The field campaign was performed at the permanent station Milano Pascal of the ARPA Lombardia Air Quality Network from 6 July until 12 December 2023. This is an urban background site located in the eastern side of Milan, in the University campus area called “Città Studi” (45.478° N, 9.231° E; 122 ma.s.l.); the station is placed in a public park about 130 m from road traffic. The metropolitan city of Milan is the second most densely populated area in Italy (ca. 2300 inhabitantskm-2, almost doubled by daily commuters) and is located in the Po Valley, a well-known European pollution hotspot. The site is characterized by wintertime episodes of high pollutant concentrations, due to emissions from a variety of sources (e.g., residential heating, traffic, and industries) and prolonged atmospheric stability conditions related to the presence of the mountain chains of the Alps and the Apennines (Vecchi et al., 2007, 2009). Moreover, in Milan, more than 80 % of the days in a year are characterized by wind speed lower than 2 ms-1 (Vecchi et al., 2019). The site is well-documented with respect to gas-phase pollutants (e.g., NOx, SO2, O3), PM10 and PM2.5 chemical characterization, and source apportionment (e.g., Amato et al., 2016; Altuwayjiri et al., 2021).

The Xact^® 625i Ambient Metals Monitor (Cooper Environmental Services (CES), Beaverton, OR, USA) is an online energy dispersive XRF spectrometer, designed for continuous measurements of the elemental composition of ambient aerosol. The device operates using a reel-to-reel filter tape sampling technique, followed by the analysis of metals in the resulting PM spot through energy dispersive X-ray fluorescence (ED-XRF). Ambient air is drawn inside the instrument through a PM size-selective inlet, which was PM10 in this study, with a flow rate of 16.7 Lmin-1, and the PM is collected onto a Teflon filter tape. After each sampling interval is completed, the filter tape is automatically advanced to the XRF system, where the resulting PM deposit is irradiated with an X-ray tube (rhodium anode; max voltage: 50 kV; current: 1 mA) with three excitation conditions (see Table S1 in the Supplement), and the fluorescence X-rays are measured by a silicon drift detector (SDD). In the meantime, the next sample is collected on a clean spot of the filter tape, and the process is repeated during each sampling interval, which was set at 60 min for this study. The XRF spectra thus produced are automatically analyzed by a proprietary software for spectral analysis and elemental quantification, which is installed on the built-in computer. The software, through a linear least-squares deconvolution algorithm, fits each measured spectrum with a library of pure element reference spectra to obtain the concentration data for each calibrated element in ngm-3. Data can then be downloaded and monitored remotely with an internet connection. Sampling and XRF analysis are performed continuously and simultaneously, except for the time required for tape advancement (∼ 20 s). Quality assurance (QA) checks are performed every day at midnight for 30 min and consist of an energy calibration (using a rod coated with Cr and Nb) and an upscale measurement to monitor the stability of the instrument response (for Cr, Pb, and Cd). Therefore, the sample following midnight is collected with a sampling interval limited to 30 min (00:30–01:00 LT).

The instrument was located inside a temperature-controlled cabinet outside the ARPA Lombardia monitoring station. If any errors are detected during operation, the system halts sampling, ramps the X-rays down for safety, and displays the cause of the error. The instrument was configured to quantify 36 elements: Al, Si, P, S, Cl, K, Ca, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Ge, As, Se, Br, Rb, Sr, Y, Zr, Cd, In, Sn, Sb, I, Ba, Hg, Tl, Pb, and Bi; in addition, Nb was also detected for daily QA checks. Before the beginning of the experimental campaign, each of these elements was calibrated with a reference standard sample. For each element, 1σ interference-free MDLs (MDL1σ) for 1 h of sampling are reported in Table S2 in the Supplement, provided for Xact^® 625i following the approach reported in Currie (1977). In XRF analyses, MDLs are inversely proportional to the square root of the irradiation time, which in the case of Xact^® 625i corresponds to the sampling interval. Therefore, lower MDLs are reached for longer sampling durations.

Daily PM10 samples were collected on mixed cellulose ester membrane filters (47 mm diameter) with a SWAM Dual Channel Monitor (FAI Instruments, Rome, Italy) equipped with PM10 and PM2.5 inlets. The elemental composition of PM samples was determined offline by ED-XRF spectrometry in the laboratories of the Environmental Monitoring Sector of ARPA Lombardia. An Epsilon 4 spectrometer from Malvern Panalytical (Monza, Italy) was used for the ED-XRF analysis. Four different irradiation conditions, which are reported in Table S3 in the Supplement, were chosen to optimize the measurement of 19 elements, i.e., Al, Si, P, S, Cl, K, Ca, Ti, V, Cr, Mn, Fe, Ni, Cu, Zn, Br, Rb, Sr, and Pb. For these measurements, MDLs based on a 24 h sampling time were evaluated as 3 times the square root of the counts in the background below the peak of the element divided by the corresponding sensitivity in the blank filter (MDL3σ) (Jenkins et al., 1981; Lindgren, 2006) and are reported in Table S4 in the Supplement.

The Xact^® 625i measurements started on 6 July 2023 16:00 LT (local time) and ended on 12 December 2023 22:00 LT. The sampling interval of the instrument was set to 1 h. During the summer, in July and August, Xact^® 625i suffered from high temperatures during heat waves, causing the X-ray tube to reach temperatures above 45 °C. This led to automatic shutdowns of the measurements and to subsequent manual restarts, mostly performed remotely. The issue was mainly observed in the central hours of the day, from 13:00 to 16:00 LT. Nevertheless, it was still possible to attain a data coverage above 80 % for Xact^® 625i data in the central hours of the day during summertime. As a precautionary measure to avoid heat damage to the X-ray tube, Xact^® 625i was switched off from 12 to 24 August. During those days, a power failure in the ARPA Lombardia cabin caused also the interruption of daily measurements. Another power failure occurred from 22 October to 8 November, leading to a long pause of hourly measurements. The X-ray tube of Xact^® 625i started malfunctioning on 6 December. The issue could not be resolved, and the X-ray tube had to be replaced, resulting in a premature end to the experimental campaign.

Overall, the Xact^® 625i dataset consists of 2693 valid 1 h samples out of 3822 possible samples, attaining a coverage of 70 %. For the daily filters, the dataset consists of 149 samples out of 157 possible samples, reaching a coverage of 95 %. A timeline of the periods in which data are missing is reported for both hourly and daily measurements in Fig. S1 in the Supplement. A summary of the periods of interruption of the measurements lasting more than 12 h is reported in Table S5 in the Supplement. The number N of overlapping days with validated data is reported in Table  for each element considered for the intercomparison.

The Xact^® 625i data, which were originally reported in LT, were synchronized to daily samples and expressed in UTC + 1 time zone.

Following the approach of Furger et al. (2017) and Tremper et al. (2018), for the intercomparison study presented here, the MDL3σ was considered also for Xact^® 625i; indeed, the MDL3σ ensures high statistical confidence (99.7 %) and better comparability with previous literature works. Hereafter, MDL3σ will be referred to simply as MDL.

All the elements measured on the daily filters by offline ED-XRF have less than 35 % of their data points below their MDL. Among the elements detected by Xact^® 625i, 13 of them (P, Co, Ga, Ge, Y, Cd, In, Sn, Sb, I, Hg, Tl, and Bi) have more than 90 % of their data points below the MDL; therefore, these elements were excluded from the intercomparison analysis. V and Rb have > 70 % of their data points below the MDL, leading to a less robust intercomparison with offline ED-XRF (see Fig. S2 in the Supplement). In Table S6 in the Supplement, the number of data points with concentrations above the MDL is reported for each element measured by Xact^® 625i and by offline ED-XRF.

The intercomparison between Xact^® 625i and daily PM10 elemental concentrations was finally performed on 16 elements (Al, Si, S, Cl, K, Ca, Ti, Cr, Mn, Fe, Ni, Cu, Zn, Br, Sr, and Pb), which were measured by both techniques and were consistently above their MDLs.

An overview of the data recorded during the experimental campaign is given in Fig. , taking into account all available valid data of the elements considered for the intercomparison. To account for seasonal differences in terms of meteorology and emissions, data were divided into three periods: July–August, September–October, and November–December. The basic statistics of the dataset, including the mean, median, standard deviations, and 25th and 75th percentiles, are reported in Table S7 in the Supplement. As previously mentioned, the Xact^® 625i data coverage for July and August was impacted by the loss of data mainly related to the time interval 13:00–16:00 LT when hot temperatures caused the X-ray tube to switch off; therefore, the statistical robustness of the comparison is lower than in the other represented periods.

For the intercomparison between the two methods, the Xact^® 625i hourly data were averaged to 24 h to be comparable to the corresponding daily filter samples measured by offline ED-XRF. Every day, during QA checks performed from 00:00 to 00:30 LT, Xact^® 625i generated one sample with a 30 min temporal resolution so that this sample could be added to the 23 hourly samples to calculate 24 h daily means. This procedure implicitly assumes that the half-hour sample collected during the first hour of sampling is representative of the entire hour. The hypothesis was tested conducting 23.5 h weighted means on a small number of samples, following the method of Furger et al. (2017). Tests showed a difference of less than 3 % between the 23.5 h weighted mean and the 24 h mean, which was then chosen as the calculation method. For this reason, the Xact^® 625i data were aggregated to 24 h daily means.

As previously stated, during the campaign, summer days experienced heat waves, which caused Xact^® 625i to stop during the central hours of the day, leading to missing data. For this reason, the data coverage of Xact^® 625i was evaluated for each day of the experimental campaign. In order to avoid misestimation of daily Xact^® 625i concentrations, days with less than 18 h of valid data (75 % coverage) were excluded from the intercomparison. In addition, Xact^® 625i daily means were not calculated when more than 6 h of data were under the MDL for 1 day. In all comparisons, data under the MDL were replaced by 0.5 ⋅ MDL.

The comparisons between the daily PM10 elemental concentrations retrieved by ARPA Lombardia through offline ED-XRF and the daily means calculated from the Xact^® 625i hourly data were carried out using the Deming regression (Deming, 1943). This regression approach minimizes the sum of distances between the regression line and the X and Y variables, considering the experimental uncertainties in both variables. For the offline ED-XRF measurements, the uncertainty included contributions of 5 % from calibration standard uncertainty (U.S. EPA, 1999) and, for each spectrum, the contribution of counting statistics and fitting errors. For the Xact^® 625i measurements, the uncertainty included contributions of 5 % from calibration standard uncertainty (U.S. EPA, 1999) and an element-specific uncertainty derived from the spectral deconvolution calculated by the instrument software for each spectrum, which includes also the contribution of the flow and the sample deposit area. The mean relative uncertainties registered during the experimental campaign are reported for each element and for both online and offline methods in Table S8 in the Supplement.

The results of the intercomparison between the PM10 elemental concentrations retrieved offline and online are reported in Table . The Deming regression parameters are reported along with their uncertainties and the coefficient of determination of the linear regression; the amount of data (N) considered for the comparison after data reduction is also reported.

The scatterplots of the intercomparisons are presented in Figs. –. The time plots of the time series obtained by the two measurements methods are reported in Figs. S3–S6 in the Supplement. The 16 selected elements are compared by dividing them into three groups based on data characteristics.

The first group, Group A (Figs.  and ), includes K, Ca, Ti, Fe, Zn, Br, Sr, and Pb. This group shows excellent correlation between the two measurements methods (R2 > 0.95) and is characterized by slopes compatible with unity within 3 times the uncertainty of the fitted slope (3σ). For Ca, Ti, and Zn also, the intercepts are compatible with 0 within 3σ. Among this group, K, Ca, Ti, Fe, and Zn are measured by Xact^® 625i with relative uncertainties in the range 10 %–20 % (see Table S8). Br, Sr, and Pb are instead measured by Xact^® 625i with a higher uncertainty, on average 30 %–50 % (see Table S8), and Sr and Pb hourly data are also more frequently under the MDL (20 % of data).

The second group, Group B (Fig. ), consists of the elements Si, S, Mn, and Cu. This group is characterized by excellent correlation between the two measurements methods (R2 > 0.95), but, in contrast to Group A, the slopes of the regressions are not compatible with 1 within 3σ. Si and S are among the lightest elements measured by Xact^® 625i and, along with Al, can be tricky to measure with ED-XRF because of absorption effects due to the presence of air in the irradiation chamber (e.g., as typically occurs in the XRF online measurements) and/or self-absorption inside the coarse particles themselves (Hunter and Rhodes, 1972a, b; Van Grieken and Markowicz, 1993); these effects can lead to an underestimation of low-Z element concentrations. Nevertheless, looking at the results for Al, Si, and S, absorption effects seem not to be the cause of the observed discrepancy, as the Xact^® 625i data are typically higher than the offline ED-XRF analysis. Moreover, it should be noted that Si is detected by Xact^® 625i with mean uncertainties of 30 %, while S is detected with mean uncertainties of 10 %. In the case of Mn and Cu, concentrations provided by Xact^® 625i are constantly lower than the daily offline measurements by approximately 15 %.

A possible explanation for the observed discrepancies is related to the fact that, despite all samples being measured through the ED-XRF technique, the spectra analysis for quantitative analysis is different and – more importantly – the two instruments are not calibrated with the same set of certified standards, which can lead to different quantification of concentrations. However, the Xact^® 625i data of the elements of this group can still be validated when compared to an offline measurement technique and used for high-temporal-resolution elemental concentration assessment, after harmonization of the datasets.

The third group of elements, Group C (Fig. ), is composed of Al, Cl, Cr, and Ni. This group shows less comparability between the two methods, with R2 in the range of 0.67–0.87. Cl, Cr, and Ni are frequently close or under the MDL for both experimental techniques and are characterized by mean relative uncertainties in the range of 30 %–50 %. For these elements, the comparison could be improved by carrying out the Xact^® 625i measurements on a 2 h timescale. Among the 16 elements evaluated for the intercomparison, Al is the one with the highest MDL for the Xact^® 625i hourly measurements, and its hourly concentrations are under the MDL for nearly 35 % of data points, while Al offline data are always above the MDL. Al is also measured by Xact^® 625i with mean uncertainties of 50 %. As can also be seen in Fig. S6a, the Xact^® 625i time series of Al is characterized by a constant upward shift in background concentrations, which is not observed for the other elements. The measurement of Al with Xact^® 625i is complicated by the fact that the instrument uses an Al filter to carry out the analysis, as reported in Table S1; another possible issue could be that the X-rays hit some internal parts of the instrument, causing a significant increase in the background. Al concentrations cannot thus be corrected in a reliable way, and further improvements in the instrument should be considered to enhance Al detection. In the case of Cl, which shows quite scattered data, concentrations obtained by Xact^® 625i are on average higher than the ones measured offline on daily samples. This could be explained by the volatility of Cl. Xact^® 625i ED-XRF measurements are performed immediately after the collection of the sample, while daily PM10 filters are stored in the sampler at the monitoring station for up to 2 weeks before being taken to the laboratory for the offline ED-XRF analysis.

The results of this study represent a significant step forward from Park et al. (2014), which is – as far as we know – the only previous study available in the literature presenting a comparison between Xact^® hourly data and offline ED-XRF daily data. Park et al. (2014) conducted an experimental campaign with a forerunner version of Xact^® (Xact^® 620) in Gwangju, South Korea. The campaign was carried out during February 2011 and lasted only 1 month, focusing on the PM2.5 fraction. The Xact^® 620 model was the first commercially available near-real-time ambient metal monitor; it was able to detect elements starting from K and had higher detection limits (details can be found in Park et al., 2014), required more manual intervention for calibration and quality assurance processes, and had a more basic interface with limited remote access capabilities. The daily filters were measured offline with an Epsilon 5 ED-XRF spectrometer (Malvern Panalytical). The study compared the online and offline concentrations of 12 elements (K, Ca, Ti, V, Mn, Fe, Ni, Cu, Zn, As, Ba, Pb), nine of which were also analyzed in our study. For the nine common elements (K, Ca, Ti, Mn, Fe, Ni, Cu, Zn, Pb), they observed a mean R2 of 0.89 and a slope of 1.31, with Xact^® measurements on average 30 % higher than offline ED-XRF. In our study, for these nine elements, we found a much better correlation, with a mean R2 of 0.96 and slope of 0.94, which is closer to unity. Moreover, our study included seven elements (Al, Si, S, Cl, Cr, Br, Sr) that were not taken into account by Park et al. (2014), and the measurement campaign lasted for a longer period (6 months), giving more robustness to the results.

Overall, considering all 16 elements evaluated in this study, we found good correlation (mean R2 of 0.93) between the online and offline ED-XRF, with a mean slope of 1.11. The results are also in agreement with Tremper et al. (2018), who compared Xact^® measurements to ICP-MS daily measurements at three sites in the United Kingdom. They observed a mean R2 of 0.93 and a slope of 1.07 for the elements As, Ba, Ca, Cr, Cu, Fe, K, Mn, Ni, Pb, Se, Sr, Ti, V, and Zn. In the study by Furger et al. (2017), they found instead that the elemental measurements by an Xact^® 625i were, on average, 28 % higher than ICP-OES and ICP-MS measurements for S, K, Ca, Ti, Mn, Fe, Cu, Zn, Ba, and Pb.

A summary of previous literature studies and their characteristics is reported in Tables S9 and S10 in the Supplement. In these studies, several reasons for the differences observed between the Xact^® data and the offline techniques are described, and some of them are shortly reported below. In general, as specified in Tremper et al. (2018), the measured elements are chosen to represent a range of source categories (i.e., regulatory, traffic, industry), plus the internal standard (Pd for Xact^® 625 and Nb for Xact^® 625i). The number of elements that are actually quantified and thus included in the intercomparison results for each study depends on the ambient air concentrations and thus the site and MDL of the techniques.

In the U.S. EPA (2012) work, intercomparison results are available only for six elements, as the others were under the MDL of ICP-MS analysis and/or Xact^® measurements; weak regression parameters for Cu are explained by concentrations frequently close to the MDL of both techniques. In Furger et al. (2017), Xact^® and ICP-MS data showed high linearity and little scatter in the regressions for the elements S, K, Ca, Ti, Mn, Fe, Cu, Zn, Ba, and Pb; the relative mean difference of 28 % they found was attributed to many possible causes such as differences in the inlets used for the Xact^® instrument and the high-volume samplers for ICP-MS filter samples, a slightly different location of the samplers, possible calibration issues with the Xact^® device for S, values next to MDLs for one or both techniques, XRF particle-size-dependent self-absorption effects for the lighter elements, and line interferences or contaminations during the ICP-MS digestion and analysis procedures. Tremper et al. (2018) and Bhowmik et al. (2022) both mentioned similar reasons for the differences observed between Xact^® and ICP-MS filter data; in addition, blank filters were found to be variable, the standards used for Xact^® calibration had a much higher concentration than ambient air, and the calibration matrix differed from the sample matrix.