Quantitative Infrared-to-Terahertz Nanospectroscopy of Semiconductors Authors: Lukas Wehmeier, Zijian Zhou, Makoto Tsuneto, Heze Zhao, Stephanie N. Gilbert Corder, Hans A. Bechtel, Vladimir A. Martinez, Christopher C. Homes, D. N. Basov, Mengkun Liu, G. L. Carr Published: 2026 ABSTRACT Semiconductor technology now employs few-nanometer features, necessitating tools probing electronic properties on the same length scale. While the concentration of free charge carriers is routinely measured, the scattering rate remains challenging to access at the nanoscale. Here, we present ultrabroadband (5–50 THz) synchrotron infrared nanospectroscopy as a quantitative metrology tool for semiconductors. This technique can determine both the charge carrier concentration and scattering rate with percent-level accuracy, and it is inherently capable of ∼10 nm spatial resolution. We study silicon with different doping levels and confirm the method’s accuracy by statistical analysis and comparison with established far-field infrared spectroscopy. Near-field measurements systematically reveal charge-carrier concentrations ∼30% lower than far-field values, consistent with increased surface sensitivity and surface depletion. Our work establishes synchrotron infrared nanospectroscopy as a precise tool for quantitative nanoscale semiconductor characterization and paves the way toward all-optical characterization of surface depletion effects. Introduction A semiconductor’s electrical properties are largely determined by the concentration and scattering rate of its free charge carriers. Therefore, techniques that can determine these two parameters with good accuracy and precision are important for semiconductor characterization. While recent semiconductor technology employs structuring with a pitch of a few 10 to 100 nm and minimal structure sizes of only a few nm, only limited techniques provide information on both the charge carrier concentration and scattering rate at the required nanoscale spatial resolution. While some nanoscopic techniques exist for determining the charge carrier concentration, the scattering rate is particularly challenging to extract. Conventional infrared spectroscopy is a well-established and accurate method to determine both parameters, − but optical diffraction limits its spatial resolution to the micrometer range. Consequently, it cannot characterize modern semiconductor devices at the relevant nanometer length scale. Techniques based on atomic force microscopy (AFM) provide the necessary spatial resolution of a few to a few 10 nm. Several AFM techniques enable characterization of the local charge carrier concentration, including scanning capacitance microscopy (SCM, measuring capacitance), , conductive atomic force microscopy (c-AFM, measuring electrical resistance/conductivity), and Kelvin probe force microscopy (KPFM, measuring work function). , While these AFM-based methods measure the charge carrier concentration, they do not usually provide accurate information on the local charge carrier scattering rate. Scattering-type scanning near-field optical microscopy (s-SNOM) and infrared nanospectroscopy combine infrared spectroscopy with the high spatial resolution of atomic force microscopy. s-SNOM at infrared and terahertz frequencies has therefore been employed for numerous studies of semiconductors. − The technique effectively determines the charge carrier concentration of semiconductors with either relatively low charge carrier concentration (≲1·1017 cm–3, via terahertz nanoscopy at frequencies ≲3 THz including well below 1 THz) , or relatively high charge carrier concentration (≳1019 cm–3, via near- to mid-infrared spectroscopy at frequencies ≳20 THz). , Studies of intermediate charge carrier concentrations in the range of 1017 to 1019 cm–3 are less common, mainly due to technical challenges accessing the so-called terahertz gap (1–10 THz), where the relevant spectroscopic features of semiconductors with the above-mentioned intermediate doping are found. While limited s-SNOM studies at selected terahertz (THz) frequencies have demonstrated the large potential of nanospectroscopy for studying semiconductor samples, broadband nanospectroscopy studies of semiconductors in the spectral range around 10 THz remain elusive. Technical challenges also limit infrared nanospectroscopy studies of the charge carrier scattering rate, particularly for semiconductors with intermediate charge carrier concentrations. While the charge carrier concentration can be evaluated at selected frequencies or in a narrow spectral range, evaluation of the mobility or scattering rate requires spectroscopy in a broader spectral range. In the mid infrared, nanospectra covering a limited spectral width provide good precision , for carrier concentrations of around ≳1019 cm–3. However, due to the large width of the spectral features involved, ,, precision benefits from broader spectral coverage. Also, mid-infrared studies usually only provide upper limits for the charge carrier concentrations of samples with doping levels ≲1018 cm–3. To summarize, for semiconductor characterization, the ideal infrared nanospectroscopy simultaneously covers lower frequencies (e.g., into the terahertz gap to cover intermediate doping levels) and a broad spectral range in a single measurement (to efficiently evaluate scattering rates). However, such capabilities are not typically available in a single setup. Here, we address these challenges via ultrabroadband synchrotron-based infrared nanospectroscopy, which covers a spectral range from approximately 5 to 50 THz, i.e., about 1 order of magnitude. , For silicon, this spectral range implies highest sensitivity for doping levels around 1018 to 1020 cm–3, with reduced sensitivity beyond this range. We analyze our experimental nanospectroscopy data by combining analytical modeling with dedicated statistical data analysis. We present infrared nanospectroscopy spectra of well-characterized homogeneous bulk silicon samples with varying doping levels demonstrating that this technology can precisely (within a few percent) determine both the charge carrier concentration and scattering rate. As demonstrated in previous work, the technique consistently provides a spatial resolution of ∼10 nm. , Here, we confirm the accuracy and precision of infrared near-field nanospectroscopy by comparison with established far-field spectroscopy measurements. Results The optical properties of a simple semiconductor can be accurately characterized by the Drude model. Notably, it describes the broadband optical response of a material with only three parameters, two of which reflect the contributions of free charge carriers. These two parameters can be chosen either as the charge carrier concentration and mobility or, equivalently, as the plasma frequency ω p and scattering rate γ. Experimentally, ω p and γ are conveniently accessed via infrared spectroscopy, including both far- and near-field techniques. Figure a shows the setup of a far-field Fourier transform infrared (FTIR) spectrometer. Infrared radiation is created with a thermal lab-based source and guided through a symmetric Michelson interferometer, consisting of a beam splitter, a fixed reference mirror and a second, moveable reference mirror. After passing through the interferometer, the light is reflected off the sample at near-normal incidence and detected with a suitable detector while the moveable mirror is scanned. Details are given in Methods. For comparison, Figure b shows a synchrotron infrared nanospectroscopy setup. Infrared radiation is generated by a synchrotron light source (here, 22-IR-2 MET beamline of the National Synchrotron Light Source II, NSLS-II), and guided to an asymmetric Michelson interferometer. Like the symmetric Michelson, it consists of a beam splitter and a moveable reference mirror, but the fixed reference mirror is replaced by a near-field microscopy setup. The infrared radiation is focused onto an atomic force microscopy (AFM) tip. The light scattered from the tip is combined with the light from the reference arm and detected (details in Methods). , The advantage of the asymmetric Michelson interferometer is that it allows accessing the full complex optical response, i.e., both the real and imaginary parts of the optical signal (equivalently, optical amplitude and phase). Comparison between far-field and near-field spectroscopy of semiconductors. (a) Sketch of experimental far-field Fourier transform infrared (FTIR) reflectance spectroscopy setup with a thermal infrared source. (b) Sketch of infrared nanospectroscopy setup with a synchrotron light source (similar version previously published in reference, available under a CC-BY-NC-ND 4.0 license). (c) Calculated far-field FTIR reflectance spectrum and (d) near-field amplitude spectrum of model semiconductors with different charge carrier concentrations of 1016, 1017, 1018, 1 × 1019, 3 × 1019, and 5 × 1019 cm–3. The calculation uses an effective electron mass ,, m eff = 0.26 me, with bare electron mass me and scattering rate γ depending on the charge carrier concentration as described in reference . (e) Calculated far-field reflectance spectrum and (f) near-field amplitude spectrum of model semiconductors with constant charge carrier concentration of 1 × 1019 cm–3 and different scattering rates γ = 250, 300, 350, 450, 600, and 950 cm–1. For comparison, the spectral range covered in this work’s experiments is marked by black arrows in (c–f). We next describe how semiconductor properties can be derived from their infrared (nano)­spectra. In the Drude model, the dielectric function of a simple semiconductor is expressed as ε(ω)=ε∞−ωp2ω2+iγω with optical frequency ω and high-frequency dielectric constant ε ∞ = 11.7. , The plasma frequency, ω p , and charge carrier concentration n are directly related to each other as ωp2=ne2ε0meff with effective charge carrier mass m eff and vacuum permittivity ε 0. Figure c,d illustrate how the charge carrier concentration impacts the calculated infrared spectra acquired via far-field FTIR and infrared nanospectroscopy, respectively. Specifically, Figure c shows FTIR reflectance spectra of a semiconductor with a free-carrier concentrations 1016, 1017, 1018, 1 × 1019, 3 × 1019, and 5 × 1019 cm–3. The reflectance values are relative values that are normalized to the reflectance of a gold reference sample. As a common feature of all calculated spectra, we find a high reflectance close to 1 at low frequencies (<ω p *) and a lower reflectivity of ∼0.3 at high frequencies (>ω p *). The reflectance transitions from high to low in a relatively broad frequency window around the screened plasma frequency ωp*=ωp/ε∞ . The analytically calculated near-field amplitude spectra (finite-dipole model, , see Methods) show features that resemble the ones observed in the far-field spectra, albeit with a somewhat sharper transition, as is typical when comparing far- and near-field spectra. ,,, Figure e,f show calculations of how the scattering rate impacts the infrared spectra. For both the far- and near-field spectra, a larger scattering rate (lower mobility) broadens and flattens the transition region, which typically covers a spectral width of a few 100 cm–1. This highlights the need for sufficiently broad spectral coverage when estimating the scattering rate. In the experiments, we measure four semiconductor samples (commercial prime-grade silicon wafers from University Wafer) with varying doping concentration and type. We chose one sample with lower charge-carrier concentration (sample #1, Boron doping, p-type, vendor ID 1575, specified resistivity range ρ: 10–20 mΩ·cm), two with intermediate concentration (samples #2 + #3, Arsenic doping (n-type), vendor ID 2270, ρ: 1–5 mΩ·cm), and one with higher concentration (sample #4, Boron doping (p-type), vendor ID 2631, ρ: 1–5 mΩ·cm). Samples #2 and #3 are cut from different wafers with the same specifications. As such, we expect only small variations between these two samples, providing a good test for the precision of our experimental method. While the resistivity specifications of the commercial samples allow approximate estimates of doping and plasma frequency, the values are only specified in a relatively broad range. Using far-field FTIR spectroscopy (Figure a) as a well-established experimental technique, we obtain a more accurate sample characterization. Figure shows the FTIR reflectance spectra of the four samples. Symbols represent experimental values, and solid lines show Drude-model fits. The fit uses three free parameters, the two material parameters – the plasma frequency ω p,FF and scattering rate γFF – and a scaling constant to compensate for potential sample-reference misalignment. To indicate values from far-field (FF) FTIR reference measurements, we use the index “FF” for ω p,FF and γFF. As illustrated in Figure , the Drude model accurately describes the sample response. For samples #1 through #4, we find ω p,FF of 1316, 3060, 3262, and 6014 cm–1 and γFF of 487, 437, 435, and 763 cm–1, respectively, with fit uncertainties of 1 cm–1 (Table ). As expected, samples #2 and #3 have very similar properties. They have basically the same scattering rate and the plasma frequency of sample #2 is only 6% lower than that of sample #3. This similarity permits an efficient test of our nanospectroscopy technique. Experimental semiconductor far-field FTIR spectra. Symbols represent experimental data, and lines show an analytical fit using the Drude model to describe sample properties. While the experimental data was taken with a spectral resolution of 3.3 cm–1, for enhanced clarity, it is represented with fewer symbols here. 1Experimental Semiconductor Sample Parameters Derived from Near- and Far-Field Measurements sampledoping, vendor IDω p,NF [cm–1]γNF [cm‑1]ω p,FF [cm–1]γFF [cm–1]ρFF (ρvendor)[mΩ·cm] #1 p-type (B), 15751202 ± 105473 ± 191131648717 (10–20) #2 n-type (As), 22702547 ± 101380 ± 5030604372.8 (1–5) #3 n-type (As), 22702719 ± 149427 ± 5432624352.5 (1–5) #4 p-type (B), 26315085 ± 2611093 ± 18260147631.3 (1–5)aThe table shows the plasma frequency, ωp, and scattering rate, γ, both for the near-field (NF) and far-field (FF) measurements. The silicon samples are either p-type or n-type doped with Boron (B) or Arsenic (As), respectively. ρFF and ρvendor refer to the resistivity derived from the far-field measurements and the vendor-specified range, respectively. The near-field values, ωp,NF and γNF, are represented by their average value ± standard deviation. We note that the uncertainty of the average near-field values will be smaller than the standard deviation by a factor of √n≳4 where n refers to the number of measurements included in each average. The uncertainty of the far-field values, ωp,FF and γFF, is ± 1 cm–1 for all values and has been derived from the uncertainty of the fit. Experimental Semiconductor Sample Parameters Derived from Near- and Far-Field Measurements The table shows the plasma frequency, ωp, and scattering rate, γ, both for the near-field (NF) and far-field (FF) measurements. The silicon samples are either p-type or n-type doped with Boron (B) or Arsenic (As), respectively. ρFF and ρvendor refer to the resistivity derived from the far-field measurements and the vendor-specified range, respectively. The near-field values, ωp,NF and γNF, are represented by their average value ± standard deviation. We note that the uncertainty of the average near-field values will be smaller than the standard deviation by a factor of √n≳4 where n refers to the number of measurements included in each average. The uncertainty of the far-field values, ωp,FF and γFF, is ± 1 cm–1 for all values and has been derived from the uncertainty of the fit. The sample parameters derived above allow extraction of additional properties. Using the relation n=ωp2ε0meffe2 with m eff = 0.26 m e , (0.37 m e ) for n-doped (p-doped) samples, the plasma frequency yields carrier concentrations of 7.0 × 1018, 2.7 × 1019, 3.0 × 1019, and 1.5 × 1020 cm–3. Also, based on ω p,FF and γFF, we calculate the samples’ resistivity ρ using the formula ρFF=1ε0γFFωp,FF2 . In wavenumber units (γ̃ and ω̃ p 2 in cm–1), this yields the simple rule-of-thumb ρFF[Ω·cm]=60γ̃[cm−1]ω̃p2[cm−1] (see Methods). For samples #1, #2, #3, and #4, we find resistivities of 17, 2.8, 2.5, and 1.3 mΩ·cm, respectively, which is within the vendor’s specified range of 10–20 mΩ·cm for sample #1 and 1–5 mΩ·cm for samples #2-#4. Figure shows representative near-field measurements of the four silicon samples. While the far-field reflectance spectra are real-valued, the near-field spectra are complex-valued, and each sample yields a full spectrum for both the real (Figure a) and imaginary part (Figure b). Symbols in Figure show the experimental results. Using a combination of the Drude model for the sample and the finite dipole model for the tip–sample interaction, we analytically fit the measured spectra. The fits, shown as solid lines in Figure , generally provide a very good match to our experimental data. As described in the Methods section, our fitting procedure includes experimental parameters such as the tip length and angle of incidence, as well as the two relevant material parameters, the plasma frequency and scattering rate. To reduce the number of free parameters, many model parameters (e.g., tip radius) and material parameters (e.g., ε ∞ ) are kept fixed. We note that in our fitting procedure we consider the full, complex data set, i.e., we simultaneously fit the real and imaginary part of the near-field data. Compared to analyzing only the real-valued near-field amplitude spectrum, using the complex spectrum significantly enhances the quality and reliability. In total, our model uses 6 free parameters, 4 experimental parameters (Methods) and 2 sample parameters (plasma frequency and scattering rate). Note that the number of free parameters per spectrum is the same for the far- and near-field spectra, namely 3 parameters per real-valued spectrum (6 parameters for the complex-valued spectrum). Exemplary experimental semiconductor near-field spectra. (a) Real and (b) imaginary part of the complex near-field spectra of four different semiconductor samples. The symbols represent experimental data, and the lines show the analytical fit. The latter uses the Drude model to describe the sample properties and the finite dipole model (FDM) to describe the near-field interaction (additional detail in Methods). The colors refer to the sample, green: sample #1, blue: sample #2, purple: sample #3, orange: sample #4. The near-field data shown above is demodulated at the second harmonic of the tip frequency. Figure shows near-field spectra taken with a specific AFM tip type (RockyMountain nanotechnology), at a specific tip tapping amplitude (100 nm), with the detector signal being demodulated at the second harmonic (n = 2) of the tip tapping frequency Ω. Demodulation of the detector signal at a higher harmonic nΩ enables separation of the near-field signal from the far-field background. , To improve statistics on accuracy and precision, we repeated the measurement with different tip tapping amplitudes (100, 112 nm), a different tip type (Attocube nanoFTIR tip), and evaluated the different harmonics n = 2,3,4 of the signal. For each sample, this leads to a few 10s of near-field spectra. For each complex spectrum, we performed the analytical fit described above and retrieved a specific value for plasma frequency and scattering rate. Values resulting from individual spectra are depicted as points in Figure , with plasma frequency on the horizontal axis and scattering rate on the vertical axis. Overview of experimental semiconductor sample parameters derived from near- and far-field measurements. The colors refer to the sample, green: sample #1, blue: sample #2, purple: sample #3, orange: sample #4. Dots and crosses represent sample parameters retrieved from near-field spectra, and triangles represent sample parameters retrieved from far-field spectra. For better statistics, near-field spectroscopy has been repeated with different AFM tip types, different AFM settings, and on different measurement days. Also, spectra from different harmonics n = 2,3,4 have been included. Each dot represents a single near-field measurement with crosses representing the weighted average value of the near-field data for each sample. For clearer visualization, the plot shows a reduced number of data points, which are derived from the 50% of individual spectra with the lowest fitting errors. Values from all spectra are included in the weighted average. The shaded areas represent the kernel density estimation (KDE) of the fitting results, illustrating the distribution range of the parameters for each sample. Despite the variation of the experimental parameters, for all samples the plasma frequencies derived from the different near-field spectra show a standard deviation of less than 10%. Considering the strong general dependence of the near-field response on experimental parameters such as the AFM tip shape, tapping amplitude, and demodulation order, this represents good agreement and indicates robustness of our method against variation of experimental parameters. We also investigated possible systematic dependencies of the derived sample properties on the experimental parameters, e.g., a dependence of the plasma frequency on harmonic demodulation order. No systematic trend has been found, likely due to the experimental uncertainty. The scattering rates for samples #2 and #3 both show a standard deviation of 13%. For samples #1 and #4 the derived scattering rates vary more strongly with standard deviations of 40% and 17%, respectively. The larger variance for samples #1 and #4 compared to samples #2 and #3 is explained by the spectral range of highest sensitivity of our nanospectroscopy setup. The characteristic features of each spectrum, particularly the trough of the real part and the peak of the imaginary part, are in the spectral range where our nanospectroscopy setup is most sensitive for samples #2 and #3, but closer to the edges of our sensitivity range for samples #1 and #4. The cross symbols in Figure mark the weighted averages of the sample parameters derived from the near-field nanospectroscopy data for each sample. As summarized in Table , we find average plasma frequencies of 1202, 2547, 2719, and 5085 cm–1, and scattering rates of 473, 380, 427, and 1093 cm–1 for samples #1, #2, #3, and #4, respectively. It is interesting to compare the sample parameters obtained via infrared near-field nanospectroscopy with the far-field FTIR reference measurements. For all samples, the scattering rate derived from the nanospectroscopy measurements agrees with the far-field reference within two standard deviations. The far-field reference also lies within the scattering rate range estimated via kernel density estimation (shaded region in Figure ). For samples #1, and #3, the near-field measurements of the scattering rates agree with the far-field result within 3%, which is remarkable. For samples #2 and #4, the deviation between near- and far-field measurements is −13% and +43%, respectively, corresponding to 1–2 standard deviations. While this might indicate a systematic deviation, the deviation might also stem from experimental uncertainty. Sample #4 shows a relatively large variance of scattering rates. However, as our setup had been optimized for lower frequencies, this is not a limitation of the method and higher accuracy for samples with larger plasma frequency (such as sample #4) could be achieved by optimizing the setup for higher-frequency sensitivity. To summarize, the scattering rates obtained via infrared nanospectroscopy and far-field reference agree well within the uncertainties, and our measurements do not indicate systematic deviation in scattering rate. For the plasma frequency, the comparison between near-field and far-field values shows a clear and systematic deviation. We find the plasma frequency from near-field measurements is consistently below the far-field FTIR reference. The near-field value is 9%, 17%, 17%, and 15% smaller for samples #1, #2, #3, and #4, respectively (Table ), resulting in an average deviation of 14%. This difference exceeds the standard deviation of the near-field measurements for all samples. For samples #2–#4, the difference exceeds the standard deviation by more than a factor of 3. We therefore find a systematically lower plasma frequency from the near-field method. However, both the near-field and far-field method provide very comparable relative differences between the samples. For example, the near-field (far-field) measurements indicate that the plasma frequencies for sample #1, #2, and #4 differ from sample #3 by −56% (−60%), −6.3% (−6.2%), and +87% (84%), respectively. The excellent agreement of the relative values between the near-field method and the far-field reference suggests a remarkable precision of the near-field method. Discussion For the scattering rate, near-field and far-field measurements agree within the measurement uncertainty, with particularly good agreement for samples #1 to #3. However, for the plasma frequency, we find that the near-field values systematically deviate from the far-field reference results. On average, the near-field results for the plasma frequency are about 14% lower than the FTIR reference measurements, exceeding the measurement uncertainty. Two main explanations may account for this:1. Model Inaccuracies There might be a systematic error in our data analysis. As the Drude model is well established for far-field FTIR data, we expect any model inaccuracies to arise mainly in the near-field analysis. For example, there could be a spectral shift in the near-field response that is not well represented by the finite-dipole model we use. This spectral shift would need to be different from the common observation that the peak of the near-field signal appears shifted relative to the plasma frequency since this is considered in the finite dipole model and the simpler point dipole model. ,,, By employing a layered finite dipole model (Methods), , we can exclude dielectric loading due to possible surface oxidation as the cause of the observed spectral shift. Assuming physical explanations representing actual sample properties can also be excluded, it will be straightforward to correct for any remaining model-induced systematic deviation between near- and far-field results. For example, multiplying the derived plasma frequency by a constant factor could compensate for the known deviation. As a more advanced approach, the analytical model used to describe the tip–sample interaction could be improved. Other analytical models exist such as the point-dipole model or the lightning-rod model. Also, numerical simulations , or machine learning , may provide improved descriptions of the tip–sample interaction. Finally, any systematic deviation can be corrected by calibrating the method using a reference semiconductor sample with known properties. The applicability of a reference sample for calibration is confirmed by the good relative precision of the near-field method as demonstrated by the accurate reproduction of the relative plasma-frequency differences between samples #2 and #3.2. Surface Depletion Layer While inaccuracies of our analytical description are one possible explanation, the finite-dipole model used in the present study is well established and validated in several works. ,, It is therefore important to consider physical explanations as well. Near-field studies probe less well-defined optical polarization states (Methods) and are much more surface sensitive than their far-field counterparts. While depending on experimental and sample parameters, near-field studies typically probe depths of a few 10 nm up to ∼100 nm. ,,, For most of the spectral range, this is more than an order of magnitude smaller than the probing depth of the far-field FTIR reference measurements (calculation in Methods). Due to the smaller probing depth of the near-field technique, the lower plasma frequency observed via near-field spectroscopy might be explained by a lower charge carrier concentration at the surface of the semiconductors (surface depletion). , Although an increased effective mass could partially account for differences in ω p , possibly via the different polarization states probed via near- and far-field experiments, we do not explore issues related to the effective mass anisotropy of the transport properties here but note that this could be a subject of further studies. Focusing on changes related to the charge carrier concentration n∝ω p 2, the observed ∼14% reduction in ω p suggests an effective surface charge carrier depletion of about 30%, averaged within the probing depth of the near-field technique. Considering prior reports and earlier infrared nanospectroscopy experiments, , surface depletion appears a plausible explanation for the observed decrease in plasma frequency. For the doping concentrations studied here (calculations in Methods), the width of the depletion layer of doped Si in contact with a SiO2 surface layer is expected to be a few nm to a few 10 nm. This matches the probing depth of our near-field technique and suggests infrared nanospectroscopy can probe semiconductor surface depletion effects. In principle, a thin surface depletion layer should more strongly influence the near-field signal at higher harmonic demodulation orders. However, we estimate this dependence is below our experimental uncertainty and we did not observe a systematic trend with demodulation order. Nevertheless, the expected dependence may provide opportunities for detailed nanospectroscopy studies of surface depletion layers in future measurements with further improved experimental precision and accuracy. We note that the smaller probing depth of near-field techniques increases the relative importance of effects related to the finite charge-carrier mean free path. However, such effects are more relevant for lower charge-carrier concentrations than studied here and primarily impact γ rather than ω p (Methods). Also, the observed surface depletion effect could be enhanced beyond the expected intrinsic depletion at the Si surface (Methods), e.g., by tip-induced band-bending. Nonlinear effects due to strong field enhancement at the tip might increase the scattering rate. However, such effects are expected primarily at lower frequencies or stronger light fields than those studied here. 1. Model Inaccuracies There might be a systematic error in our data analysis. As the Drude model is well established for far-field FTIR data, we expect any model inaccuracies to arise mainly in the near-field analysis. For example, there could be a spectral shift in the near-field response that is not well represented by the finite-dipole model we use. This spectral shift would need to be different from the common observation that the peak of the near-field signal appears shifted relative to the plasma frequency since this is considered in the finite dipole model and the simpler point dipole model. ,,, By employing a layered finite dipole model (Methods), , we can exclude dielectric loading due to possible surface oxidation as the cause of the observed spectral shift. Assuming physical explanations representing actual sample properties can also be excluded, it will be straightforward to correct for any remaining model-induced systematic deviation between near- and far-field results. For example, multiplying the derived plasma frequency by a constant factor could compensate for the known deviation. As a more advanced approach, the analytical model used to describe the tip–sample interaction could be improved. Other analytical models exist such as the point-dipole model or the lightning-rod model. Also, numerical simulations , or machine learning , may provide improved descriptions of the tip–sample interaction. Finally, any systematic deviation can be corrected by calibrating the method using a reference semiconductor sample with known properties. The applicability of a reference sample for calibration is confirmed by the good relative precision of the near-field method as demonstrated by the accurate reproduction of the relative plasma-frequency differences between samples #2 and #3. Model Inaccuracies There might be a systematic error in our data analysis. As the Drude model is well established for far-field FTIR data, we expect any model inaccuracies to arise mainly in the near-field analysis. For example, there could be a spectral shift in the near-field response that is not well represented by the finite-dipole model we use. This spectral shift would need to be different from the common observation that the peak of the near-field signal appears shifted relative to the plasma frequency since this is considered in the finite dipole model and the simpler point dipole model. ,,, By employing a layered finite dipole model (Methods), , we can exclude dielectric loading due to possible surface oxidation as the cause of the observed spectral shift. Assuming physical explanations representing actual sample properties can also be excluded, it will be straightforward to correct for any remaining model-induced systematic deviation between near- and far-field results. For example, multiplying the derived plasma frequency by a constant factor could compensate for the known deviation. As a more advanced approach, the analytical model used to describe the tip–sample interaction could be improved. Other analytical models exist such as the point-dipole model or the lightning-rod model. Also, numerical simulations , or machine learning , may provide improved descriptions of the tip–sample interaction. Finally, any systematic deviation can be corrected by calibrating the method using a reference semiconductor sample with known properties. The applicability of a reference sample for calibration is confirmed by the good relative precision of the near-field method as demonstrated by the accurate reproduction of the relative plasma-frequency differences between samples #2 and #3. 2. Surface Depletion Layer While inaccuracies of our analytical description are one possible explanation, the finite-dipole model used in the present study is well established and validated in several works. ,, It is therefore important to consider physical explanations as well. Near-field studies probe less well-defined optical polarization states (Methods) and are much more surface sensitive than their far-field counterparts. While depending on experimental and sample parameters, near-field studies typically probe depths of a few 10 nm up to ∼100 nm. ,,, For most of the spectral range, this is more than an order of magnitude smaller than the probing depth of the far-field FTIR reference measurements (calculation in Methods). Due to the smaller probing depth of the near-field technique, the lower plasma frequency observed via near-field spectroscopy might be explained by a lower charge carrier concentration at the surface of the semiconductors (surface depletion). , Although an increased effective mass could partially account for differences in ω p , possibly via the different polarization states probed via near- and far-field experiments, we do not explore issues related to the effective mass anisotropy of the transport properties here but note that this could be a subject of further studies. Focusing on changes related to the charge carrier concentration n∝ω p 2, the observed ∼14% reduction in ω p suggests an effective surface charge carrier depletion of about 30%, averaged within the probing depth of the near-field technique. Considering prior reports and earlier infrared nanospectroscopy experiments, , surface depletion appears a plausible explanation for the observed decrease in plasma frequency. For the doping concentrations studied here (calculations in Methods), the width of the depletion layer of doped Si in contact with a SiO2 surface layer is expected to be a few nm to a few 10 nm. This matches the probing depth of our near-field technique and suggests infrared nanospectroscopy can probe semiconductor surface depletion effects. In principle, a thin surface depletion layer should more strongly influence the near-field signal at higher harmonic demodulation orders. However, we estimate this dependence is below our experimental uncertainty and we did not observe a systematic trend with demodulation order. Nevertheless, the expected dependence may provide opportunities for detailed nanospectroscopy studies of surface depletion layers in future measurements with further improved experimental precision and accuracy. We note that the smaller probing depth of near-field techniques increases the relative importance of effects related to the finite charge-carrier mean free path. However, such effects are more relevant for lower charge-carrier concentrations than studied here and primarily impact γ rather than ω p (Methods). Also, the observed surface depletion effect could be enhanced beyond the expected intrinsic depletion at the Si surface (Methods), e.g., by tip-induced band-bending. Nonlinear effects due to strong field enhancement at the tip might increase the scattering rate. However, such effects are expected primarily at lower frequencies or stronger light fields than those studied here. Surface Depletion Layer While inaccuracies of our analytical description are one possible explanation, the finite-dipole model used in the present study is well established and validated in several works. ,, It is therefore important to consider physical explanations as well. Near-field studies probe less well-defined optical polarization states (Methods) and are much more surface sensitive than their far-field counterparts. While depending on experimental and sample parameters, near-field studies typically probe depths of a few 10 nm up to ∼100 nm. ,,, For most of the spectral range, this is more than an order of magnitude smaller than the probing depth of the far-field FTIR reference measurements (calculation in Methods). Due to the smaller probing depth of the near-field technique, the lower plasma frequency observed via near-field spectroscopy might be explained by a lower charge carrier concentration at the surface of the semiconductors (surface depletion). , Although an increased effective mass could partially account for differences in ω p , possibly via the different polarization states probed via near- and far-field experiments, we do not explore issues related to the effective mass anisotropy of the transport properties here but note that this could be a subject of further studies. Focusing on changes related to the charge carrier concentration n∝ω p 2, the observed ∼14% reduction in ω p suggests an effective surface charge carrier depletion of about 30%, averaged within the probing depth of the near-field technique. Considering prior reports and earlier infrared nanospectroscopy experiments, , surface depletion appears a plausible explanation for the observed decrease in plasma frequency. For the doping concentrations studied here (calculations in Methods), the width of the depletion layer of doped Si in contact with a SiO2 surface layer is expected to be a few nm to a few 10 nm. This matches the probing depth of our near-field technique and suggests infrared nanospectroscopy can probe semiconductor surface depletion effects. In principle, a thin surface depletion layer should more strongly influence the near-field signal at higher harmonic demodulation orders. However, we estimate this dependence is below our experimental uncertainty and we did not observe a systematic trend with demodulation order. Nevertheless, the expected dependence may provide opportunities for detailed nanospectroscopy studies of surface depletion layers in future measurements with further improved experimental precision and accuracy. We note that the smaller probing depth of near-field techniques increases the relative importance of effects related to the finite charge-carrier mean free path. However, such effects are more relevant for lower charge-carrier concentrations than studied here and primarily impact γ rather than ω p (Methods). Also, the observed surface depletion effect could be enhanced beyond the expected intrinsic depletion at the Si surface (Methods), e.g., by tip-induced band-bending. Nonlinear effects due to strong field enhancement at the tip might increase the scattering rate. However, such effects are expected primarily at lower frequencies or stronger light fields than those studied here. Conclusion Our findings demonstrate that broadband infrared nanospectroscopy provides precise and accurate measurements for both the plasma frequency and scattering rate of a semiconductor. The near-field measurements provide a standard deviation (precision) of about 10% for the plasma frequency and a few 10% for the scattering rate. For samples #2 and #3, for which the characteristic spectral features are in our spectral region of highest sensitivity, the precision for both parameters is better than 10%. While the standard deviation captures the precision of our method, we confirm its accuracy by comparing near-field results to established far-field FTIR reflectance measurements. For the scattering rate, near-field and far-field measurements agree within the measurement uncertainty, suggesting the accuracy is primarily limited by precision. However, for the plasma frequency, the near-field values systematically deviate from the far-field reference results. While relative sample differences are derived with remarkable accuracy of a few percent, the near-field results for the absolute plasma frequency are on average ∼14% lower than the FTIR reference measurements, exceeding the measurement uncertainty. While this deviation might be explained by a systematic error in our analytic description of the near-field interaction and could in principle easily be corrected for, the deviation may also be explained by surface depletion reducing charge carrier by ∼30%. The latter explanation is supported by previous literature results and calculations (see Methods, Figure ) suggesting depletion widths of a few nm to a few 10 nm for our samples, matching the typical ≲100 nm probing depth of infrared near-field nanospectroscopy. This indicates that infrared near-field nanospectroscopy is well suited to study surface-depletion effects in semiconductor devices. Our systematic study paves the way toward synchrotron infrared nanospectroscopy as a versatile and quantitative tool for nanoscale semiconductor characterization. Beyond surface-sensitive probing of the plasma frequency, the demonstrated accuracy and precision in characterizing charge-carrier scattering rate is a remarkable advantage since few nanoscopic techniques can access it reliably. Quantitative infrared-to-terahertz nanospectroscopy therefore provides a versatile platform for nanoscale charge-carrier analysis, applicable not only to silicon and conventional semiconductors but also to two-dimensional and van der Waals materials and their heterostructures. Calculated Infrared Penetration Depth and Depletion Width. (a) Calculated Infrared Penetration Depth δ=λ4πIm(ϵ) providing an approximation for the probing depth of far-field infrared reflectance spectroscopy. The colors refer to the sample, green: sample #1, blue: sample #2, purple: sample #3, orange: sample #4. (b) Calculated depletion width at the sample surface as a function of doping concentration. For the doping concentrations relevant for our samples of >1018 cm–3, the depletion width is in the order of a few 1 nm to a few 10 nm, remaining well below the probing depth of far-field FTIR spectroscopy as described by the penetration depth. Methods Sample Preparation Commercial silicon (Si) wafers have been purchased through University Wafers. The wafers are [100]-oriented and the surface is polished. The wafer thickness is 0.53 mm for sample #1, and 0.25 mm for samples #2–#4. The wafers have diameters of >50 mm. Samples of about 5 mm × 5 mm were cut from the wafers using a diamond scribe. The smaller sample size is more convenient for the near-field measurements and guarantees that near-field and far-field measurements have been taken at approximately the same location. Far-Field Fourier Transform Infrared (FTIR) Spectroscopy For Fourier transform infrared (FTIR) spectroscopy, we use a commercial Bruker Vertex 80v spectrometer with its internal thermal infrared light source. The sample is illuminated in a reflection geometry at near-normal incidence (Figure a). Due to the near-normal incidence and weak focus, the FTIR setup predominantly probes the response polarized within the sample surface plane. Due to the unpolarized thermal light source, the polarization within the sample surface plane is uniformly probed. To reduce infrared absorption in air, the whole optical setup is evacuated for the measurement. The aperture size is 1.5 mm, which avoids spilling of the infrared light beyond the sample. The light reflected from the sample is measured with a deuterated l-alanine doped triglycine sulfate (DLaTGS) detector with a diamond window, enabling ultrabroadband detection across the full infrared spectral range. Spectra are taken with a resolution of 3.3 cm–1. For measurements at low frequencies (<900 cm–1), we employ a silicon (Si) beamsplitter. For measurements at higher frequencies (>500 cm–1), we employ a potassium bromide (KBr) beamsplitter. To reduce the effect of potential instrument drift, we take a gold reference spectrum both before and after the sample measurement. The sample spectrum is then normalized to the average of the two reference spectra. After multiplying the low-frequency (Si beamsplitter) spectrum with a constant, the low- and high-frequency (KBr beamsplitter) show very good overlap in the spectral range from 500 to 900 cm–1. The two spectra are merged by linearly increasing the relative weight of the high-frequency spectrum from 0 at ≤ 500 cm–1 to 1 at ≥ 900 cm–1. The resulting data is shown in Figure . Synchrotron-Based Infrared Nanospectroscopy Scattering-type scanning near-field microscopy (s-SNOM) employs the field enhancement at the tip of an atomic force microscope to provide optical and infrared information with a wavelength-independent spatial resolution of a few 10 nm. , When combined with broadband illumination and Fourier transform infrared (FTIR) spectroscopy, s-SNOM allows one to acquire broadband infrared spectra at the nanoscale. Here, we perform synchrotron infrared nanospectroscopy , with a commercial s-SNOM setup (NeaSnom by attocube, Germany) at the MET beamline of the National Synchrotron Light Source II (NSLS-II) at Brookhaven National Laboratory, Upton. The setup has previously been described in references. ,, For broad spectral coverage in a single measurement, it employs a diamond beam splitter and a liquid-helium-cooled mercury cadmium telluride (MCT) detector. To avoid absorption of infrared radiation in air, the near-field microscopy setup is kept in a nitrogen gas environment (>99% N2). At the source, the synchrotron light is predominantly linearly polarized within the synchrotron electron orbit plane. Using reflective optics, the polarization is rotated to be vertical relative to the optical table and parallel to the tip axis before being focused by the parabolic mirror. As the parabolic mirror focuses the light onto the tip at slightly oblique angle and strong focus (numerical aperture ∼0.5), the incident light polarization at the tip position contains some in-plane components but remains predominantly polarized along the tip axis (out-of-plane relative to the sample surface). The tip and its near-field interaction with the sample further modify the relative importance of in-plane vs out-of-plane field components. , For separation of the near-field signal from far-field background, we demodulate the optical signal at higher harmonics nΩ of the mechanical tip oscillation frequency Ω. The near-field spectra used in the present publication are demodulated at n = 2, 3, 4. We disregarded spectra with n = 1 due to possibly insufficient background suppression, particularly at higher frequencies. We disregarded spectra with n ≥ 5 due to lower signal-to-noise ratio. All spectra have been normalized to reference spectra obtained on a commercial gold substrate (Platypus Technologies). Spectroscopy was performed at tapping amplitudes of 100 and 112 nm. The specific tapping amplitude was verified via measuring AFM retract curves and tracking the behavior of the mechanical amplitude as a function of tip–sample distance. Measurements were performed with two AFM tip types (nanoFTIR tip by Attocube, and 25PtIr200B–H by RockyMountainNanotechnology). After changing measurement parameters (AFM tip type, tapping amplitude), a full set of measurements was taken on the reference sample and all silicon samples. Measurements positions on the samples were chosen randomly. Assuming uniform samples, random measurement positions should not affect the measurement results. If local inhomogeneities exist, random measurement positions might contribute to an artificially increased standard deviation of our nanospectroscopy experiments. However, considering that the observed ω p difference between near- and far-field measurements well exceeds the standard deviation, it appears unlikely that random measurement positions could explain the consistently reduced ω p of the near-field measurements. Analytical Description and Fitting Procedure To extract Drude parameters from nano-FTIR spectra, we use the Python library Snompy in combination with a differential evolution algorithm. The dielectric function calculated from the Drude model was used as input to Snompy’s near-field signal simulation, and model parameters were fitted by minimizing the discrepancy between simulated and experimental signals. The dielectric function under the Drude model is expressed as ε(ω)=ε∞−ωp2ω2+iγω where ε ∞ is the high-frequency dielectric constant, ω p and γ denote the plasma frequency and scattering rate, respectively. Near-field signals were simulated using the Finite Dipole Model (FDM), where the AFM tip is modeled as a conducting ellipsoid forming a finite dipole under quasi-electrostatic approximation. The FDM considers only light components that are polarized along the tip axis, assumed vertical to the sample surface. While the FDM neglects in-plane polarization components, this approximation appears reasonable for our setup that predominantly employs linearly “p-polarized” incident light as well as due to the tip antenna effect. The appropriateness of the FDM is further supported by the good agreement between the analytical description and our experimental data. While surface layers may be included in advanced versions of the FDM, we employed the established FDM for bulk samples, thereby ignoring the effects of any potential native surface SiO2 layer. The good agreement between our calculated response and our experimental data suggests that the effect of a potential native SiO2 surface layer on the Si wafer may indeed be negligibly small for our current approach and that our approach using the bulk FDM is appropriate. This is further confirmed by the absence (within our experimental signal-to-noise) of clear spectral signatures corresponding to SiO2 phonon modes in the experimental spectra, which would be expected at around 500, 800, and 1100 cm-1. , In terms of quasistatic reflection coefficient of the sample β(ω)=ε(ω)−1ε(ω)+1 , the effective tip polarizability is given by , αeff=1+f0β2(1−f1β) with f 0 and f 1 determined by the tip geometry (length L, radius r), vibration amplitude A, minimum tip–sample distance z, and empirical parameter g representing the fraction of totally induced charge located within the spheroid. The effective polarizability is proportional to the near-field signal, and by demodulating αeff, we obtain different harmonic orders of the near-field signal, which matches the procedure in our experiment. As in the experiment, we utilize the second-, third-, and fourth-order demodulated data. Since light also reflects from the sample back onto the tip, we account for this far-field effect in the final near-field signal by including S(ω)=(1+Crrp)2αeff where r p defined as follow is the Fresnel’s reflection coefficient of p-polarized light of the sample as a function of ϵ­(ω) and the angle of incidence of the light θin, assuming light incident from air or vacuum. C r is an empirical scaling factor. rp=ϵ(ω)cos⁡(θin)−ϵ(ω)−sin2⁡(θin)ϵ(ω)cos (θin)+ϵ(ω)−sin2⁡(θin) Once the parameters mentioned above are specified, Snompy can compute the near-field signal across the desired frequency range. We note that the inclusion of the far-field reflection factor, ,, (1 + C r r p )2, in S(ω), implicitly assumes a sample that is locally homogeneous on the scale of the incident wavelength. While this is a useful theoretical description for the set of homogeneous samples that we are studying here, this step of the model would need to be adapted for a sample with features varying on a smaller length scale, e.g., by assuming a spatially averaged value for r p or by employing on-pixel normalization. Where experimentally feasible, a local reference measurement provides another alternative to reduce or eliminate the impact of the far-field reflection on the normalized signal. For fitting of our model parameters, we employ a differential evolution algorithm that is well-suited for identifying global optima, making it ideal for extracting parameters like ω p and γ from experimental data. To make fitting robust against outliers, we define a Huber loss function as follows to measure the discrepancy between experimental and simulated data. Lδ(ŷ)={12|y−ŷ|2(|y−ŷ|≤δ)δ(|y−ŷ|−12δ)(|y−ŷ|>δ) The optimal parameters are identified when this loss reaches a minimum. Since the setup has lower sensitivity and signal-to-noise at certain frequencies (close to the edges of the detected spectral range), those parts of the spectrum are less reliable. We therefore use a weight function that is calculated from a reference spectrum taken on gold to prevent the algorithm from overfitting these regions. w(ω)={r(ω)(r(ω)≥0.05)0(r(ω)<0.05),wherer(ω)=Sn(Au)(ω)max(Sn(Au)) The raw experimental spectra were first reconstructed using B-spline interpolation to ensure data continuity. To further increase the stability of our fitting algorithm against noise and outliers, we apply a Gaussian filter with a standard deviation around 20 cm–1 to the extracted signal. To compensate for potential temporal drift between the measurement of the sample and gold reference spectrum, we introduce a phase correction parameter in the simulated data, θcorrected(ω)≡θoriginal(ω)+cpc×ω Here c pc is the phase correction parameter and this is used to compensate for long-time phase drift. We emphasize that we fit the full complex near-field data (simultaneously fitting the real and imaginary part) as this enhances the quality of the fit, e.g., compared to only using the signal amplitude or fitting amplitude and phase separately. At the same time, we separately fit the complex spectra for each measurement and each demodulation order, which allows us to check our data for any significant dependencies on experimental parameters. For comparison, we also tested simultaneously fitting multiple harmonics (e.g., n = 2,3,4). However, for our data set, this did not noticeably improve results compared to our regular procedure of separately fitting the different demodulation orders and then averaging the results. Physical constraints were imposed to ensure meaningful results. Specifically, ε ∞ was fixed at 11.66 (silicon), and g = 0.7e 0.06i , consistent with literature and Snompy recommendations. The tip radius was set to 20 nm, and the FDM dipole length, treated as a fitting parameter, was limited up to 800 nm to satisfy the quasi-static condition. The reflectivity parameter C r was allowed to vary within reasonable small ranges due to unknown and potentially varying incidence angles: The experiment defines a range of relevant incidence angles via the shape and position of the parabolic mirror used to focus the light, but the FDM only uses a specific angle of incidence. While the angle of incidence should be within the range of possible angles of incidence of the experiment, it is not clear which angle of incidence will be most relevant and should, hence, be used for modeling. For clarity, we list all our parameters and their value range in Table below. 2Parameters of the Finite Dipole Model and Drude Model parameter name (symbol)value/value rangeplasma frequency (ω p )100–10,000 cm–1 scattering rate (γ)20–6000 cm–1 high-frequency dielectric constant (ε ∞ )11.66tip length (L)200–800 nmtip radius (r)20 nminduced charge parameter (g)0.7e 0.06i minimum tip–sample distance (z)0 nmtip tapping amplitude (A)taken from experimentincident angle (θ in )50–80 degreesreflectivity parameters ( C r )0.2–1.0phase offset correction parameter (c pc)–0.001–0.001aThe table shows the parameters of the finite dipole model and Drude model that we use for the analytical description of our nanospectroscopy data. For each parameter, the table either provides a specific value (for a fixed parameter) or a value range (for a fitting parameter). Parameters of the Finite Dipole Model and Drude Model The table shows the parameters of the finite dipole model and Drude model that we use for the analytical description of our nanospectroscopy data. For each parameter, the table either provides a specific value (for a fixed parameter) or a value range (for a fitting parameter). In the present work, each near-field spectrum is fitted independently, which is appropriate for the discrete measurements analyzed here. The current fitting framework is therefore designed for locally homogeneous spectra and does not impose spatial continuity constraints between different measurement points. Extending this approach to spatially resolved hyperspectral data sets will require a more advanced analysis framework, in which adjacent pixels are not fully independent, and some finite-dipole model parameters are either fixed globally or constrained to vary smoothly across space. As discussed above, simultaneously including multiple harmonics in the fitting procedure did not improve results for our homogeneous bulk samples, but this should be re-evaluated for locally varying samples, e.g., using on-pixel normalization using ratios of different harmonics or machine-learning approaches that may simultaneously consider different harmonics. While our experimental approach will be applicable to spatially varying samples without modification, extending our data analysis approach to cover quantitative analysis of such samples is beyond the scope of the present study. Nevertheless, the quantitative agreement achieved here for well-characterized bulk samples indicates that such extension is feasible in future work. Probing Depth of FTIR Reflectance Measurements and Depletion Layer Width at Si Surface The probing depth of far-field reflectance studies at near-normal incidence is given by the penetration depth , δ=λ4πIm(ϵ) . For the full spectral range, the probing depth of the far-field studies is in the order of a few 100 nm to >1 μm (Figure a). As expected, the probing depth is larger (smaller) for the lower (higher) doped samples as increasing the doping typically increases Im­(√ϵ). The penetration depth needs to be compared to the expected width of the depletion layer at the sample surface. In a one-dimensional abrupt junction formulation, the width of the depletion layer at a nonbiased Si-SiO2 interface (we assume that there will be a native SiO2 layer at the Si surface) can be described by the formula W=2ϵSienψs with the static dielectric constant of Si ϵSi ∼ 11.7 and the surface potential ψs=2ψB=2kTeln(nni) with Boltzmann constant k, temperature T, and intrinsic charge carrier concentration n i , which for Si is n i ≈ 1010 cm–3. Figure b shows the resulting calculated depletion width at the sample surface as a function of doping concentration. For the doping concentrations relevant for our samples of >1018 cm–3, the depletion width is in the order of a few 1 nm to a few 10 nm. A comparison with the calculated infrared probing depth for each sample (Figure a) shows that that the latter typically is about 2 orders of magnitude larger. This illustrates why far-field infrared reflectance spectroscopy is more sensitive to bulk sample properties and will be less sensitive to surface effects such as the surface depletion discussed here. However, for the near-field method, the probing depth is typically a few 10 nm up to around 100 nm, ,, which is much more similar to the surface depletion width. This explains why near-field nanospectroscopy, compared to conventional far-field spectroscopy, is expected to be much more sensitive to surface effects such as the depletion discussed here. Effects of the Finite Mean Free Path The mean free path length l = vτ can be calculated from the average carrier velocity v and the scattering time τ=2πcγ , with speed of light c. For sufficiently high doping, v is approximately given by the Fermi velocity v≈vF=ℏmeff(3π2n)1/3 . Using the experimental values provided in the manuscript, we calculate mean free paths of 2 to 5 nm for our samples. This is smaller than the typical probing depth of both the far- and near-field measurements, suggesting that the measurements are generally dominated by the bulk scattering rates and explaining why we observe largely consistent γ values between far- and near-field measurements. Rule of Thumb for Calculating Resistivity from Plasma Frequency and Scattering Rate We mention the simple rule-of-thumb ρFF[Ω·cm]=60γ̃[cm−1]ω̃p2[cm−1] for the resistivity ρFF in units of Ω·cm, which applies if one uses wavenumber units for the scattering rate γ̃ [cm –1] and plasma frequency ω̃ p 2 [cm–1]. This section explains how this rule can be derived from the Drude model. The rule only applies to simple semiconductors that can be described via the Drude model. In the Drude model, the conductivity σ = 1/ ρ is given as σ=ne2meffγ=ωp2ε0γ where for the second step we used the expression ωp2=ne2ε0meff for the plasma frequency. In the expression above, the scattering rate γ is commonly expressed via the scattering time τ = 1/γ. In this work we use wavenumber units, cm–1, for the plasma frequency and scattering rate. Wavenumbers can be derived from frequencies in Hz by multiplication with a factor 2πc with speed of light c. By assuming wavenumber units for the plasma frequency ω̃ p = 2πcω p and scattering rate γ̃ = 2πcγ, we find the expression: ρ=1σ=12πcε0γ̃ω̃p2≈60Ω·mγ̃[m−1]ω̃p2[m−1]=60Ω·cmγ̃[cm−1]ω̃p2[cm−1] . Calculations with the Layered Finite Dipole Model to Evaluate Impacts of a Surface Oxide Layer To evaluate the possible influence of inevitable surface oxidation on the FDM results, we applied the layered FDM (Q-ave in Snompy , ) to some experimental data from Samples 2 and 3, which previously showed the best agreement with far-field data. Since no clear spectral peaks of the SiO2 phonon modes were observed, suggesting a weak impact of these modes on the near-field response, the assumed oxide layer was treated as a dielectric layer with constant ϵ = 2.2. Fitting with 2 and 10 nm oxide thicknesses resulted in a reduction of ω p by ∼2% and ∼4%, and γ by ∼25% and ∼50%, respectively. This further reduction in ω p when including the oxide layer suggests that the original underestimation of ω p is not an artifact of neglecting oxidation. We note, however, that the layered model implementation itself is not as well established as the bulk finite dipole model and may require further exploring.