Real-Time Quantification of Hydrogen and Deuterium Using Laser-Induced Breakdown Spectroscopy (LIBS) in the Off-Gas of an Engineering-Scale Pumped Molten Salt Loop

Abstract

Laser-induced breakdown spectroscopy (LIBS) is being considered for elemental analysis of radionuclides transported into the off-gas system of molten salt reactors. A mobile LIBS system was calibrated for gas-phase measurements of hydrogen (i.e., protium, denoted here as H) and deuterium (D) in argon (Ar). The system was then used to quantify the concentrations of H and D in the off-gas of an engineering-scale molten salt loop (∼200 kg salt) following the addition of H₂ or D₂ as a 4% blend in Ar. The limits of detection for H and D were determined to be 0.0013 and 0.0011 mol% (13 and 11 parts per million), respectively. This research was carried out at the U.S. Department of Energy’s Oak Ridge National Laboratory in the Facility to Alleviate Salt Technology Risks (FASTR) pumped salt loop. Previous molten salt LIBS research has predominantly been performed at the bench scale, thus this study represents a major stride toward deploying LIBS technology for monitoring the performance of molten salt reactors. During the FASTR testing, the LIBS system operated continuously for several days. This online monitoring campaign demonstrates advancement in the application of LIBS for online monitoring of off-gas in molten salt systems.

Graphical abstract

This is a visual representation of the abstract.

Keywords

Laser-induced breakdown spectroscopy LIBS hydrogen deuterium hydrogen isotopes quantification real-time monitoring in-line monitoring long-duration monitoring calibration gas LIBS molten salts

Introduction

Interest in molten salts has grown in the past decade due to their excellent thermal properties, which provide access to high temperatures at ambient pressures. This ability leads to improved efficiency and safety for applications in heat transfer and thermal energy storage (e.g., molten salt reactors,¹ concentrating solar power²) and unlocks new chemical reactions (e.g., methane pyrolysis,³ pyroprocessing of used nuclear fuel⁴). Molten salts have additional advantages for fission and fusion reactors. In molten salt reactors (MSRs), the fuel can be dissolved into the salt, which enables a liquid-fueled design.¹ In fusion reactors, a molten salt blanket can be used to produce tritium fuel (i.e., breeding).⁵

Online monitoring of hydrogen isotopes (this article uses H for protium, D for deuterium, and T for tritium) in molten salts is necessary for a variety of applications, including salt purification, T breeding for fusion reactors using molten salt blankets, and methane pyrolysis. A critical aspect of molten salt systems is controlling the closely related properties of salt purity and corrosivity. An increase of hydrogen in the off-gas could indicate the formation of hydroxides and increased salt corrosivity.⁶ Therefore, sensors for hydrogen are relevant to all applications of molten salts. The ability to differentiate hydrogen isotopes is crucial for monitoring T production in fusion salt breeding blankets, and the detection of D in MSRs is important due to the isotope’s effects on neutron moderation. Therefore, there is a need to detect and quantify hydrogen isotopes in the gas phase in real-time and at trace concentrations.⁷ However, this measurement can be challenging for traditional gas analysis techniques such as mass spectrometry and optical spectroscopy, e.g., Raman, ultraviolet–visible (UV–Vis) absorbance, Fourier transform infrared (FT-IR). Online monitoring of ambient or low-pressure (100–200 kPa) systems poses challenges for these sensors because mass spectrometry requires a high vacuum and traditional optical techniques benefit from high pressures or long path lengths to increase sensitivity.

One technique that is well suited for this application is laser-induced breakdown spectroscopy (LIBS). In LIBS, a high-power laser pulse ionizes the sample (in this case, a gas), which forms a plasma composed of excited ions from the sample. As the plasma cools, each ion or atom emits photons of characteristic energies, and by recording the spectrum of light emitted from the plasma, the elemental composition can be determined. The LIBS technique is rapid, and no sample preparation is required. Previous studies have demonstrated LIBS for detection of H, D, and T in gases,^8–11 liquids,¹² and solids.^13,14 The concentrations of H and D have been quantified in solid samples,¹⁵ but gas-phase quantification has focused on isotope ratios.¹⁶ However, for online monitoring of molten salt off-gas, quantification of the concentration of individual hydrogen isotopes in the gas phase is required; this was the focus of the present study.

This study built upon previous research progress in developing LIBS as a sensor technology for monitoring molten salt reactors. Previous studies focused on analyzing salt aerosol particles,¹⁷ detecting and discriminating hydrogen isotopes (without quantification),¹¹ investigating bulk gas effects (e.g., Ar versus He), and quantifying noble gases (Kr, Xe).¹⁸ In contrast, this study shifted from small, controlled, lab-scale systems to an engineering-scale system. The goals of this study were to identify key considerations for online monitoring of gas-phase analytes using LIBS, to develop calibrations for discriminating and quantifying gas-phase concentrations of hydrogen isotopes using LIBS, to establish limits of detection (LODs), and to demonstrate progress in the application of LIBS for online monitoring of MSR off-gas. To achieve those goals, this study investigated the effects of laser energy, pressure, matrix effects, normalization, peak fitting, and spectrometer resolution. Univariate and multivariate calibration models were compared for H and D quantification. Tests were carried out on an engineering-scale pumped salt loop in real-time. This study is a critical step in demonstrating sensor feasibility and realizing the application of LIBS for off-gas monitoring in molten salt systems.

Experimental

LIBS Measurements

A mobile LIBS platform was previously developed for deployment on bench- and engineering-scale test systems.^11,17,18 The platform was based on a commercial modular LIBS unit (LIBS-8, Applied Photonics) with a gas sample chamber attachment. The focusing optics consisted of a 3× beam expander, with a maximum beam diameter of ∼19 mm, and a 100 mm focal length lens with a diameter of 25.4 mm, resulting in an f-number of ∼4. The collection optics module includes eight separate optics; each coupled to a fiber port. Six of these fibers are connected to a built-in six-channel spectrometer (AvaSpec 4096CL, Avantes), hereafter referred to as the multichannel spectrometer (MCS). The remaining two fiber ports can be connected to auxiliary spectrometers for enhanced sensitivity or resolution. The laser was a Q-switched 1064 nm neodymium-doped yttrium aluminum garnet (Nd:YAG) laser (Litron Nano SG-150-10, Litron Lasers) capable of producing approximately 9 ns pulses of up to 150 mJ at frequencies up to 10 Hz. The laser, power supply unit, sample chamber, MCS, and control computer, were mounted on a custom cart.

The sample chamber on the mobile LIBS platform consisted of a roughly 130 × 130 × 120 mm cube. The gas analyte flowed into the chamber within a stainless-steel tube with an inner diameter (ID) of 6 mm and was then focused through a nozzle with a 0.8 mm aperture and directed across a 2 cm gap into a 10 mm ID stainless steel outlet tube. The laser focal point bisected the gap and allowed the plasma to form without interacting with the nozzle or outlet tube. Before each test, the chamber was purged with the same gas that was used for the sample carrier gas (i.e., Ar) using a chamber purge inlet located at the back of the chamber.

A second mobile spectrometer system consisting of a double echelle monochromator (DEMON, Laserteknik, Berlin) mounted on a custom cart was connected to the LIBS platform to enable high-resolution measurements through an auxiliary fiber port. A digital delay generator (Sapphire 9200, Quantum Composers) was used to synchronize the DEMON to the MCS so that spectra were recorded simultaneously and at the same delay and integration times.

The mobile LIBS system and DEMON spectrometer carts were transported 0.5 km between buildings, and no recalibration was required after moving, demonstrating that the system is robust and reliable.

An initial parameter optimization study found that a laser energy of 50 mJ, delay time of 15 μs, and integration time of 1 ms gave the best signal-to-background ratio for the H-alpha line (656.6 nm) using the MCS. Despite finding 50 mJ to be the optimal energy, calibrations were performed at 50 and 150 mJ laser energies to allow flexibility during online testing. Delay time and integration time were fixed at the optimal values of 15 μs and 1 ms, respectively. Spectra were recorded every 10 s as 100-shots accumulate to optimize the time-resolution versus signal-to-noise ratio.

Residual Gas Analyzer Measurements

The LIBS measurements were compared to an in-line residual gas analyzer (RGA; ThinkSRS UGA 100, Stanford Research Systems). An RGA is a standard technique for gas composition analysis and is based on mass spectrometry. The mass-to-charge ratio channels for H₂, D₂, Ar, HCl, and H₂O were monitored. Potential interference can occur between D₂ and He in the RGA. Although this was not an issue in this test because Ar was used for the cover gas, MSRs are likely to use He for a cover gas, which would make it very challenging to monitor D₂ with an RGA. The RGA requires reduced pressures to operate. These pressures were achieved by connecting the RGA sample inlet in a bypass of the LIBS sample line with a stainless-steel capillary to limit gas flow and enable the vacuum conditions necessary for the RGA.

Facility to Alleviate Salt Technology Risks

Tests were carried out at the U.S. Department of Energy’s Oak Ridge National Laboratory in the Facility to Alleviate Salt Technology Risks (FASTR), which is an engineering-scale pumped molten salt loop containing approximately 250 kg of MgCl₂–KCl–NaCl eutectic.^19,20 FASTR consists of a storage tank in which salt can be kept frozen or molten, a transfer line to allow for pressure-driven transfer of molten salt between the storage tank and loop, a pump tank and mechanical pump to drive molten salt through the loop, an upper plenum in which gases can escape the salt, a main heater, and a heat exchanger for heat rejection. The salt is maintained under an inert, dry Ar (ultra-high-purity) atmosphere. The pressure in the gas space is controlled by a programmable logic controller and mass flow controllers for adding or removing gas. Additional gas lines are available on the storage tank for adding gas to the headspace or sparging through the salt. Gas sampling ports are available on the storage tank and on the pump tank. The experimental setup is shown in Figure 1.

Figure 1.

Diagram of FASTR and experimental setup. Calibration samples were prepared by diluting 4% mixtures of H₂ or D₂ gas in Ar using a gas manifold. For tests on the storage tank, gas could be introduced to the salt through the sparge gas inlet. The off-gas could be sampled from the outlets on the storage tank or pump tank. The mobile LIBS platform incorporated the laser, power supply, and MCS. The DEMON was connected to one of the auxiliary optical fiber ports on the LIBS cart for high-resolution measurements.

Calibration Sample Preparation

Calibration samples were prepared using a gas manifold to control the mixing ratio of H₂ or D₂ gas with Ar. The H₂ and D₂ gases were supplied in the form of a 4% mixture in Ar. These were further diluted in ultra-high-purity Ar to produce calibration samples in the 0.01%–4% range. All gases were obtained from Airgas LLC. The gas manifold had a bleed valve to allow for mixing gases at a high flow rate while delivering a lower flow rate to the LIBS. This utilized more gas but greatly reduced the mixing time required between calibration samples to allow the concentration to equilibrate. The gas mixture was passed through a Drierite desiccant before entering the LIBS sample chamber. A rotameter downstream of the outlet allowed a typical flow rate of 0.3 standard liters per minute (SLPM) to be maintained through the LIBS sample chamber, even if gas was added to FASTR at a higher flow rate. In that case, excess gas was vented automatically by the FASTR mass flow controllers. The gas manifold is diagrammed in Figure 1.

As LIBS is an elemental analysis technique, for the purposes of calibration, the elemental concentration of H or D in terms of mole percent (mol%) was considered, rather than the H₂ or D₂ molecular concentration. Calibration samples were prepared with nominal concentrations of H or D of 8, 6, 4, 2, and 0.05 mol%, and a blank. To more closely investigate LODs, additional samples were prepared with D at 0.7 and 0.2 mol%. This study selected D for the additional samples because the increased atomic mass results in reduced peak broadening compared to H.¹³ The narrower peak results in increased SNR and is, therefore, expected to have a better LOD. For each calibration level, 100 replicate spectra were recorded. Each spectrum was an accumulation of 100 shots.

Real-Time Tests

To demonstrate the quantitative predictive capabilities of the calibration models under controlled conditions, a real-time test was carried out on the storage tank while H₂ or D₂ gas (4% in Ar) was added at a known rate. The LIBS system was connected to the storage tank gas sampling port with a 2.8 m flexible antistatic nylon sampling line, (ID = 4 mm), and the off-gas was continuously monitored. Data were recorded simultaneously on the MCS, DEMON, and RGA. Before the test, the tank and LIBS system were purged with Ar overnight to remove any water vapor or residual gas. During the test, the salt was sparged with H₂ at 0.3 SLPM for 1 h, then the gas was alternated between Ar and H₂ for periods of 30 min at increasing flow rates of 0.3, 0.9, and 1.8 SLPM. This was repeated with D₂ alternating with Ar. The conditions for the experiment are detailed in Table I.

Table I.

Experimental conditions for real-time test on storage tank.

Time (h)	0	1	1.5	2	2.5	3	3.5	4	4.5	5	5.5	6
Gas*	H₂	Ar	H₂	Ar	H₂	Ar	D₂	Ar	D₂	Ar	D₂	Ar
Flow rate (SLPM)	0.3	0.3	0.9	0.9	1.8	1.8	0.3	0.3	0.9	0.9	1.8	0.3

*H₂ and D₂ were added as 4% in Ar.

Immediately following this test, data collection continued overnight while the salt was sparged with Ar at a constant rate of 0.3 SLPM. The LIBS data collection was automated using a custom program written in Python 3 to collect 20 spectra every 10 min.

Finally, the LIBS system was connected to the gas sampling port of the pump tank with a 6.4 m flexible antistatic nylon sampling line (ID = 4 mm). The off-gas was monitored before, during, and after pumped operation of the loop with molten salt circulating. No D₂ or H₂ was added during these tests, but residual gas was present in the system and dissolved in the salt. A sweep gas (Ar) was used to flush the headspace of the pump tank and enhance sampling. The initial sweep gas flow rate was 0.15 SLPM; this was increased to 0.3 SLPM at t = 1.6 h, then shut off at t = 2.75 h, where t is the elapsed time after the gas sweep was started.

The tests reported here are only a subset of experiments carried out over the course of 10 days, including multiple overnight runs. In total, the LIBS system operated for over 140 h. The longest single test was 45 h, during which data were collected every 10 min. Data from selected tests are shown in the Supplemental Material. The results of other experiments will be discussed in future publications.

Data Analysis

The LIBS measurements are sensitive to shot-to-shot variability in laser–sample interactions and plasma properties, as well as matrix effects. Due to the relatively large range of H/D concentrations included in the calibration (up to 8 mol%), significant matrix effects were observed, as shown in Figure S1 in the Supplemental Material. To correct for this, spectra were normalized to the Ar signal at 696.54 nm. This peak was selected because it was most similar to the H/D peak at 656.2 nm in terms of intensity and energy levels and was captured on the same MCS channel. This signal was only recorded in the MCS spectra, but because the DEMON spectra were collected simultaneously, they could be normalized using the corresponding MCS spectra. This normalization improved the linearity of response and reduced shot-to-shot variability, as shown in Figure S1 in the Supplemental Material. The DEMON uses an intensified charge-coupled device (ICCD) detector, which is susceptible to noise spikes from cosmic rays and environmental radiation. Thus, before normalization, the DEMON spectra were processed with a spike-removal filter. Peak positions were verified using the National Institute of Standards and Technology’s (NIST) Atomic Spectra Database.²¹ All data analyses were performed using Python 3 with the statsmodels and sklearn libraries.^22–24

Results and Discussion

Peak Fitting

The shapes of LIBS emission peaks are influenced by instrumental, van der Waals (dipole), Doppler, and Stark (collisional) broadening. Each of these sources generally contributes a Gaussian and/or Lorentzian shape, so peaks are commonly fit with a Voigt profile, which is the convolution of a Gaussian and a Lorentzian. Additional sources, such as ion broadening, can result in asymmetry that is not Gaussian or Lorentzian, and therefore not accounted for in a Voigt fit.²⁵ While it is possible to account for additional broadening, this requires higher-resolution data to accommodate the greater number of parameters that must be fit and may not be reliable for modeling peaks when an isotopic shift is also present. It is also more computationally expensive, requiring more time to fit each peak, and therefore may not be suitable for real-time online monitoring. Accurately modeling the peak broadening is important for extracting peak widths to predict plasma parameters such as plasma temperature and electron density, but it can also improve calibration accuracy because the peak fitting smooths the raw data, which allows for a more accurate determination of signal (i.e., peak area).

Therefore, to determine the H and D signals (peak areas), the pure H and D spectral peak profiles were fit using a Voigt function to determine the center of each peak. Then, a combined model was constructed to model spectral peaks as the sum of two Voigt profiles based on the pure H and D peak centers. The resulting fit parameters indicated that the Lorentzian term mostly described the observed peak shapes, both on the MCS and DEMON, and the Gaussian contribution was minimal. As expected, the Voigt profile did not account for all broadening of the H-alpha line (see Figure S2 in the Supplemental Material). However, this did not appear to have a negative effect on calibration model performance.

For real-time online monitoring, data processing schemes must be robust, reliable, and fast enough to keep up with the data acquisition rate. The method in this study used a model based on the sum of two Voigt profiles and took a maximum of 60 ms to process a spectrum, which was sufficient for real-time performance even at sampling rates much higher than what would be required for MSR off-gas monitoring. Further optimization is possible, but nevertheless, peak fitting, which must be performed for each peak to be monitored on every spectrum, may not be suitable if many peaks are to be monitored and spectra are collected at a high frequency. In this case, multivariate methods, which apply a precalculated transformation to each spectrum, can offer much faster performance.

Effects of Energy and Normalization

During the initial optimization of LIBS parameters, the effects of laser energy were investigated. Previous research determined that the breakdown threshold for Ar with the LIBS platform is 15–20 mJ.¹⁸ Therefore, 50 mJ was selected as the low energy level and 150 mJ as the high energy level. Generally, higher laser energy leads to more excitation of the sample and increased plasma intensity. Figure 2 shows the raw and normalized calibration spectra for H and D in Ar at these energies. The concentrations of H₂ and D₂ were nominally 4%, corresponding to 8 mol% H or D. Although the raw intensity increased with energy, the normalized intensities were lower at increased energy due to the overwhelming effect of the increased Ar background emission. At the same time, peak broadening was notable, as well as an increased baseline at 150 mJ, indicating that under these conditions, a lower power such as 50 mJ offered superior performance. Therefore, 50 mJ was used for further tests. The improved resolution of the DEMON (Figures 2b, 2d) is evident from the narrower line profile. The signal from D was slightly higher than for H at the same concentration.

Figure 2.

Spectral profiles for H and D in Ar (8 mol%) recorded on the (a, c) MCS and (b, d) DEMON. (a, b) Raw signal and (c, d) normalized to the Ar peak at 696.5 nm. Note that DEMON spectra were normalized to the 696.5 nm Ar peak measured on the MCS.

Normalization is necessary to reduce shot-to-shot variability, and it can also mitigate matrix effects due to changes in bulk gas composition. This presents practical challenges for high-resolution measurements with limited spectral coverage. Due to the limited wavelength range of the DEMON, these spectra had to be normalized to the corresponding MCS data.

Effects of Pressure

The pressure and flow rate of the gas affect the laser-induced plasma properties and behavior, which affect the signal. Because the pressure in the FASTR gas space can vary from 130 to 240 kPa, spectra were collected throughout this range of pressures to investigate the effects on the H and D signals. Increasing pressure caused peak broadening and reduced emission intensity, but the impact was relatively small, and normalization mitigated the effects. In the range of pressures investigated, only a 15% change in normalized peak area was observed. The spectra and predicted concentrations are shown in Supplemental Material Figures S3 and S4. In this study, pressure and flow rate were fixed during calibration and storage tank measurements to limit the number of calibration samples needed. If flow rate and/or pressure are variable, more calibration measurements would be needed, and multivariate models would likely offer the best performance.

Calibrations

Approaches to calibration in spectroscopy can be broadly categorized as univariate or multivariate. Univariate models are based on a single input, such as the signal of a single wavelength channel or the total signal across a narrow range of wavelength channels (i.e., peak height or peak area). Regression analysis, such as ordinary least squares regression (OLS), is used to capture the relationship between signal and concentration using a linear or higher-order fit. Multivariate models in spectroscopy use the signal on multiple wavelength channels as separate inputs and generally employ some form of data compression by identifying underlying correlations in the input data.^26,27 Common approaches are principal components regression (PCR), which captures the maximum variance in the input signal, and partial least squares regression (PLS), which captures the maximum covariance between the input signal and the sample concentrations.

Various metrics can be used to compare the performance of different models. The value of R² is a measure of how well the model captures variation in the data. The root mean square error (RMSE) is a measure of how well the model predicts the data used for calibration (RMSEC) or cross-validation (RMSECV); this is commonly expressed as a percentage of the mean calibration sample concentration. The limit of quantification (LOQ) describes the lower concentration limit at which a model’s predictions are quantitative, which accounts for uncertainty in the model. For the OLS models in this study, LOQ was calculated based on the 95% confidence prediction interval.^28,29 For PCR and PLS, the pseudo-univariate LOQ was used.³⁰

Figure 3 shows the spectra for each calibration level (average of 100 replicates), along with the actual concentrations. Four different calibration methods were compared, first- and second-order OLS (OLS1 and OLS2), PCR, and PLS. Separate models were created for the MCS and DEMON data. OLS models were fit to the H and D peak areas calculated using the double-Voigt fit described previously. The data exhibited nonlinearity, which was captured well by the OLS2 and multivariate models, but the OLS1 model only performed well in the linear low concentration range. Therefore, to enable the best comparison between models, the OLS1 model was only evaluated for the lowest concentration samples. Because extra calibration samples were included for D, the five lowest levels were used, but for H, only the lowest four levels were used. The spectral window for PCR and PLS was optimized based on the resulting R², LOQ, and RMSECV. The window range was 654.5–657.5 nm for the DEMON data and 502.5–725 nm for the MCS data (all captured on the same channel). The number of latent variables (LVs) was selected for each model by increasing the number of LVs until including an additional LV improved the RMSECV by less than 10%. The number of LVs ranged from two to four for PCR models and from three to four for PLS models. PCR and PLS models were tested using leave-one-group-out cross-validation, in which each calibration concentration level was treated as a group, and one group was removed at a time and used to test the model constructed from the remaining groups. Model performance was compared based on R², percent RMSEC (%RMSEC) or percent RMSECV (%RMSECV), and LOQ. These figures of merit are summarized in Table II.

Figure 3.

Calibration spectra collected at 50 mJ laser energy for (a, b) H and (c, d) D recorded using the (a, c) MCS and (b, d) DEMON.

Table II.

Figures of merit for comparison of calibration models.

		MCS			DEMON
Method	Gas	LOQ(mol%)^a	%RMSEC or%RMSECV^b	R²	LOQ(mol%)^a	%RMSEC or%RMSECV^b	R²
OLS1	D	0.11%	5.7%	0.9976	0.14%	6.9%	0.9966
OLS1	H	0.22%	4.0%	0.9988	0.30%	5.4%	0.9978
OLS2	D	0.23%	2.3%	0.9996	0.27%	2.7%	0.9994
OLS2	H	0.25%	2.0%	0.9995	0.31%	2.4%	0.9993
PCR	D	0.95%	6.6%	0.9985	0.77%	5.4%	0.9990
PCR	H	0.63%	4.5%	0.9994	1.09%	7.7%	0.9982
PLS	D	0.95%	6.5%	0.9985	0.76%	5.4%	0.9990
PLS	H	0.47%	3.4%	0.9997	0.72%	5.1%	0.9992

LOQ for OLS1 and OLS calculated based on 95% confidence interval of prediction. LOQ for PCR and PLS calculated based on a pseudo-univariate approach.

Univariate models (OLS1 and OLS2) evaluated based on %RMSEC. Multivariate models (PCR and PLS) evaluated based on %RMSECV with leave-one-group-out cross-validation.

All models had R² > 0.99, with many >0.999. Similarly, all models performed well based on %RMSEC or %RMSECV as well as LOQ, although the OLS models performed better than PCR and PLS. This was likely due to the overlap between H and D peaks, which the multivariate models can capture in the LV loadings but was more reliably and accurately modeled by the peak fitting approach used with the OLS models. The lowest LOQs were for the OLS1 model; this was expected because the LOQ was calculated based on the prediction band, which accounted for the uncertainties of all calibration samples. As the uncertainty scaled with signal, the high-concentration samples had higher uncertainty, and excluding them reduced the total uncertainty, resulting in a lower LOQ. The best %RMSEC or %RMSECV was for the OLS2 models, which performed better than the OLS1 models at higher concentrations because they could better describe nonlinearities in the calibration data. Therefore, to provide the best overall performance, a hierarchical model was used which primarily relied on the OLS2 model unless the OLS2-predicted concentration fell below the LOQ for that model (see Table II), in which case the OLS1 model was used. The OLS2 calibration curves are shown in Figure 4. Although univariate models had the best performance in this case, as mentioned previously, multivariate models would likely be needed if pressure and/or flow rate are variable, or if faster computation times are required.

Figure 4.

Second-order OLS calibration curves for (a, b) H and (c, d) D based on (a, c) MCS data and (b, d) DEMON data.

In general, the calibrations made from the MCS data slightly outperformed their DEMON counterparts. This result is not intuitive because the DEMON has better sensitivity and resolution; however, the DEMON data had to be normalized to the MCS data. Although this normalization made a significant improvement to the noise and linearity of the DEMON data, it also introduced uncertainty related to the MCS signal, resulting in similar performance by models constructed for each spectrometer. Minimal difference was noted between the performance of H and D models, indicating that LIBS can reliably be used to quantify either isotope.

Limits of Detection

The traditional method of estimating the LOD for linear, univariate calibrations is based on the calibration slope (m) and standard deviation of the blank (σ): LOD = 3σ/m. Although this equation is most commonly used, it fails to account for uncertainty in the calibration fit or errors in calibration sample concentrations and is frequently extrapolated far outside the range of calibration samples. It is also not suitable for comparing univariate and multivariate calibration models. Nevertheless, this traditional approach still holds value because it describes the lower limit at which calibration should be possible with a given technique. Therefore, according to this approach, the LODs for the MCS OLS1 models were 0.0013 mol% for H and 0.0011 mol% for D (additional values are reported in Table S2, Supplemental Material). These values correspond to 13 and 11 ppm (parts per million), which is comparable to LODs reported for H and D in gases and in solids (10–20 ppm),^8,15 indicating that LIBS is suitable for monitoring hydrogen isotopes at low concentrations. However, further research into alternate techniques, such as double-pulse LIBS, is needed to improve sensitivity for trace detection.

Storage Tank Test Results

A combined LIBS prediction model was created from the hierarchical OLS models based on the MCS data because these had the best performance. This combined model was used to predict the H and D concentrations for the real-time test on the storage tank. These predictions are shown in Figure 5, along with the data recorded simultaneously by the RGA. Although concentration can be determined using the RGA, this requires a gas-specific sensitivity factor, which was not available. Instead, the RGA signal was scaled to match the range of the concentrations measured using LIBS to serve as a qualitative comparison.

Figure 5.

(a) LIBS-predicted H and D concentrations using the MCS OLS model versus RGA signal for the test on the FASTR storage tank. Changes in the flow rate and composition of the sparge gas are denoted by the dashed lines and labels above the plot. (b) Predicted H and D concentrations were monitored overnight. The expanded region shows detection of H and D at concentrations below the LOQ and approaching the LOD (the H and D LODs are indistinguishable on this scale).

The LIBS model quickly responded to increases and decreases in analyte concentration when the sparge gas or flow rate was changed; any latency can most likely be attributed to the response time of FASTR (i.e., time for gas to flow to LIBS sensor location). The moving percent RSD (%RSD) of the LIBS predictions was roughly 1%–2% for H and D, which corresponded with the model %RMSEC (see Figure S5, Supplemental Material). Overall, the LIBS results agreed with the RGA signal, with proportional increases for each H or D addition. However, the RGA data was delayed by approximately 20 min, and the RGA failed to register decreases in H₂ or D₂ concentration when Ar was added. The time lag in the RGA data can be attributed to the capillary sampling inlet used to maintain a reduced pressure in the RGA, which introduced a sampling delay. The delay in response to decreasing concentration is likely due to H₂ or D₂ accumulating or being retained within the RGA. A turbomolecular pump was used to maintain near-vacuum conditions within the analyzer, which allowed for very sensitive measurements. However, the pump was less efficient at removing very light atoms or molecules, which can lead to accumulation of H₂. These limitations of the RGA resulting from the low-pressure requirements make it less suitable for this application.

The RGA specifically measures molecular H₂ and D₂, but LIBS measures total H or D, regardless of phase or molecular species. This can be an advantage, especially considering the excellent sensitivity and response time of LIBS. In particular, LIBS could be useful for monitoring the total D content in MSRs, which can affect reactor neutronics. However, if molecular information is required, LIBS can be paired with other techniques, such as RGA or Raman spectroscopy.³¹

Following the test, monitoring continued overnight, as shown in Figure 5b. The H and D concentrations decreased steadily throughout the test, which was consistent with slow dilution of the large gas volume of the storage tank (143 L). The D concentration fell below the OLS1 LOQ at t ≈ 8 h and continued to drop below 0.01 mol%, approaching the LOD of 0.0011 mol%. At t ≈ 4 h, the H concentration decreased below the OLS1 LOQ before leveling off at approximately 0.09 mol%. The expected level of H in the background was <0.001 mol% due to residual impurities in the Ar sparge gas. The elevated level of H may have been caused by residual moisture held up in the LIBS sample chamber or in the sampling lines. To enable quantitative predictions in this concentration range, further calibrations are required. The lowest concentration that could be produced by the setup used in this study was 0.02 mol%, and the lowest concentration calibration samples used were 0.04 mol% for H and 0.07 mol% for D. Future calibrations will require calibration samples in the 0.0001–0.01 mol% range or lower, which will require new methods of sample preparation.

Pump Tank Test Results

After the test on the storage tank, the salt was transferred to the pump tank, and LIBS was used to monitor the off-gas from the pumped loop. Because the H and D concentrations in this test were much lower than in the test on the storage tank, the OLS1 models were used for quantification. The D concentrations measured in this test fell below the model’s LOQ (see Table II) and were outside the concentration range of the calibration samples. However, they were above the LOD of 0.0013 mol%. The results of this test are shown in Figure 6.

Figure 6.

Predicted H (left axis) and D (right axis) concentrations for the pump tank test. The H signal was elevated due to residual air and moisture in the sample line. Both H and D signals increased when the pump was turned on as dissolved gases were released from the salt.

At the start of the test, detected H and D concentrations briefly spiked, likely due to air and residual gas held up in the sampling lines from previous tests. Residual moisture in the sample chamber also contributed to an elevated baseline H concentration which decreased throughout the test until the sweep gas flow was shut off at t = 2.75 h; when the sweep gas was stopped, flow through the LIBS stopped, allowing the gases in the sample chamber to mix freely, resulting in a decrease in D and corresponding increase in H within the analyzed volume.

When the pump was turned on, both H and D signals increased as dissolved gas was released from the salt. Although the H concentration reached its peak only 0.25 h after the pump was turned on, the D concentration continued to increase for about 0.75 h. The total amount of H and D dissolved in the salt was not controlled in this experiment, but more D likely would have been dissolved because the previous experiment ended with sparging the salt with D₂. The slower release of D was likely due to the decreased diffusion coefficients for D₂ compared to H₂; the diffusion coefficient is inversely proportional to the square root of molecular mass (i.e., the diffusion of D₂ is approximately 71% that of H₂). A detailed analysis of these data is planned as part of a future article. The pump flow rate did not appear to affect the rate of H or D release, but this factor should be examined more closely over longer timescales and at higher concentrations of dissolved gas.

Conclusion

This study identified the importance of normalization to correct for shot-to-shot variability and mitigate the impact of pressure and matrix effects. Although high resolution enables better discrimination of isotopes, the results of this experiment indicated the importance of broad spectral coverage over high resolution for quantification because normalization requires a spectral range that includes peaks for the analyte(s) of interest as well as the bulk gas. Lower laser energy and longer delay times (50 mJ and 15 μs) offered the best performance in an Ar bulk gas. Future work should focus on combining the sensitivity and resolution of the DEMON with the broadband advantages of the MCS by using different spectrometers, such as an echelle, paired with sensitive detectors such as an ICCD. Calibrations were constructed for real-time discrimination and quantification of gas-phase hydrogen isotopes, with LOQs of 0.22 mol% for H and 0.11 mol% for D. This study found that univariate (OLS) models outperformed multivariate models (PCR, PLS) when pressure and flow rate were kept constant. Multivariate models may be required if pressure and flow rate vary, and they may also offer faster computation compared to univariate models that require peak-fitting of every spectrum. The LODs for H and D with the current LIBS platform were established at 0.0013 and 0.0011 mol% (13 and 11 ppm). Future research should focus on improving sensitivity (e.g., using double pulse-LIBS). This study demonstrated the robust, reliable performance of the LIBS platform for deployed and automated testing, and the sensitivity and selectivity for hydrogen isotopes were also demonstrated through long-duration, real-time tests on an engineering-scale pumped molten salt loop.

Supplemental Material

sj-docx-1-asp-10.1177_00037028261454365 - Supplemental material for Real-Time Quantification of Hydrogen and Deuterium Using Laser-Induced Breakdown Spectroscopy (LIBS) in the Off-Gas of an Engineering-Scale Pumped Molten Salt Loop

Supplemental material, sj-docx-1-asp-10.1177_00037028261454365 for Real-Time Quantification of Hydrogen and Deuterium Using Laser-Induced Breakdown Spectroscopy (LIBS) in the Off-Gas of an Engineering-Scale Pumped Molten Salt Loop by Zechariah B. Kitzhaber, Daniel Orea, Joanna McFarlane, Kevin Robb and Hunter B. Andrews in Applied Spectroscopy

Footnotes

Acknowledgments

This work was funded by the US Department of Energy’s Office of Nuclear Energy, Advanced Reactor Development Program, Molten Salt Reactor Program.

Author Note

This manuscript has been authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the U.S. Department of Energy (DOE). The US government retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan ().

CRediT Author Statement

Conceptualization (Z.K., H.A.), Data curation (Z.K., H.A., D.O.), Formal analysis (Z.K.), Funding acquisition (H.A., J.M., K.R.), Investigation (Z.K., H.A.), Methodology (Z.K., D.O., H.A.), Visualization (Z.K., H.A.), Writing – original draft (Z.K., H.A.), Writing – review and editing (All authors).

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability

All relevant data that support these experimental findings are available from the corresponding author upon reasonable request. A Supplemental Information file is available containing additional details and figures regarding peak fitting, normalization, pressure effects, limits of detection, prediction RSD, and additional calibration data.

Supplemental Material

All supplemental material mentioned in the text accompanies this paper online.

ORCID iDs

Zechariah B. Kitzhaber

Joanna McFarlane

Kevin Robb

Hunter B. Andrews

References

Serp

Allibert

Beneš

Delpech

, et al. “The Molten Salt Reactor (MSR) in Generation IV: Overview and Perspectives”. Prog. Nucl. Energy 2014. 77: 308–319. 10.1016/j.pnucene.2014.02.014

Kurley

J.M.

Halstenberg

P.W.

McAlister

Raiman

, et al. “Enabling Chloride Salts for Thermal Energy Storage: Implications of Salt Purity”. RSC Adv. 2019. 9(44): 25602–25608. 20190815. 10.1039/c9ra03133b

Nikolis

Patzschke

C.F.

Nabizade

Tan

, et al. “Methane Pyrolysis for H₂ Production in Molten Salts: Simulation of the Reaction, Transport Phenomena, and the Industrial-Scale Process”. Chem. Eng. J. 2025. 524: 168104. 10.1016/j.cej.2025.168104

Laidler

J.J.

Batles

J.E.

Miller

W.E.

, et al. “Development of Pyroprocessing Technology”. Prog. Nucl. Energy. 1997. 31: 131–140. 10.1016/S0149-1970(96)00014-5

Boullon

Jaboulay

J.-C.

Aubert

. “Molten Salt Breeding Blanket: Investigations and Proposals of Pre-Conceptual Design Options for Testing in DEMO”. Fusion Eng. Des. 2021. 171: 112707. 10.1016/j.fusengdes.2021.112707

Sulejmanovic

Kurley

J.M.

Robb

Raiman

. “Validating Modern Methods for Impurity Analysis in Fluoride Salts”. J. Nucl. Mater. 2021. 553: Article no. 152972. 10.1016/j.jnucmat.2021.152972

Taylor

C.N.

. “Hydrogen and Its Detection in Fusion and Fission Nuclear Materials: A Review”. J. Nucl. Mater. 2022. 558: Article no. 153396. 10.1016/j.jnucmat.2021.153396

Ball

A.J.

Hohreiter

Hahn

D.W.

. “Hydrogen Leak Detection Using Laser-Induced Breakdown Spectroscopy”. Appl. Spectrosc. 2005. 59(3): 348–353. 10.1366/0003702053585282

Jõgi

Ristkok

Butikova

Raud

Paris

. “LIBS Plasma in Atmospheric Pressure Argon, Nitrogen and Helium: Spatio-Temporal Distribution of Plume Emission and Hα Linewidth”. Nucl. Mater. Energy. 2023. 37: 101543. 10.1016/j.nme.2023.101543

10.

Doucet

F.R.

Lithgow

Kosierb

Bouchard

Sabsabi

. “Determination of Isotope Ratios Using Laser-Induced Breakdown Spectroscopy in Ambient Air at Atmospheric Pressure for Nuclear Forensics”. J. Anal. At. Spectrom. 2011. 26(3): 563–541. 10.1039/c0ja00199f

11.

Andrews

H.B.

Kitzhaber

Z.B.

Orea

McFarlane

. “Real-Time Elemental and Isotopic Measurements of Molten Salt Systems Through Laser-Induced Breakdown Spectroscopy”. J. Am. Chem. Soc. 2025. 147(1): 910–917. 20241204. 10.1021/jacs.4c13684

12.

Cremers

D.A.

Beddingfield

Smithwick

Chinni

R.C.

, et al. “Monitoring Uranium, Hydrogen, and Lithium and Their Isotopes Using a Compact Laser-Induced Breakdown Spectroscopy (LIBS) Probe and High-Resolution Spectrometer”. Appl. Spectrosc. 2012. 66(3): 250–261. 10.1366/11-06314

13.

Harilal

S.S.

Shaik

A.K.

Kautz

E.J.

Devaraj

, et al. “Detection of Tritium Using Ultrafast Laser-Induced Breakdown Spectroscopy”. J. Anal. At. Spectrom. 2024. 39(3): 699–703. 10.1039/d3ja00439b

14.

Mittelmann

Touchet

Mao

Park

, et al. “Hydrogen Isotope Analysis in W-Tiles Using fs-LIBS”. Sci. Rep. 2023. 13: 2285. 10.1038/s41598-023-29138-2

15.

Pardede

Lie

T.J.

Iqbal

Bilal

, et al. “H-D Analysis Employing Energy Transfer from Metastable Excited-State He in Double-Pulse LIBS with Low-Pressure He Gas”. Anal. Chem. 2019. 91(2): 1571–1577. 10.1021/acs.analchem.8b04834

16.

D’Ulivo

Onor

Pitzalis

Spiniello

, et al. “Determination of the Deuterium/Hydrogen Ratio in Gas Reaction Products by Laser-Induced Breakdown Spectroscopy”. Spectrochim. Acta, Part B. 2006. 61(7): 797–802. 10.1016/j.sab.2006.06.001

17.

Kitzhaber

Orea

McFarlane

Manard

Andrews

. “Sparge Sampling of Molten Salts for Online Monitoring via Laser-Induced Breakdown Spectroscopy”. ACS Omega. 2025. 10(33): 37889–37897. 10.1021/acsomega.5c04988

18.

Andrews

H.B.

Kitzhaber

Z.B.

McFarlane

. “Real-Time Monitoring of Trace Noble Gases Using Laser-Induced Breakdown Spectroscopy: An Investigation of the Impact of Bulk Gas on Plasma Properties and Sensitivity”. Spectrochim. Acta, Part B. 2025. 230. 10.1016/j.sab.2025.107237

19.

Robb

Baird

Massengale

Kappes

Mulligan

P.L.

. “Facility to Alleviate Salt Technology Risks (FASTR): Design Report”. https://info.ornl.gov/sites/publications/Files/Pub188897.pdf [accessed 11 May 2026].

20.

Robb

Kappes

. “Facility to Alleviate Salt Technology Risks (FASTR): Commissioning Update”. https://info.ornl.gov/sites/publications/Files/Pub191132.pdf [accessed 11 May 2026].

21.

Kramida

Ralchenko

Yu.

. “NIST Atomic Spectra Database”. https://www.nist.gov/pml/atomic-spectra-database [accessed 11 May 2026].

22.

Python Software Foundation. “The Python Language Reference”. https://docs.python.org/3/reference/index.html [accessed 11 May 2026].

23.

Seabold

Perktold

. “Statsmodels: Econometric and Statistical Modeling with Python”. https://proceedings.scipy.org/articles/Majora-92bf1922-011 [accessed 11 May 2026].

24.

Pedregosa

Varoquax

Gramfort

Michel

, et al. “Scikit-Learn: Machine Learning in Python”. J. Mach. Learn. Res. 2011. 12: 2825–2830. 10.5555/1953048.2078195

25.

Diaz-Soriano

Alcaraz-Pelegrina

J.M.

Sarsa

Dimitrijevic

M.S.

Yubero

. “Simple and Analytical Function for the Stark Profile of the Hα line and its Application to Plasma Characterization”. J. Quant. Spectrosc. Radiat. Transfer. 2018. 217: 111–115. 10.1016/j.jqsrt.2018.05.031

26.

Naes

Martens

. “Multivariate Calibration .2. Chemometric Methods”. TrAC, Trends Anal. Chem. 1984. 3(10): 266–271. 10.1016/0165-9936(84)80044-8

27.

Wold

Martens

Wold

. “The Multivariate Calibration-Problem in Chemistry Solved by the PLS Method”. In: Kågström

Ruhe

, editors. Matrix Pencils: Lecture Notes in Mathematics. Berlin; Heidelberg: Springer, 1983. Pp. 286–293.

28.

Mocak

Bond

A.M.

Mitchell

Scollary

. “A Statistical Overview of Standard (IUPAC and ACS) and New Procedures for Determining the Limits of Detection and Quantification: Application to Voltammetric and Stripping Techniques (Technical Report)”. Pure Appl. Chem. 1997. 69(2): 297–328.

29.

Mermet

J.-M.

. “Limit of Quantitation in Atomic Spectrometry: An Unambiguous Concept?”. Spectrochim. Acta, Part B. 2008. 63(2): 166–182. 10.1016/j.sab.2007.11.029

30.

Allegrini

Olivieri

A.C.

. “IUPAC-Consistent Approach to the Limit of Detection in Partial Least-Squares Calibration”. Anal. Chem. 2014. 86(15): 7858–7866. 10.1021/ac501786u

31.

Felmy

H.M.

Cox

R.M.

Espley

A.F.

Campbell

E.L.

, et al. “Quantification of Hydrogen Isotopes Utilizing Raman Spectroscopy Paired with Chemometric Analysis for Application Across Multiple Systems”. Anal. Chem. 2024. 96(18): 7220–7230. 10.1021/acs.analchem.4c00802

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

4.91 MB