diluted natural uranium stock solution (10 mg/L) and measuring the uptake difference by ICP-MS.

For radiation therapy treatment, three nanoparticles were investigated; freeze-dried spherical gold NPs (Aurovist) with a mean diameter of 1.9 ± 0.6 nm (Nano- probes Inc.) freeze-dried gadolinium-based nanoparticles (AGuIX) with a mean diameter of 3 ± 1.5 nm (NH TherAguix, France); and magnetic dextran iron oxide nanoparticles (RCL-01) in water with a mean (Z-average) diameter of 140 ± 4 nm and a polydispersity index of 0.25 ± 0.02 (Resonant Circuits Limited), stored at 4 °C. In the first experiment, two cell lines were investigated (MCF-7 human breast adenocarcinoma cell line and U87 human glioblastoma cell line) (1) for the three types of nanoparticles while a second experiment focused on the gadolinium product (Dotarem) using a panel of six different cancer cell lines (H460 [lung], DU145 [prostate], MiaPaCa2 [pancreas], MCF7 [breast], U87 [brain], and HEPG2 [liver]) (8,9). Both experiments aimed to relate the cell survival (clonogenic assay) and DNA damage (immunofluorescence assay) induced by radiation in cells loaded with differ- ent amount of nanoparticle.

Determination of Metal Impurities Present in Graphene

The growing industry uptake of graphene has driven a need to develop international standards that ensure end users have available robust and reproducible measurement protocols. Currently, NPL is involved in the preparation of several international graphene standards, which are focused on the measurement of the material properties of graphene and related two-dimensional materials through different characterization techniques, including Raman spectroscopy and X-ray photoelectron spectroscopy (XPS). However, there are a limited number of standardized methods available that specifically describe sample treatment and measurement of the chemical properties of graphene by ICP-MS. Because of the limited number of standardized methods available, an international interlaboratory VAMAS study was developed to address the limits currently faced with ICP-MS analysis of the metal impurities present in graphene (3). The validation of a suitable protocol and the quantification of the associated uncertainties will directly influence the development of the international standard ISO/PWI 23359 in ISO/TC229 “Nanotechnologies, for the chemical characterization of graphene,” that includes the application of ICP-MS. As such, the results from the inter- laboratory study will help form the basis for future standards.

For the initial work completed for this VAMAS study to analyze quantitatively species and concentrations of metal impurities contained in graphene samples, a one-step digestion protocol developed was effective in ensuring complete disintegration of certain graphene samples, with no solid material visible in the digest. ICP-MS analysis of the sample also indicated that the digestion method enabled successful ionization and detection of select metal impurities present in such graphene samples (see Table II). However, because of the harsh sample treatment applied to graphene flakes for complete disintegration, it was difficult to assess if the metal impurity was bound to graphene or in free solution. Further work will explore the design of sample preparation conditions and quantitative ICP-MS analysis to determine how strongly the metal impurity was incorporated on to graphene.

Radionuclide Uptake of Functionalized Graphene Oxide (GO) for Use in Nuclear Waste Treatment 

The Nuclear Decommissioning Authority (NDA) has identified the need to research graphene materials (2). Therefore, in this study, the suitability of functionalized-GO materials was assessed by determining two solution parameters by ICP-MS. The first is sorption (S) and describes the process which target radionuclide species undergo attachment to the surface of the material. The second is the distribution coefficient (

), which indicates the relative affinity of a material for a target radionuclide species and is represented as a mass-weighted ratio between the solution and the material.

The performance of each material was assessed by determining the uranium uptake by using equation 1 with the corresponding distribution coefficient, 

 refer to the initial counts per second detected by ICP-MS prior to the addition of the sorbent material and final counts per second detected by ICP-MS, respectively; 

 refers to the volume of standard solution used (MES) in mL; and 

 refers to the mass of sorbent material used in g.

Figure 1 illustrates the difference in sorption capabilities between unmodified Fe

 nanoparticles, COOH-f GO, and MGO. It was found that Fe

 has a low affinity for uranium, with 17.8% sorption and a distribution coefficient (

 mL/g observed. In contrast, it was found that attaching Fe

 NPs to the surface of COOH-f GO to form MGO lead to an increased retention of uranium; with 96.8% sorption and a 

 mL/g. The high retention suggests that MGO is suitable for use in uranium remediation as the 

 value obtained is comparable to the values reported previously for GO (6). It was found that studying the solution parameters by ICP-MS allows for the increased radionuclide sorption rates seen in graphene oxide-modified materials to be effectively measured.

Cell Uptake Mechanism of High-Z NPs for Dosimetry Assessments

Nanoparticles are known to be able to locally enhance radiation dose deposition levels and cause tumors to be more sensitive to ionizing radiation damage. However, there are differences in results between studies (for example with regards to different cell lines, nanoparticle types and concentrations, incubation times, and irradiation parameters). Two studies focused on the use of nanoparticles for increasing the level of cell kill by ionizing radiation when introduced into tumor cells. A range of analytical techniques were used in these studies (1,8), with ICP-MS used to investigate the dynamics of NP uptake in different cell lines.

The counts per second (CPS) found by ICP-MS in solutions obtained from lysed cells exposed to nanoparticles were compared to a calibration curve determined using bulk solutions containing known concentrations of nanoparticles. Figure 2 illustrates an example of the calibration curve obtained for two nanoparticle systems studied—Dotarem and gold NPs.

Analysis was performed to determine the average mass of NP per cell. The main conclusion from the data is the cell line specific nature of NP uptake. In both studies, there was a measurable difference in cell survival between the different nanoparticles and cell lines. The ICP-MS results improve the understanding of the biological impact of nanoparticles for radiation treatment, by comparing the variation in biological effects across cell line and nanoparticle type, as well as relating these findings with cellular uptake. A limitation of these measurements is that there is no indication of mechanism of uptake, and where in the cell the NP is found, which can potentially be overcome by a time-based study, such as lysing the cells at different times post-exposure to NPs.

ICP-MS has been used across several groups at the National Physical Laboratory for application in material standardization, nuclear waste treatment, and nuclear medicine. In this article, three case-studies were used to highlight how ICP-MS can be used to help resolve issues that are currently faced in these applications. ICP-MS has been shown to be effective for rapid bulk sample measurement of graphene, functionalized-GO and high-Z NPs. However, challenges still remain with ICP-MS measurement at the nanoscale, specifically with respect to understanding how surface interactions, which occur between analytes of interest and the nano-material, can be further quantified. This information can potentially be a valuable tool and allow for further insights to be made into the effect of NP size, concentration or shape on the quality, elemental and cell uptake capability of the material, respectively.

One technique which can overcome this challenge is single particle inductively coupled plasma mass spectrometry (SP-ICP-MS). In this technique, a suspension of nanoparticles is introduced, which has been diluted so that only one NP at a time is measured by the instrument. This then allows for several material properties to be measured including particle size, size distribution and the total concentration of both NPs and dissolved contents present in a sample. This would previously have required the use of multiple analytical imaging techniques. As a result, SP- ICP-MS has been increasingly used over the last few years in applications including, consumer product safety and environmental analysis whereby insights into the size and bioaccumulation of nanomaterials are vital.

To date, SP-ICP-MS has not been widely used in nuclear and materials applications. As such, there is great scope to make progress in this area. For instance, most publications have reported mainly on the analysis of ideal NPs (for example, silica and titanium dioxide) (7). This indicates that SP-ICP-MS is not well-suited for the analysis of complex and irregular nanomaterial systems, such as those highlighted in this article. Therefore, further method development work is required in the future to allow for the shape of nanomaterials to be considered during SP-ICP-MS analysis. This can be potentially achieved by developing a database from offline transmission electron microscopy (TEM) measurement of nanomaterials, which can be feed directly to the instruments software. Overall, this will confirm that a fit-for-purpose measurement protocol exists for end-users of SP-ICP-MS and will ensure that this technique remains a prime candidate for nanomaterial analysis.

The work on iron, gadolinium, and gold NPs was supported by the Medical Research Council Proximity to Discovery award (grant number 530912/170684) and the Peter Samuel Royal Free Fund. The work on functionalized graphene oxide was funded from by an iCASE studentship from the Engineering and Physical Sciences Research Council (EPSRC) awarded to the National Physical Laboratory and the University of Surrey. The work on the VAMAS study was supported by the National Key Research and Development Program of China (2016YFF0204303).

The National Physical Laboratory is operated by NPL Management Ltd, a wholly owned company of the Department for Business, Energy, and Industrial Strategy (BEIS).

(1) R. Ahmad, G. Schettino, G. Royle, M. Barry, Q.A. Pankhurst, O. Tille- ment, B. Russell, and K. Ricketts, 

(4), 900411, (2020). https://doi.org/10.1002/ ppsc.201900411

(2) P. Davies, S. Walton, S. Ashley, A. Tzalenchuk, C. Williams, and P. Wiper, 

The potential applications of graphene (and related compounds) relevant to the NDA’s decommissioning mission

, NS4145-500-005, Issue 3 (Nuclear Decommissioning Authority, Cumbria, UK 2017), NDA015384.

Results from VAMAS study focusing on sample treatment and quantitative measurement by ICP-MS of metal impurities in graphene samples

(4) J. Liu, Z. Liu, C.J. Barrow, and W. Yang, 

, 1–19 (2015). https://doi.org/10.1016/j. aca.2014.07.031

(5) H. Mohamud, Ph.D. thesis, University of Surrey, Guildford (2019).

(6) H. Mohamud, P. Ivanov, B.C. Russell, P.H. Regan, and N.I. Ward, J. 

(2), 1–10 (2018). https://doi.org/10.1007/s10967-018- 5741-4

(9), (2019). https://doi.org/10.1039/c9ja00206e

(8) E. Russell, C. McGarry, G. Schettino, B. Russell, and K.M. Prise, 

Evaluating uptake and toxicity of gadolinium-based contrast agents in combination with radiotherapy for future MR-guided therapy

. Poster presentation, ICRR, Manchester, 25–29 (2019).

(9) E. Russell, S.J. McMahon, B. Russell, H. Mohamud, C.K. McGarry, G. Schettino, K.M. Prise, Effects of Gadolinium MRI Contrast Agents on DNA Damage and Cell Survival when used in Combination with Radiation. 

. http://www.vamas. org/twa41/index.html (accessed 19/05/2020).

 are with the National Physical Laboratory in Teddington, Middlesex, United Kingdom. 

 are with the CAS Key Laboratory of Standardization and Measurement for Nanotechnology, National Center for Nanoscience and Technology in Beijing, China. Direct correspondence to: ben.russell@npl.co.uk.

Peer-Reviewed Research

Inductively Coupled Plasma–Mass Spectrometry (ICP-MS) Analysis of Nanomaterials for Use in Nuclear and Material Applications

As part of our ongoing review of infrared spectroscopic theory, we will discuss why different functional groups have different peak heights and widths. We find that Beer’s law determines peak height, and that the absorptivity in Beer’s law depends upon a mysterious quantity called the dipole moment, and is in large part responsible for why infrared peaks vary in size, from large to non-existent. Infrared peak widths can vary from a few wavenumbers to thousands of wavenumbers in width. We explore the role of intermolecular interactions in this phenomenon. We will wrap up by discussing the importance of integrating peak position, height, and width information to properly interpret spectra.

When I first introduced infrared spectral interpretation theory many columns ago (1), it was interspersed with discussions of the interpretation of specific functional groups. This approach made sense at the time, but, by not presenting all the theory of a piece, some important concepts that you need to know to interpret spectra successfully might have gotten short shrift. Hence the need for this Grand Review.

There are three important pieces of information in an infrared spectrum; peak positions, peak heights, and peak widths. All these pieces of information can be used to distinguish different functional groups from each other, and I have emphasized repeatedly in my columns the importance of integrating all these pieces of information to interpret spectra. It then makes sense for you to learn the reasons why different functional groups have different peak position, heights, and widths.

In my last column, I discussed why different functional groups have different peak positions (2). Here, we will discuss why different functional groups have different peak heights and widths. We will find that Beer’s law determines peak height, and that the absorptivity in Beer’s law depends upon a mysterious quantity called the 

, and is in large part responsible for why infrared peaks vary in size from large to nonexistent. Infrared peak widths can vary from a few wavenumbers to thousands of wavenumbers in width. We will explore the role of intermolecular interactions in this phenomenon.

When plotted in absorbance, the peak heights in an infrared spectrum are proportional to concentration, allowing infrared spectra to be used to determine the concentration of chemical species in samples (3). The relationship between absorbance and concentration is summarized in Beer’s law (3):

 equals absorbance, ɛ equals absorptivity, 

The pathlength is simply the thickness of sample seen by the infrared beam. The absorptivity is the proportionality constant between absorbance and concentration. The absorptivity also has physical meaning; it is a constant for a given pure molecule at a given wavenumber. For example, the absorptivity of pure acetone at 1716 cm

 (if not stated, assume all peak positions are in cm

 going forward) under standard conditions of temperature and pressure is a constant. This means that, under these conditions and for a fixed pathlength, a sample of pure acetone will always absorb the same amount of light at 1716.

Since different molecules have different absorptivities, this quantity must somehow be related to chemical structure. Consider the electronic structure of the CO

, the electronegativity difference between the carbon and the oxygen atoms causes the electrons to be shared unevenly (4), leading to a partial positive charge on the carbon, and partial negative charges on the oxygens. In Figure 1, a partial positive charge is represented by a δ+, and a partial negative charge by a δ-.

A dipole moment can be defined as two charges separated by a distance. Since chemical bonds, such as the ones in CO

, can consist of charges separated by a distance, chemical bonds have what are called 

. The bond dipole is a vector quantity; it has a magnitude and direction. The arrows in Figure 1 show the dipole moment vectors for CO

. Each arrow’s length represents the magnitude of the dipole moment for that bond, while the arrow head denotes its direction.

The magnitude of a bond dipole is simply the size of the charges multiplied by their distance apart, and can be calculated as such:

The bond dipole is a measure of charge asymmetry, and bonds with large charges held far apart having bigger dipole moments than bonds with small charges held close together. Thus, polar bonds, such as O-H and C=O, have large charge asymmetry, and therefore large dipole moments. Covalent bonds such as C-C bonds have small change asymmetry, and therefore small dipole moments.

The net dipole moment for a molecule is the sum of the bond dipoles for that molecule. Note in Figure 1 that, for CO

 at equilibrium, the two bond dipoles are equal in magnitude and opposite in direction. When added together, these two dipole moments cancel each other, giving a net dipole of zero.

So what does all this dipole moment stuff have to do with molecular vibrations and peak heights? Consider the symmetric and asymmetric stretching peaks of CO

During the symmetric stretch, the length of the bonds changes, but, since they stretch in phase with each other, the two bonds are always the same length, and always point in opposite directions. We say the symmetry of the molecule is maintained during the symmetric stretch (hence its name), and further meaning that the net dipole of the molecule is always zero at all points during this stretch. We can then say that, for the symmetric stretch of CO

, with respect to the bond length, dx, is zero. In other words:

An inspection of the infrared spectrum of CO

 shows the intensity of its symmetric stretching peak is zero, which means there is no peak corresponding to the symmetric stretch in CO

’s spectrum. Notice that there is a connection between dμ/dx being zero and the peak intensity being zero.

 undergoes an asymmetric stretch, as seen in Figure 2, the bond’s lengths are different at the same point in time. This means at points during this vibration, the two bond dipoles have different values, and that the net dipole is non-zero during this vibration. In this case, we say the symmetry of the molecule is broken, hence the name 

. Since the two bond dipoles values vary during the asymmetric stretch, this means the net dipole during this vibration has non-zero values. In other words:

. In fact, dμ/dx is quite large for this vibration, since we are stretching large dipole moments asymmetrically. The asymmetric stretching peak in the spectrum of CO

, and is quite intense, as seen in Figure 3. In fact, it is the intensity of this absorbance that makes CO

You may have seen 2350 peak before in sample spectra measured with a Fourier-transform infrared (FT-IR) spectrometer. This peak occurs when the CO

 concentration inside the spectrometer increases between the measurement of the background and sample spectra (please consult my book on FT-IR to learn how to avoid this problem [5]).

Note that dμ/dx for the asymmetric stretch of CO

 is large, and the vibration’s corresponding peak is large. What, then, is the connection between dμ/dx and intensity, and how does this new discovery of ours fit in with Beer’s law?

I will skip the derivation here, but quantum mechanics tells us that (6):

where ɛ equals absorptivity, dμ equals change in dipole moment during a vibration, and dx equals change in bond length during a vibration.

Therefore, the absorptivity depends upon the change in dipole moment with respect to bond length during a vibration squared. In other words, the absorptivity depends upon how the electronic structure of a molecule or functional group changes during a vibration. It is this equation that links peak height to molecular structure, and is why different functional groups have different peak heights.

A vibration such as the symmetric stretch of CO

 for which dμ/dx = 0 and the resultant peak intensity is zero is called an 

. On the other hand, vibrations such as the asymmetric stretch of CO

 for which dμ/dx ≠ 0 and the resultant peak’s intensity is non-zero are called infrared active peaks. One of the reasons that the number of normal modes a molecule has is greater than its number of infrared peaks is that not all vibrations are infrared active.

Polar functional groups such as C-O or C=O have large values of 

, hence have vibrations with large values of dμ/dx, and generally have intense peaks. On the other hand, for non-polar functional groups, such as C-C bonds that have small values of μ and their vibrations have small values of dμ/dx, their peaks are generally small. This means that infrared spectroscopy is better at detecting some functional groups than others, and that polar functional groups are generally the easiest to see. In other words, infrared spectroscopy is not a democracy, and, from the point of view of an infrared spectrometer, not all functional groups are created equal. This actually limits us as to what functional groups infrared spectroscopy can see. Fortunately, other molecular spectroscopy techniques, such as Raman scattering, are excellent at seeing the functional groups infrared (IR) cannot see, and thus can take up the slack.

In addition to peak positions and heights, peak widths can also be used to distinguish functional groups. Infrared peak widths vary greatly across the spectrum of known molecules. Figure 4 shows the infrared spectrum of liquid water.

Infrared peaks are generally narrow at their tops, and wider at their bottoms. How, then, do we measure peak widths in a standard fashion? One way spectroscopists do this is to measure what is called the full width half maximum (FWHM) of a peak (6). To measure a FWHM, take the height of a peak, go to where it is half this height, and then measure its width there. For example, if a peak is 1.0 absorbance unit (AU) high, we would measure this peak’s FWHM at 0.5 AU. The FWHM os the O-H stretching peak of water at 3300 cm

, as illustrated in Figure 4. This is an unusually broad peak for an infrared spectrum.

Figure 5 shows an example of a spectrum with narrow peaks; it is of the molecule benzonitrile.

In Figure 5, the FWHM of the C≡N stretching peak at 2228 is less than 10 cm

Integrating peak position, height, and width information is important to properly interpret infrared spectra. But do you understand why peak heights and widths vary? We explain.

January 2021

Using Raman Spectroscopy for the Characterization of Zeolite Crystals

A Genetic Algorithm–Optimized Extreme Learning Machine Model for Process Ethylene Analysis Robustness Enhancement

Inductively Coupled Plasma–Mass Spectrometry (ICP-MS) Analysis of Nanomaterials for Use in Nuclear and Material Applications

The Grand Review II: Why Do Different Functional Groups Have Different Peak Heights and Widths?

Investigating a Laser-Induced Titanium Plasma Under an Applied Static Electric Field

Narrowband Visible Light: The OPO Reigns

Vol 36 No 1 Spectroscopy January 2021 Regular Issue PDF