Accelerating X-ray photoemission spectroscopy measurements using Bayesian super-resolution

This study applies Bayesian super-resolution to X-ray photoelectron spectroscopy (XPS), achieving up to a 20-fold reduction in measurement time while preserving data quality. Traditional XPS, crucial for surface analysis, typically requires extensive measurement durations. Our methodology significantly accelerates the process, as demonstrated with glass and Polytetrafluoroethylene samples, where we reduced measurement times by up to 1/20th without compromising spectral accuracy. This approach decreases noise levels and retains spectral integrity, offering a highly efficient solution for XPS. This innovation is particularly valuable in material science, enabling rapid, reliable surface analysis.

X-ray photoelectron spectroscopy (XPS) is a widely used measurement technique in the fields of materials science, chemistry, and chemical engineering.[19][20][21] The application of XPS has broadened to include techniques like depth profiling through sputtering and mapping measurements with focused X-ray scanning.However, a major challenge often encountered in this field is the prolonged duration of measurements.This issue can limit the method's efficiency in fast-paced research and industrial environments.Recently, we successfully applied Bayesian super-resolution techniques to spectroscopic data, enhancing both the resolution and the signal-to-noise ratio, which enabled precise characterization of spectral peaks and detailed analysis of spectral shapes and positions with increased accuracy. 22)In this study, we focus on employing the analysis method using Bayesian super-resolution to shorten the XPS measurement time.It is important to note that the term "resolution" can have different meanings across various fields.In digital image processing, "resolution" typically refers to the pixel interval length, contrasting with "spectral resolution," which denotes the ability to distinguish between closely spaced spectral lines.In our manuscript, "super-resolution" refers to narrowing the data interval, rather than enhancing spectral resolution.
Bayesian super-resolution is a technique for reconstructing high-resolution data from a set of low-resolution data.Initially developed as an image super-resolution algorithm, 23,24) it has been successfully applied to spectral data for high-precision evaluation. 22)The process of Bayesian super-resolution involves applying Bayes' rule under a predetermined prior distribution to infer sub-pixel shifts, effectively creating a narrow-interval spectrum from a broader data range.In this study, we have applied the Bayesian super-resolution to XPS data exactly as reported in our previous work, 22) without modification.In Bayesian super-resolution for XPS for example, spectra are initially acquired with energy shift.The technique focuses on reconstructing a high-resolution spectrum from these low-resolution inputs, achieved by defining a prior distribution for a narrow-interval spectrum and a conditional probability for each observation, inferring the energy shifts.In our previous study, we have successfully reconstructed a spectrum with a 0.01 cm −1 data interval from spectra obtained at approximately 0.8 cm −1 intervals by applying Bayesian super-resolution to Raman spectra.In the present study, we aim to use Bayesian super-resolution in XPS scanning measurements to get the required results faster.As the XPS measurement time in scanning mode is roughly proportional to the number of measurement points, expanding the energy steps can reduce the measurement time.Therefore, by conducting measurements with wider energy steps and reconstructing to the desired resolution using Bayesian superresolution, it is possible to maintain the resolution of the measurements while reducing the measurement time.
XPS data were acquired using a PHI5000 (Ulvac PHI Co. Ltd.) system.The measurement specimens for XPS were glass substrate and Polytetrafluoroethylene (PTFE) sheet.To obtain XPS spectra with relative shifts, the samples were measured with a slight charging caused by incomplete charge neutralization.Specifically, measurements were taken without ion neutralization and with electron neutralization only.The pass energy was set to 46.95 V for acquiring superresolution data, and measurements were conducted at a wide energy step of 0.4 eV.For comparison, measurements were also performed at a finer step of 0.05 eV.
Figure 1(a) shows the Si 2p XPS spectra obtained from a glass substrate for the purpose of super-resolution analysis, and Fig. 1(b) shows the changes in the peak center energies as determined by Gaussian fitting.It is observed that the peak center energies consistently appear at higher than the actual energy across all measurement data, which is thought to be due to incomplete charge neutralization.With an increasing number of measurements, a consistent shift of the peak positions towards the lower energy side is apparent.Notably, the extent of the energy shift surpasses the energy step (0.4 eV) used in the data acquisition.This indicates that the dataset is suitable for narrowing the data interval beyond the initial data interval, thereby enabling precise super-resolution processing.Figure 1(c) contrasts the XPS spectrum reconstructed using Bayesian super-resolution from a set of 80 spectra measured at 0.4 eV steps with the spectrum obtained from a tenfold accumulation at 0.05 eV steps, specifically for the Si 2p XPS spectra from a glass substrate.The spectrum reconstructed through super-resolution closely mirrors the accumulated measurements, indicating a reduction in noise and thereby resulting in higher quality data.Table I summarizes the values of the standard deviation of the noise in the background region, along with the measurement conditions.Notably, the standard deviation for the reconstructed spectrum is less than half that observed for the accumulated spectrum.Additionally, performing 80 measurements at 0.4 eV steps and accumulating data 10 times at 0.05 eV steps results in approximately the same measurement duration.Under these conditions, the spectrum obtained by applying super-resolution to the data collected at 0.4 eV steps and then reconstructing it to the 0.05 eV step resolution exhibits a significant noise reduction.This suggests that super-resolution can achieve data quality comparable to that of standard measurement methods, even with markedly reduced measurement times.
In a subsequent experiment, the dwell time was markedly reduced to just 10 ms, merely one-fifth of the standard duration.Consequently, the overall measurement time was also reduced to one-fifth of its usual length.Figure 2 displays the F 1s XPS spectrum obtained from a PTFE sheet.The data, reconstructed via spectral super-resolution, exhibit  narrower peak widths, attributable to the shortened dwell time and the consequent reduction in charge-up effects.Notably, despite the abbreviated measurement time, the quality of the data was effectively maintained.The standard deviation of the noise in the reconstructed spectrum was found to be 27, which is lower than the 32 observed in the accumulated spectrum.This result, achieved even with a shorter measurement duration, highlights the potential of super-resolution techniques in enhancing the efficiency of spectroscopic analyzes without compromising the quality and integrity of the data.
To explore the possibility of accelerating the XPS measurement process using Bayesian super-resolution, we conducted reconstructions with reduced numbers of datasets (N). Figure 3(a) shows the median standard deviation of noise when performing super-resolution with N datasets randomly selected from the original 80 measurements.For comparison, the standard deviation of noise for data accumulated over a 50 ms dwell time and the expected standard deviation for a 10 ms accumulation time are also presented.Even when the number of data sets is reduced to one-fourth (N = 20), the noise remains smaller than that of a 10 ms accumulation; however, further reduction in datasets results in a rapid increase in noise.Figures 3(b) and 3(c) display ten spectra, each reconstructed from N datasets (20 and 10, respectively) randomly selected from the original 80 measurements.The display of ten different spectra serves to illustrate the variability in reconstruction outcomes depending on the specific dataset chosen.When N is set to 20, the reconstruction results are predominantly consistent; however, with N at 10, some of the reconstructions exhibit significant noise.This variation highlights the trade-off between reducing the number of datasets and maintaining the integrity of superresolution reconstructions.These results demonstrate that reducing the datasets to a certain extent does not significantly degrade the quality.From these findings, it becomes clear that for an 8-fold super-resolution, approximately 20 measurement datasets are sufficient for effective reconstruction.By using about 20 measurement datasets and reducing the dwell time to 10 ms, which is one-fifth of the usual dwell time, it is possible to shorten the total measurement time by up to 1/20th without significant degradation of reconstructed spectrum.In this case, although the noise in the data increases, the instability in the reconstruction results due to having too few data points is less likely to occur.In addressing the accuracy of energy along the horizontal axis, it is crucial to acknowledge the integration of spectral data with inherent peak position shifts.Our methodology ensures the relative accuracy of spectral data.However, absolute accuracy in peak positions remains a challenge due to the inherent nature of XPS data and incomplete charge neutralization effects.Although our method makes it challenging to determine the precise locations of peak positions, it is considered applicable for quantitative analysis of elements.
Significant reduction in XPS measurement time achieved through Bayesian super-resolution, present a substantial leap forward in enhancing research and development throughput.By substantially shortening the time required for each measurement, this technique enables a higher number of experiments to be conducted within the same timeframe, effectively accelerating the pace of scientific discovery and material analysis in industrial application.Furthermore, the ability to perform depth profiling and mapping measurements with greater precision within limited time windows marks a remarkable advancement.This not only improves the efficiency of experiments but also ensures that more detailed and accurate data are obtained.Such precision in mapping is crucial for understanding the composition and structure of materials at a microscopic level, which has far-reaching implications in material science, nanotechnology, and other related fields.In essence, this technological breakthrough not only streamlines the measurement process but also elevates the quality of data, thereby paving the way for more rapid and accurate analyzes in various research and development domains.048001-3 © 2024 The Author(s).Published on behalf of The Japan Society of Applied Physics by IOP Publishing Ltd

Fig. 1 .Fig. 2 .
Fig. 1.(a) XPS spectra of Si 2p obtained from a glass substrate intended for super-resolution analysis at an energy step of 0.4 eV.(b) The energy shift determined by Gaussian fitting across the series of XPS measurements.(c) Comparison of the spectrum reconstructed using Bayesian super-resolution (SR) from 80 spectra at 0.4 eV steps against the spectrum from tenfold accumulated measurements at 0.05 eV steps.

Fig. 3 .
Fig. 3. (a) Median standard deviation of noise in XPS spectra reconstructed using Bayesian super-resolution with varying numbers of datasets (N), compared to noise levels in data accumulated over 50 ms and expected noise for 10 ms accumulation.Ten different spectra each, reconstructed from (b) 20 and (c) 10 datasets randomly selected from the original 80 measurements, demonstrating the effect of dataset reduction on reconstruction consistency and noise levels.