An industrial molten metal chemical analyzer based on laser-induced breakdown spectroscopy (LIBS) has been used to perform concentration analysis of important trace elements (Fe,Si,Cr,Mn,Ti, 20-3000 ppm) in liquid aluminum. In order to rule out accidental correlations between different elements in the spectral analysis, the impurity concentration of the measured samples was fully randomized. Reference concentration measurements were performed using arc-spark optical emission spectroscopy (spark-OES) on solid samples cast from the full volume of the LIBS-analyzed melt. For elements Fe, Cr, Mn and Ti, correlation coefficients and prediction uncertainty of LIBS measurements, using a linear correlation model, are shown to be determined mainly by random measurement error, which was of the order of 1% for both the LIBS analysis and the spark-OES analysis for concentrations above 100 ppm for all of the investigated elements. In the case of silicon, we postulate that inhomogeneous solidification is leading to a reduced absolute accuracy of spark-OES reference measurements and a corresponding increase in the minimum prediction uncertainty. The results confirm that LIBS analysis of molten aluminum can be used for on-line process control in the aluminum industry by providing measurement accuracy comparable of current industry-standard laboratory analysis methods.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
Currently, the most actively researched topic in the field of atomic spectrometry of ferrous and non-ferrous metals is laser-induced breakdown spectroscopy (LIBS), in large part because of its potential for rapid on-line chemical analysis in metals production and processing . Historically, the application of LIBS in analytical chemistry has been limited by its strong matrix dependence as well as its intrinsically small sampling volume, resulting in poor sensitivity and measurement results that are strongly influenced by sample inhomogeneity . Nevertheless, LIBS remains an actively researched area with significant potential for application within a wide range of industries .
A promising industrial application of LIBS is the real-time chemical analysis of molten aluminum in primary production, alloying, casting and recycling. Previous efforts to quantify minor elements in molten aluminum using LIBS date back at least 25 years, where it was already noted that the observed correlation between LIBS signals and the reference concentration of the melt was better in the liquid phase than in the solid phase . Later work has focused mainly on the development of immersive probes for in-situ analysis of liquid aluminum, where installation of prototypes in several industrial plants has also been reported [5,6]. Published data from on-site tests is still quite limited and has mainly focused on quantification of alloying elements (present in concentrations of 0.1-10 wt%) rather than trace elements . Recent work by the authors has confirmed that for many trace elements, the repeatability of LIBS measurements on as-produced liquid aluminum, performed on-site within a primary aluminum smelter, can be similar to or better than laboratory analysis on corresponding solid process samples, performed using industry-standard spark optical emission spectroscopy (spark-OES). Furthermore, in an Al-Si-Mg alloy, where the potential for formation of various intermetallic phases during solidification is larger, the advantage of measuring chemical composition in the liquid phase was found to be even more pronounced .
Any validation of the accuracy of LIBS analysis requires comparison against trusted reference measurements. Considering that the LIBS measurement and the reference sample are typically collected at different points in space and time, it becomes necessary to ensure a spatially and temporally homogeneous melt. The latter is particularly important in the case of the more volatile impurities or alloying elements . In practice, these criteria may not be easily fulfilled, especially in an industrial environment. Consequently, there may be no “true” reference to which to compare the LIBS analysis and neither measurement will necessarily be representative of the actual averaged chemical composition of the total volume of liquid metal in question. Furthermore, the industrial environment often offers limited possibilities of tailoring the melt chemistry to obtain a suitable range and distribution of concentration values for a given element to properly establish the accuracy of LIBS quantification, or for testing randomized concentrations of different elements in order to eliminate the possibility of accidental correlations  in the spectral analysis.
In the present work, we address these complications by presenting LIBS analysis of homogenized molten aluminum with intentionally added non-volatile impurities (Fe,Si,Cr,Ti,Mn) having uncorrelated concentrations, using an industrial molten-metal LIBS analyzer. We compare the results from the LIBS analysis with spark-OES analysis of solid samples, cast from the full volume of the same material. The correlation between the two conclusively demonstrates the potential of LIBS to deliver accurate chemical analysis of various minor elements in molten aluminum.
As-produced aluminum base material was obtained from Century Aluminum (Nordural smelter, Grundartangi, Iceland). The material was not otherwise purified, thus defining certain minimum impurity levels. Six melt compositions were prepared by weighing predetermined amounts of pure elements (>99% Fe and Si) or mixtures (AlTi80, AlCr80 and AlMn80, Rio Tinto) in granulated form and compacting the granulated materials, together with Al shavings, into holes drilled into solid blocks of the base material. The aluminum blocks were subsequently heated up to 950°C in a furnace and mechanically stirred at regular intervals. The actual impurity concentration of the resulting melts was somewhat lower than the initial weight fractions, as some of the added material may have formed insoluble phases and/or may have been disproportionately segregated into a layer of dross that was removed prior to performing the measurements.
Concentration measurements (discussed in more detail below) confirmed acceptable distributions of impurity levels, as shown in Table 1, with no significant correlation between the concentration values of any two elements (R2<0.5 in all cases). As is apparent from the table, the concentration intervals varied from approximately 200 ppm to approximately 2000ppm.
Samples of molten aluminum (190 ± 20 g for each melt composition) were transferred from the furnace to the sampling ladle of an industrial LIBS analyzer with robotic feeding (Model EA-2500, DT-Equipment). The ladle was coated with a non-wetting layer of boron nitride to prevent any cross-contamination between successive samples. An identical analyzer and sample feeding mechanism have been employed for on-line measurements of aluminum melt composition in an aluminum smelter . The LIBS measurement was carried out at a melt temperature of 730°C and consisted of 900 laser pulses at 30 Hz repetition rate (1064 nm wavelength, <15 ns pulse length) focused using a 500 mm lens. Spectral analysis of the collected plasma emission was carried out using an Echelle spectrometer and CCD camera. Additional measurement details are described in Ref. 9. For improved statistics, LIBS measurements were repeated 10 times for each melt sample. It should be noted that ablation of the liquid metal does not permanently alter the sample surface and repeated measurements therefore do not suffer from any measurable drift  in contrast to, e.g., crater effects observed in LIBS analysis of solid samples. Likewise, we did not observe any systematic drift over the total measurement time (5-6 minutes) used in the present work.
Immediately following the LIBS measurements, the molten aluminum in the sampling ladle was poured into an iron mold and cast into 2-3 cylindrical samples of 40 mm diameter (Fig. 1). The samples were subsequently milled to a depth of 1.2 mm (from the bottom surface), prior to spark-OES analysis, in accordance with sampling standards . Spark-OES analysis (Bruker Magellan Q8) was repeated 5 times for each solid sample at different positions, within a 10-mm radius from the center of the sample. The OES spectrometer was pre-calibrated using standard reference materials.
3. Results and discussion
A part of the LIBS spectrum obtained from a sample of liquid aluminum (melt no. 6) is shown in Fig. 2, together with two of the spectral peaks used for analysis (Fe II 260.708686 nm, Cr II 284.324559 nm). The resolving power (λ/Δλ) of the spectrometer was approximately 9000 and the measured peaks were well approximated with Voigt profiles. Other peaks used for the concentration analysis were Si I 250.690 nm, Mn II 294.9205 nm and Ti II 332.29339 nm (peak wavelengths obtained from the NIST database ). The choice of peaks for analysis was based on a large number of factors, including their relative intensity for the given plasma temperature and charge density, potential spectral overlap with other relevant peaks, positions relative to the Echelle orders, and the observed measurement-to-measurement stability of the peaks.
The standard deviation of LIBS measurements of individual elements in the aluminum melt (previously confirmed to be normally distributed ) is shown in Fig. 3. The data points were evaluated from the 10 repeated measurements for each melt composition. The random error for all the elements, across the investigated range of concentrations, can be approximately represented as 1.5 ppm + 1% of the measured concentration, as indicated by the dashed line in the figure. The differences in observed variance between individual elements can be mainly attributed to the signal-to-noise ratio of the relevant peaks in the LIBS spectrum. Other peaks for the same elements will therefore be expected to exhibit different %RSD values.
In the case of Fe and Si, repeated spark-OES measurements on individual solid samples showed a slightly higher standard deviation than observed in the LIBS measurements of the melt, in agreement with previous reports . Conversely, for Cr, Ti and Mn, the spark-OES measurements of the presently studied samples yielded somewhat lower standard deviation (typically around or below 1% for concentrations above 50 ppm) than observed in the LIBS analysis. However, the difference in OES-measured concentrations of a particular element between individual solid samples cast from the same melt was, in some cases, several times larger than the measurement standard deviation within each individual sample, indicating an uneven distribution of elements between samples. In order to get the best estimate of the averaged melt concentration from spark-OES measurements, the mass of the individual solid samples was therefore measured and a weighted average of the spark-OES results for a given melt was taken as being most representative for the average melt chemistry. The spark-OES measurement error (in Figs. 4 and 6) was taken as the intra-sample standard deviation or half of the maximum inter-sample deviation, whichever was larger.
The widest concentration interval and strongest correlation between spark-OES and LIBS results was obtained in the case of Cr, depicted in Fig. 4. A least-squares linear fit to the data was calculated, taking into account the uncorrelated random error in x and y values (York method ), yielding a Pearson correlation coefficient R=0.99997.
It is instructive to evaluate the significance of this metric, commonly used in the literature, by considering a hypothetical set of six equally spaced data points with perfectly correlated mean values and normally distributed (uncorrelated) errors in x and y. The probability of observing a given value of the correlation coefficient for such a dataset, evaluated by simulating 10.000 data sets using %RSD values typical of the measured data (Fig. 3), is shown in Fig. 5. The simulations show that in the case of 1% measurement error the probability of observing a correlation coefficient R>0.9995 is ≈80%, with the most likely value being 0.9998-0.9999. As expected, the distribution widens and shifts towards lower R values with increased random error. Increasing the number of data points, however, also shifts the distribution of correlation coefficients towards lower R values. Likewise, the probability of observing a correlation R>0.9999 decreases exponentially with the number of data points in the set, even though the determination of the fitting parameters of the regression becomes statistically more accurate. This can be reflected the determination of another commonly used metric, the limit of detection (LoD). One method to determine the LoD is to calculate the uncertainty of the y-intercept of the least-squares fit . Other parameters being equal, an increased number of (properly distributed) data points can therefore yield significant improvements in LoD, despite the value of the correlation coefficient being simultaneously reduced. The data shown in Fig. 4, for example, yields a LoD around 10 ppm for Cr, while a linear fit to approximately 100 data points for Cr concentration in the range 5-100 ppm from Ref. 9 yields a lower correlation coefficient but an order of magnitude improvement in LoD. #
Correlation analysis was carried out for the other investigated elements, as show in Fig. 6. The linear correlation coefficients of the measured data were >0.999 for Mn, Ti and Fe. As discussed above, this is consistent with the value of the correlation coefficient being mainly determined by the random error, indicating that other (non-systematic) errors arising, e.g., from the preparation of samples, inhomogeneous solidification or the effects of cross-correlation between different impurities can be considered negligible by comparison. Only the case of Si (R=0.9976) shows potential indications of deviations that cannot be accounted for by random measurement error in LIBS or OES alone. Other emission peaks for Si in the LIBS spectra were investigated, confirming that the residuals (deviations from linear relationships) were largely independent of the selected emission peak. We conclude that the most likely source of error in this case can be traced to inhomogeneous solidification prior to spark-OES measurement, in other words that for Si the measurements of the solid samples do not provide a fully accurate representation of the average melt concentration. The authors are presently carrying out a more detailed analysis on silicon segregation in cast reference samples for different silicon concentrations. The results of this analysis will be published elsewhere.
The linearity of the data in Figs. 4 and 6 confirms the absence of any significant non-linear effects in the LIBS analysis for the investigated emission lines. A non-zero constant offset (positive or negative) was generally observed in the fitting parameters and analysis of multiple LIBS peaks of the same element revealed this shift (and its sign) to be peak-dependent, pointing to systematic offsets in the fitting of the local spectral background in the LIBS analysis.
In a practical context, neither the correlation coefficient nor the LoD provide the most important measure of the usefulness of the LIBS analysis. Rather, the most important metric is the predictive power of the LIBS analysis within concentration ranges of the different elements expected to arise in realistic situations. Therefore, 95% confidence limits were determined for the linear fits to the data sets in Figs. 4 and 6. The half-width (in the x-direction) of the respective confidence intervals is plotted in Fig. 7. For Cr, Mn, Ti and Fe, the lowest predication uncertainty is of the same order as the typical random measurement error. For silicon, however, the lowest prediction uncertainty is about 5% of the measured concentration, which is significantly higher than the relative standard deviation of both the LIBS and the OES measurements. Additional studies of the actual distribution of silicon in the cast reference samples are needed to establish the source of reduced correlation between results on solid and liquid samples and the resulting increase in prediction uncertainty. Although it has not been the focus of the present paper we emphasize that, given the low measurement uncertainty and high correlation coefficients, the prediction uncertainty of the LIBS analysis for a given concentration range can, of course, be lowered further by using larger and/or specifically tailored calibration data sets.
We have shown that in-situ LIBS analysis of molten aluminum can provide accurate information of its impurity content, using melt samples with randomized amounts of elements Cr, Mn, Ti, Fe and Si. By using randomized impurity concentrations between samples, we eliminate the possibility of accidental correlations between different elements. Apart from Si, the correlation between LIBS results on the molten metal and reference OES analysis of solid samples was consistent with the correlation coefficient being limited only by the random error in the measurements, which for the present LIBS analysis was found to be of the order of 1.5 ppm plus 1% of the measured concentration. Based on these calibration sets, a prediction uncertainty of the same order as the random error was demonstrated, except for Si, where prediction uncertainty was significantly higher and attributed to increased uncertainty in the OES reference measurements. Our results also underline the importance of careful sample preparation and analysis for obtaining true reference measurements in the solid phase for demonstrating the full potential of LIBS analysis of metal in the liquid phase.
Icelandic Centre for Research (176618-0611).
SHG and KL: DT-Equipment (I,E,P). JM: DT-Equipment (P).
1. S. Carter, R. Clough, A. Fisher, B. Gibson, B. Russell, and J. Waack, “Atomic spectrometry update: review of advances in the analysis of metals, chemicals and materials,” J. Anal. At. Spectrom. 34(11), 2159–2216 (2019). [CrossRef]
2. V. Palleschi, Spectrosc. Eur., 2017, 29, 6.
3. S. Legnaioli, B. Campanella, F. Poggialini, S. Pagnotta, M. A. Harith, Z. A. Abdel-Salam, and V. Palleschi, “Industrial applications of laser-induced breakdown spectroscopy: a review,” Anal. Methods 12(8), 1014–1029 (2020). [CrossRef]
4. L. Paksy, B. Német, A. Lengyel, L. Kozma, and J. Czekkel, “Production control of metal alloys by laser spectroscopy of the molten metals. Part I. Preliminary investigations,” Spectrochim. Acta, Part B 51(2), 279–290 (1996). [CrossRef]
5. A.K. Rai, F.-Y. Yueh, and J.P. Singh, “Laser-induced breakdown spectroscopy of molten aluminum alloy,” Appl. Opt. 42(12), 2078 (2003). [CrossRef]
6. S.W. Hudson, J. Craparo, R. de Saro, and D. Apelian, “Applications of Laser-Induced Breakdown Spectroscopy (LIBS) in Molten Metal Processing,” Metall. Mater. Trans. B 48(5), 2731–2742 (2017). [CrossRef]
7. J. Herbert, J. Fernandez, R. De Saro, and J. Craparo, in: C. Chesonis (ed.), Light Metals, 2019, The Minerals, Metals & Materials Series, Springer, Cham, 945.
8. S.H. Gudmundsson, J. Matthiasson, and K. Leosson, in: A. Tomsett, ed. Light Metals, 2020. The Minerals, Metals & Materials Series. Springer, Cham, 860.
9. S.H. Gudmundsson, J. Matthiasson, B.M. Björnsson, H. Gudmundsson, and K. Leosson, “Quantitative in-situ analysis of impurity elements in primary aluminum processing using laser-induced breakdown spectroscopy,” Spectrochim. Acta, Part B 158, 105646 (2019). [CrossRef]
10. J. El Haddad, L. Canioni, and B. Bousquet, “Good practices in LIBS analysis: Review and advices,” Spectrochim. Acta, Part B 101, 171–182 (2014). [CrossRef]
11. ASTM International, Standard Practices for Sampling and Sample Preparation of Aluminum and Aluminum Alloys for Determination of Chemical Composition by Spark Atomic Emission Spectrometer E716, 2016.
12. A. KramidaYu Ralchenko., J. Reader and NIST ASD Team (2019). NIST Atomic Spectra Database (version 5.7.1). National Institute of Standards and Technology, Gaithersburg, MD. DOI: https://doi.org/10.18434/T4W30F
13. D. York, N.M. Evensen, M. López-Martínez, and J. De Basabe-Delgado, “Unified equations for the slope, intercept, and standard errors of the best straight line,” Am. J. Phys. 72(3), 367–375 (2004). [CrossRef]