Energy level determination method and energy level determination system

Abstract

To provide an energy level determination method and an energy level determination system which can highly-accurately and easily determine an energy level of a solid from an electron spectroscopy spectrum or an optical absorption spectrum.SOLUTION: An energy level determination method determines a level of energy, such as electron affinity and ionization energy, of a solid from an electron spectroscopy spectrum or an optical absorption spectrum. The onset of the electron spectroscopy spectrum or the optical absorption spectrum is automatically determined by such a learned machine learning model that a measured value data set of the electron spectroscopy spectrum or the optical absorption spectrum has been learned.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present invention relates to an energy level determination method and system for determining energy levels such as the electron affinity and ionization energy of a solid from an electronic spectroscopy spectrum or a light absorption spectrum. [Background technology]

[0002] In organic semiconductor elements such as organic electroluminescence elements and organic solar cells, functions are expressed by the movement of holes in the energy level derived from the highest occupied molecular orbital (HOMO) and electrons in the energy level derived from the lowest occupied molecular orbital (LUMO). For this reason, the electron affinity (EA), which represents the LUMO level based on the vacuum level of the organic semiconductor used, and the ionization energy (IE), which represents the HOMO level, are very important parameters in the design and analysis of the operation of organic semiconductor elements.

[0003] As shown in Patent Document 1, for example, the EA of an organic semiconductor can be precisely determined from the rising edge of the LEIPS spectrum obtained by low-energy inverse photoelectron spectroscopy (LEIPS). In detail, in LEIPS, the EA can be determined based on the kinetic energy of the electrons incident on the organic semiconductor sample and the law of conservation of energy of the generated light, without damaging the sample, by narrowing down the electron energy to about 0 to 5 eV. It should be noted that LEIPS can also determine the EA of various solids such as metals and inorganic semiconductors in addition to organic semiconductors.

[0004] In addition, the IE can be precisely determined from the rise of the spectrum in ultraviolet photoelectron spectroscopy (UPS), photoelectron yield spectroscopy (PYS), etc. In this article, LEIPS, UPS, PYS, etc. are collectively referred to as electron spectroscopy. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent No. 6108361 (pages 4 to 6, Figure 1) Summary of the Invention [Problem to be solved by the invention]

[0006] In Patent Document 1, the EA of a solid in LEIPS is determined by a manual analysis of the LEIPS spectrum, specifically, the EA of the solid is determined from the intersection of the background in the LEIPS spectrum and a straight line fitted to the rising edge of the spectrum. However, this manual analysis of the LEIPS spectrum requires high skill and takes a long time.

[0007] The present invention has been made in view of the above-mentioned problems, and has an object to provide an energy level determination method and an energy level determination system that can determine the energy level of a solid with high accuracy and ease from an electron spectroscopy spectrum or an optical absorption spectrum. [Means for solving the problem]

[0008] In order to solve the above problems, the energy level determination method of the present invention comprises the steps of: A method for determining an energy level, such as an electron affinity or an ionization energy, of a solid from an electronic spectroscopy spectrum or a light absorption spectrum, comprising the steps of: The method is characterized in that the rise of the electron spectroscopy spectrum or the optical absorption spectrum is automatically determined by a trained machine learning model that has been trained on a data set of actual measurements of the electron spectroscopy spectrum or the optical absorption spectrum. According to this feature, by using a measured data set of electron spectroscopy spectra or optical absorption spectra for learning, it can be widely applied to determining the energy levels of various solids, and the energy levels of solids can be determined with high accuracy and ease from the electron spectroscopy spectra or optical absorption spectra using a trained machine learning model.

[0009] The measured value data set is characterized in that the energy range of the electron spectroscopy spectrum or the optical absorption spectrum is uniform. According to this feature, even if the energy range of the actual measured values ​​of the electron spectroscopy spectrum or optical absorption spectrum used to train the machine learning model is narrow, the prediction accuracy of the machine learning model can be improved by supplementing the data of the electron spectroscopy spectrum or optical absorption spectrum and aligning the energy range of the electron spectroscopy spectrum or optical absorption spectrum in the actual measured value dataset.

[0010] The electronic spectroscopy spectrum or the optical absorption spectrum is shifted in the direction of the kinetic energy axis and then duplicated, thereby increasing the number of data in the actual measurement data set. According to this feature, it is possible to increase the number of actual measurement data sets that are trained by the machine learning model without compromising the characteristics of the electron spectroscopy spectrum or optical absorption spectrum, thereby improving the prediction accuracy of the machine learning model.

[0011] The shift width of the electron spectroscopy spectrum or the optical absorption spectrum is characterized by being 0.001 to 1 eV, preferably 0.01 to 0.5 eV, and more preferably 0.01 to 0.1 eV. This feature makes it possible to highly efficiently improve the predictive accuracy of the machine learning model.

[0012] The measured value data set is characterized in that a smoothing process is performed on the electron spectroscopy spectrum or the optical absorption spectrum. According to this feature, noise can be reduced without compromising the characteristics of the electron spectroscopy spectrum or optical absorption spectrum, thereby improving the prediction accuracy of the machine learning model.

[0013] The smoothing process is characterized in that it is performed multiple times. According to this feature, noise can be sufficiently reduced without impairing the characteristics of the electronic spectroscopy spectrum or optical absorption spectrum, thereby improving the prediction accuracy of the machine learning model.

[0014] The electron affinity is automatically determined from the electronic spectroscopy spectrum or the optical absorption spectrum. According to this feature, the electron affinity, which requires a high level of skill in the analysis of electron spectroscopy or optical absorption spectrum, can be determined easily with high accuracy.

[0015] The measured values ​​of the electron spectroscopy spectrum are converted based on the vacuum level of the solid, The vacuum level is characterized by being determined by fitting the first derivative of the low energy electron transmission spectrum with a Gaussian function or a third order polynomial function. According to this feature, the vacuum level can be accurately determined from the low-energy electron transmission spectrum, so that the energy level of a solid can be determined easily and with high precision from the electron spectroscopy spectrum using a trained machine learning model.

[0016] The first derivative of the low-energy electron transmission spectrum is subjected to a Fourier transform. According to this feature, by performing a Fourier transform on the first derivative of the low-energy electron transmission spectrum, the number of peaks contained in the first derivative, in other words, the presence or absence of a shoulder structure appearing in the spectrum, can be easily and automatically detected. If the low-energy electron transmission spectrum has a shoulder structure, it suggests that there is a problem with sample charging or the like, and the electron spectroscopy spectrum has not been measured correctly. Therefore, by using data with only one peak, i.e., data without a shoulder structure, for determining the vacuum level, an accurate vacuum level can be determined.

[0017] The energy level determination system of the present invention comprises: An energy level determination system for determining an energy level such as an electron affinity or an ionization energy of a solid from an electronic spectroscopy spectrum or a light absorption spectrum, comprising: A trained machine learning model trained on a data set of actual measurements of the electron spectroscopy spectrum or the optical absorption spectrum, The method is characterized in that the onset of the electronic spectroscopy spectrum or the optical absorption spectrum is automatically determined by the trained machine learning model. According to this feature, by using a measured data set of electron spectroscopy spectra or optical absorption spectra for learning, the solids for which the energy levels are to be determined can be widely used, and the energy levels of solids can be determined easily and with high accuracy from the electron spectroscopy spectra or optical absorption spectra using a trained machine learning model. [Brief description of the drawings]

[0018] [Figure 1] FIG. 1 is a diagram showing the configuration of a measurement device using low energy inverse photoelectron spectroscopy (LEIPS) in an embodiment. [Diagram 2] FIG. 1 is a diagram showing an outline of a conventional manual analysis method for LEIPS spectra. [Diagram 3] FIG. 1 shows the molecular structures of some of the organic semiconductors used in the measured LEIPS spectrum data set used for machine learning in the examples. [Figure 4] FIG. 1 shows the energy range of the LEIPS spectra constituting the actual measurement data set in the examples. [Diagram 5] FIG. 1 is an explanatory diagram showing the results of automatic analysis (predicted values) of the electron affinity (EA) of F8 (polyfluorene) in Example 1. For comparison, predicted values ​​using Random Forest and Gradient Boosting as machine learning models for the LEIPS spectrum and values ​​analyzed by a manual analysis method are shown. [Figure 6] FIG. 11 is a diagram showing the distribution of the differences (residuals) between the predicted values ​​and the analyzed values ​​of the EA obtained by automatic analysis of the test data in Example 1. [Figure 7] FIG. 2(a) is an explanatory diagram showing the results of automatic analysis of EA of pentacene in Example 1, and FIG. 2(b) is an explanatory diagram showing the results of automatic analysis of EA of SnPc (tin phthalocyanine). [Figure 8] FIG. 11 is an explanatory diagram showing the results of automatic analysis of EA of SnPc in Example 2. The results of automatic analysis of EA by a trained machine learning model in which machine learning was performed using LEIPS spectra that were subjected to no smoothing process, smoothing process three times with 25 data points, and smoothing process five times with 81 data points as training data are shown from the top. [Figure 9] This is a diagram showing the distribution of the difference between the predicted value and the analyzed value of EA obtained by the automatic analysis of the test data in Example 2. The automatic analysis results of EA for LEIPS spectra obtained from the top with no smoothing process, with smoothing process performed three times with 25 data points, and with smoothing process performed five times with 81 data points are shown. [Figure 10] FIG. 13 is a diagram illustrating a shift process in which training data of an actual measurement value data set is shifted and new data is replicated in the third embodiment. [Figure 11] This is a diagram showing the distribution of the difference between the predicted value and the analyzed value of EA obtained by automatic analysis of test data in Example 3. Note that from the top, the automatic analysis result of EA by a trained machine learning model in which machine learning was performed using LEIPS spectra that were smoothed three times with 25 data points as training data, and the second to fifth are the automatic analysis results of EA by a trained machine learning model in which machine learning was performed by increasing the number of data for training data by performing smoothing three times with 25 data points and further performing shift processing with shift widths of 0.01 eV, 0.05 eV, 0.10 eV, and 0.15 eV. [Figure 12]FIG. 1A shows the low energy electron transmission (LEET) spectrum and its first derivative of the PCBM thin film in Example 4; FIG. 1B shows the first derivative of FIG. 1A fitted with a Gaussian function; FIG. 1C shows the first derivative of FIG. 1A fitted with a third-order polynomial function; FIG. 1D shows the distribution of the difference between the fitted values ​​automatically analyzed using a Gaussian function for 50 measurements and the analytical values ​​manually determined by an expert; and FIG. 1E shows the distribution of the difference between the fitted values ​​automatically analyzed using a third-order polynomial function for 50 measurements and the analytical values ​​manually determined by an expert. [Figure 13] FIG. 1A is a diagram showing the LEET spectrum and its first derivative (without shoulder structure) of the organic semiconductor thin film in Example 5, FIG. 1B is a diagram showing the LEET spectrum and its first derivative (with shoulder structure) of the organic semiconductor thin film, FIG. 1C is a diagram showing the result of Fourier transform of the first derivative (without shoulder structure) of FIG. 1A, and FIG. 1D is a diagram showing the result of Fourier transform of the first derivative (with shoulder structure) of FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] Hereinafter, an embodiment of the present invention will be described. However, the present invention can be embodied in many different forms, and is not limited to the following embodiments and examples.

[0020] The energy level determination method and energy level determination system of the present embodiment determine the energy level of a solid from an electronic spectroscopy spectrum or an optical absorption spectrum measured by various measurement methods, and automatically determine the rise of the electronic spectroscopy spectrum or the optical absorption spectrum by a trained machine learning model that has been trained on a data set of actual measured values ​​of the electronic spectroscopy spectrum or the optical absorption spectrum.

[0021] The energy level determination method and the energy level determination system of the present embodiment can determine the energy level of any solid, whether it is a metal, a semiconductor, or an insulator, from the rise of the electronic spectroscopy spectrum or the optical absorption spectrum, as long as the solid can measure the electronic spectroscopy spectrum or the optical absorption spectrum.

[0022] In this embodiment, the electron spectroscopy spectrum may be an electron spectroscopy spectrum measured by various electron spectroscopy methods, for example, photoelectron spectroscopy (PES), photoelectron yield spectroscopy (PYS), inverse photoelectron spectroscopy (IPES), low energy inverse photoelectron spectroscopy (LEIPS), X-ray photoelectron spectroscopy (XPS), ultraviolet photoelectron spectroscopy (UPS), angle resolved electron spectroscopy (ARPES), Auger electron spectroscopy (AES), soft X-ray emission spectroscopy (SXES), etc. In addition, the optical absorption spectrum may be measured by, for example, ultraviolet-visible spectroscopy (UV-VIS), X-ray absorption spectroscopy (XAS), etc.

[0023] The energy level determination method and energy level determination system of the present embodiment can automatically determine the ionization energy (IE) as an energy level from the rising edge of the electronic spectroscopy spectrum or light absorption spectrum of a solid measured by, for example, PES or PYS, and can automatically determine the electron affinity (EA) as an energy level from the rising edge of the electronic spectroscopy spectrum of a solid measured by IPES or LEIPS. Here, the "rising edge" of the electronic spectroscopy spectrum refers to the rising portion of the intensity relative to the electron energy in, for example, PES, IPES, and LEIPS, and refers to the rising portion of the intensity relative to the light energy in PYS.

[0024] In the energy level determination method and the energy level determination system of the present embodiment, machine learning is performed so that the energy level can be automatically determined from the rising edge shape of the electron spectroscopy spectrum or the optical absorption spectrum.

[0025] In this embodiment, the machine learning is supervised learning, and the measured values ​​of the electronic spectroscopy spectrum or the optical absorption spectrum are input as input data, and the energy levels determined from the rise of the electronic spectroscopy spectrum by a manual analysis method (see FIG. 2) are output as output data, which are given to a machine learning algorithm and calculated as correct answer data. Various machine learning models may be applied to the machine learning, but for example, by applying Random Forest or Gradient Boosting using an ensemble method that constructs a more powerful model by combining multiple decision trees (regression trees), which are learning models with low accuracy, it is possible to improve the generalization ability, i.e., the prediction ability for unlearned data, and suppress overlearning.

[0026] In this embodiment, the actual measurement data set used for machine learning is created from data obtained by analyzing actual values ​​of electron spectroscopy spectra or optical absorption spectra measured by various measurement methods using a manual analysis method. In detail, the actual measurement data set includes annotations that add metadata such as the type of measurement method, the type of solid, and the analysis value of the energy level determined by the manual analysis method, i.e., ground truth data, to the actual measurement value of the electron spectroscopy spectrum.

[0027] In addition, since the tendency of the spectrum shape of the electron spectroscopy spectrum or the optical absorption spectrum differs depending on the type of solid measured by various measurement methods, the accuracy of the automatic determination of the energy level can be improved by preparing actual measurement value data sets of metals, semiconductors, and insulators and using them for machine learning. In addition, for example, in semiconductors, the tendency of the spectrum shape differs between n-type semiconductors and p-type semiconductors, so the accuracy of the automatic determination of the energy level can be improved by preparing a more finely divided actual measurement value data set and using it for machine learning.

[0028] In this embodiment, as a pre-processing of the actual measurement data set before machine learning, it is preferable that the electron spectroscopy spectrum or the optical absorption spectrum is converted according to the determined energy level, etc. For example, when determining the electron affinity or ionization energy, the electron spectroscopy spectrum is converted based on the vacuum level. The electron spectroscopy spectrum may be converted based on the Fermi level, or may be converted so as to move the rising edge of the spectrum to the center of the energy range.

[0029] In this embodiment, as a pre-processing of the actual measurement data set before machine learning, for actual measurement data in which the energy range of the actual measurement value of the electron spectroscopy spectrum or the optical absorption spectrum is narrower than a predetermined range, both ends of the spectrum data are supplemented to make it fit the predetermined energy range. Note that for actual measurement data in which the energy range is wider than the predetermined range, the pre-processing is not performed, and unnecessary parts are cut off. As a method of supplementing the spectrum data, a method of supplementing both ends of the electron spectroscopy spectrum or the optical absorption spectrum by making them continuous (see FIG. 4), a method of supplementing a straight line fitting the slope near both ends of the electron spectroscopy spectrum or the optical absorption spectrum, or the like may be applied. Note that the appropriate energy range of the electron spectroscopy spectrum or the optical absorption spectrum differs depending on the type of measurement method, the type of solid, and the like.

[0030] In this embodiment, as pre-processing of the actual measurement data set before machine learning, for example, a smoothing process is performed by the Savitzky-Golay method, etc. Note that the smoothing process can be performed multiple times to sufficiently reduce noise without impairing the characteristics of the electron spectroscopy spectrum or the optical absorption spectrum.

[0031] In this embodiment, in order to increase the number of data in the actual measurement data set, a shift process is performed in which the actual measurement values ​​of the electron spectroscopy spectrum or the optical absorption spectrum are shifted by a predetermined shift width in the direction of the electron kinetic energy axis and duplicated. Note that the electron spectroscopy spectrum or the optical absorption spectrum to be shifted may be actual measurement data that has not been subjected to smoothing processing, or may be actual measurement data that has been subjected to smoothing processing.

[0032] In addition, the appropriate shift width in the shift process is 0.001 to 1 eV, preferably 0.01 to 0.5 eV, more preferably 0.01 to 0.1 eV, to improve the accuracy of automatic determination of the energy level. In addition, in the case of LEIPS, the shift width is preferably 0.05 eV to improve the accuracy of automatic determination of the electron affinity.

[0033] In addition, in the shift process, the number of copies to be made may be appropriately selected based on the relationship between the number of data in the actual measurement data set and the number of data sets required for machine learning. However, from the viewpoint of improving the accuracy of the automatic determination of the energy level, it is preferable to make copies in both the positive and negative directions of the reference spectrum. EXAMPLES

[0034] Here, the energy level determination system according to the embodiment was actually manufactured, and the energy level of a solid was determined. Specific explanation will be given below.

[0035] In this embodiment, the actual measured value of the LEIPS spectrum of an organic semiconductor measured by the low energy inverse photoelectron spectroscopy (LEIPS) measuring device shown in Figure 1 is input into the energy level determination system, and the electron affinity (EA) is automatically determined by the trained machine learning model.

[0036] In this embodiment, two machine learning models are used: Random Forest (RF) and Gradient Boosting (GB). The variable predicted by machine learning (objective variable) is the analyzed EA value, and the variable acting on the objective variable (explanatory variable) is the intensity (LEIPS spectrum) for each energy based on the vacuum level.

[0037] As shown in Figure 2, the manual analysis of the LEIPS spectrum determines the EA from the intersection of the background of the LEIPS spectrum and the spectrum near the onset, which are fitted with a straight line (see the dashed line in Figure 2). In this example, the EA obtained by the manual analysis by an expert in LEIPS is used as the analytical value.

[0038] The measured data set for machine learning in this embodiment uses 208 measured LEIPS spectra data for 21 types of organic semiconductors (see FIG. 3), including PCBM, C60, and P3HT, which are known materials for organic solar cells, as well as non-fullerene acceptors such as ITIC and IT-M. In addition, metadata such as the type of organic semiconductor and the analytical value of EA, which is the correct answer data, are added to the measured LEIPS spectra data by annotation.

[0039] In addition, as a pre-processing of the measured data set before machine learning, the LEIPS spectrum was converted to the vacuum level standard, and in order to align the energy range of the LEIPS spectrum to a predetermined range, the values ​​at both ends of the LEIPS spectrum were supplemented by making them continuous for the data of 208 LEIPS spectra (see the dotted line in FIG. 4). In this example, the energy range was set to 1.01 to 7.41 eV as the predetermined range.

[0040] After this preprocessing, 80% of the actual measurement data set is divided into training data (teacher data) and 20% is divided into test data, and machine learning is performed.

[0041] Figure 5 shows an example of the results of automatic analysis when test data was input to two trained machine learning models (RF, GB) that had undergone machine learning using a measured value dataset (training data).

[0042] Figure 5 shows the results of automatic analysis of organic semiconductor F8 (polyfluorene, see Figure 3), showing the LEIPS spectrum and the analytical value of EA (see solid arrow), which is the correct data, the predicted value of EA by RF (see two-dot chain arrow), and the predicted value of EA by GB (see dotted arrow). For organic semiconductor F8, the difference (residual) between the predicted value and the analytical value was 0.04 eV for RF and 0 eV for GB, confirming that EA can be automatically analyzed with high accuracy.

[0043] In addition, as shown in FIG. 6, from the distribution of the difference between the predicted value and the analytical value for all the test data, the residuals were concentrated near 0 in both the RF and GB machine learning models, and it was confirmed that EA can be automatically analyzed with high accuracy. In RF, the prediction accuracy was 71% of the test data within the residual of ±0.1 eV, and 88% of the test data within the residual of ±0.2 eV, while in GB, the prediction accuracy was 76% of the test data within the residual of ±0.1 eV, and 93% of the test data within the residual of ±0.2 eV. From this, the difference between the predicted value and the analytical value was slightly better in GB, but no significant difference due to the machine learning model was confirmed. In this embodiment, R 2 The correlation coefficient is R 2 : 0.924, R at RF 2 :0.964, which is highly accurate.

[0044] Here, the results of automatic analysis for pentacene and tin phthalocyanine (SnPc), for which the residuals were large in both RF and GB in Figure 6, are shown in Figure 7. In Figure 7, in order to facilitate comparison with the analytical values, straight lines fitted to the rising edges of the spectra corresponding to the predicted values ​​for RF and GB are shown.

[0045] As shown in Figure 7(a), in the case of pentacene, the signal-to-noise ratio is poor, so it is presumed that both the RF and GB machine learning models predicted the noise as the onset. In addition, as shown in Figure 7(b), in the case of tin phthalocyanine, in addition to the signal-to-noise ratio being poor, the onset changes slightly along the way, so it is presumed that a different point was predicted as the onset. From these findings, it was confirmed that when there was a large difference between the predicted value and the analytical value, there was a problem with the signal-to-noise ratio and shape of the LEIPS spectrum. EXAMPLES

[0046] In this embodiment, in order to solve the above-mentioned problems of the SN ratio and shape of the LEIPS spectrum, a smoothing process is performed on the LEIPS spectrum constituting the actual measurement data set. Note that the same configuration as in the above-mentioned embodiment 1 and the description of the overlapping contents will be omitted.

[0047] The actual measurement data set for machine learning in this embodiment uses actual measurement data of 253 LEIPS spectra for 28 types of organic semiconductors (see FIG. 3), which is more than in Example 1. In addition to the 21 types of organic semiconductors in Example 1, seven types of organic semiconductors, mainly those with low analytical values ​​for EA (around 2), such as BCP and organic EL materials, are added.

[0048] In this embodiment, the Savitzky-Golay method is used for the smoothing process, thereby preserving the characteristics of the LEIPS spectrum shape.

[0049] As shown in Figure 8, the automated analysis results for tin phthalocyanine (SnPc) predicted noise as a rising edge in both the RF and GB machine learning models when smoothing processing was not performed. However, when smoothing processing was performed three times with 25 data points, the difference between the predicted value and the analytical value was 0.03 eV for RF and 0.11 eV for GB. When smoothing processing was performed five times with 81 data points, the difference between the predicted value and the analytical value was 0 eV for RF and 0.01 eV for GB, confirming that the prediction accuracy was improved.

[0050] As shown in Figure 9, from the distribution of the difference between the predicted values ​​and the analysis values ​​for all the test data, the residuals are concentrated near 0 for both the RF and GB machine learning models, but when smoothing processing is not performed, the distribution is broader. However, by increasing the number of data points for smoothing processing, the amount of data with residuals near 0 increases, narrowing the distribution. It can also be confirmed from the standard deviation value that the prediction accuracy is improved by smoothing processing. EXAMPLES

[0051] In this embodiment, since machine learning essentially requires 1000 or more pieces of data, a shift process is performed to increase the actual measurement data set. Note that the same configuration as in the first and second embodiments will not be described here.

[0052] As shown in Fig. 10, the shifting process of this embodiment copies the LEIPS spectrum of the training data (25 data points that have been smoothed three times) while shifting it in the kinetic energy axis direction by a predetermined shift width to generate new data. In detail, the shift width is set to 0.01 eV, 0.05 eV, 0.10 eV, and 0.15 eV, respectively, and 10 LEIPS spectra are generated in the negative direction and 10 in the positive direction of the reference spectrum for each LEIPS spectrum.

[0053] In this embodiment, in order to determine whether a specific LEIPS spectrum has been overtrained using test data, only the training data is shifted. That is, the training data is increased from 202 to 4242 while leaving the 51 pieces of test data as they are.

[0054] As shown in Figure 11, from the distribution of the differences between the predicted values ​​and the analytical values ​​for all of the test data, it was confirmed that, compared to the case where smoothing processing was performed three times on 25 data points, the residuals were concentrated near 0 for both the RF and GB machine learning models, and the residuals were mostly distributed within ±0.2 eV. In particular, it was confirmed that the standard deviation value was smaller by performing shift processing with a shift width of 0.01 eV, thereby improving the prediction accuracy. EXAMPLES

[0055] In this embodiment, a method for accurately determining the vacuum level by automatic analysis (onset analysis) of the low energy electron transmission (LEET) spectrum obtained in the process of LEIPS measurement of an organic semiconductor is described in order to convert the LEIPS spectrum to a vacuum level standard as a pre-processing of the actual measurement data set before machine learning. Note that the description of the same configuration as in the above-mentioned embodiments 1 to 3 will be omitted.

[0056] In this embodiment, the vacuum level of the organic semiconductor is determined by using a machine learning regression model from the LEET spectrum. Since the LEET spectrum is obtained during the LEIPS measurement, the vacuum level of the organic semiconductor can be determined in this embodiment under the same conditions as the LEIPS measurement without adding any additional measurement equipment. In the actual measurement, the LEET spectrum is measured as the sample electron current I = I (V), which is a function of the electron acceleration voltage V (see, for example, FIG. 12(a)).

[0057] In this embodiment, in order to determine the vacuum level of an organic semiconductor from a LEET spectrum, first, the first derivative dI / dV is calculated from the measured LEET spectrum I(V) by a method such as the five-point Savitzky-Golay method. Next, the maximum value of the first derivative dI / dV is searched for, and the top 60% of the data points are assumed to be the peak region, and a Gaussian function or a third-order polynomial function that represents the peak shape is fitted. The extreme value (i.e., the maximum value) of the fitted function is the onset of LEET. The vacuum level is obtained by adding the detected light energy of the light detection used in the LEIPS measurement to the energy of the onset.

[0058] Here, a method for determining the vacuum level of an organic semiconductor will be described with a specific example. In FIG. 12(a), the dashed line indicates the LEET spectrum of the PCBM thin film, which is an organic semiconductor, that is, the sample electron current I. In this embodiment, the curve of the first derivative dI / dV (solid line) is obtained by calculating the first derivative dI / dV from the LEET spectrum. Also, as shown by the arrow in FIG. 12(a), the maximum value of the first derivative dI / dV is set as the onset.

[0059] Next, assuming that the top 60% of the data points in the first derivative dI / dV are the peak region, fitting is performed using a Gaussian function or a third-order polynomial function. The Gaussian function shown by the dashed line in Figure 12(b) roughly fits the top 60% of the asymmetric peak of the curve of the first derivative dI / dV shown by the solid line, but it was confirmed that the fitting value (1.90 eV) automatically analyzed using the Gaussian function is 0.02 eV lower than the analysis value (1.92 eV) manually determined by an expert shown by the white arrow, because the peak shape of the curve of the first derivative dI / dV is asymmetric.

[0060] In order to solve the problem of fitting to the curve of the first derivative dI / dV with such an asymmetric peak shape, fitting using a third-order polynomial function was performed. As a result, it was confirmed that the third-order polynomial function shown by the dashed line in Figure 12(c) accurately fits the asymmetric peak shape of the curve of the first derivative dI / dV shown by the solid line, and the fitting value (1.92 eV) automatically analyzed using the third-order polynomial function matches the analytical value (1.92 eV) manually determined by an expert, as shown by the white arrow.

[0061] In addition, as shown in Figures 12(d) and (e), a comparison was made between the distribution of differences between the analytical values ​​manually determined from the results of 50 spectral measurements and the fitting values ​​automatically analyzed using a Gaussian function or a third-order polynomial function. As a result, it was confirmed that by fitting the curve of the first derivative dI / dV having an asymmetric peak shape with a third-order polynomial function, the average deviation μ changed from -0.017 eV (Gaussian function) to -0.001 eV (third-order polynomial function) and the standard deviation σ changed from 0.016 eV (Gaussian function) to 0.011 eV (third-order polynomial function), thereby improving the fitting accuracy.

[0062] In this way, when the curve of the first derivative dI / dV has an asymmetric peak shape, the fitting accuracy can be improved by adopting fitting with a third-order polynomial function. Also, when the curve of the first derivative dI / dV has a symmetric peak shape, sufficient fitting accuracy can be obtained even if fitting with a Gaussian function is adopted. For these reasons, it is preferable that either the Gaussian function or the third-order polynomial function can be selected for fitting depending on the peak shape of the curve of the first derivative dI / dV. EXAMPLES

[0063] In this embodiment, the vacuum level of an organic semiconductor is determined by performing a Fourier transform on the curve of the first derivative dI / dV of the LEET spectrum to automatically detect the number of peaks and select data to be used for determining the vacuum level. Note that the description of the same configuration as in the fourth embodiment will be omitted.

[0064] Figure 13(a) shows the LEET spectrum of an organic semiconductor thin film (dash-dotted line), and Figure 13(b) shows the LEET spectrum of an organic semiconductor thin film (dash-dotted line). Here, the curve (solid line) of the first derivative dI / dV of the LEET spectrum of the organic semiconductor thin film shown in Figure 13(b) shows two peaks (see solid line arrows) that are presumed to be caused by the charging of the organic semiconductor, and a so-called shoulder structure can be confirmed. It is difficult to determine the position of the peak in the first derivative dI / dV of the LEET spectrum with such a shoulder structure, and it may not be possible to accurately determine the vacuum level by simply fitting with a Gaussian function or a third-order polynomial function. The white arrows in Figure 13(b) are analytical values ​​determined manually by experts.

[0065] Therefore, in order to accurately determine the vacuum level, a Fourier transform is performed on the curve of the first derivative dI / dV of the LEET spectrum. It was confirmed that by performing a Fourier transform on the curve of the first derivative dI / dV of the LEET spectrum without a shoulder structure (see FIG. 13(a)), one clear peak (see the solid arrow in FIG. 13(c)) is automatically detected, and by performing a Fourier transform on the curve of the first derivative dI / dV of the LEET spectrum with a shoulder structure (see FIG. 13(b)), multiple (four in this embodiment) clear peaks (see the solid arrows in FIG. 13(d)) are automatically detected.

[0066] As described above, in order to convert the LEIPS spectrum into the vacuum level standard as a pre-processing of the measured data set before machine learning, the number of peaks included in the first derivative, in other words, the presence or absence of a shoulder structure appearing in the spectrum, can be easily and automatically detected by performing a Fourier transform on the curve of the first derivative dI / dV of the LEET spectrum. If the LEET spectrum has a shoulder structure, it suggests that there is a problem with sample charging or the like and that the LEIPS spectrum has not been measured correctly. Therefore, the data is rejected and data with one peak, i.e., data without a shoulder structure, is used to determine the vacuum level, and the vacuum level can be accurately determined by the fitting value automatically analyzed by a Gaussian function or a third-order polynomial function.

[0067] In other words, by combining the automatic detection of the number of peaks by Fourier transform of the curve of the first derivative dI / dV of the LEET spectrum in this embodiment, the automatic analysis (onset analysis) of the LEET spectrum described in Example 4, and the automatic analysis of the onset of the LEIPS spectrum described in Examples 1 to 3, the electron affinity can be completely automatically determined.

[0068] In addition, in Examples 1 to 5, the embodiment of automatically analyzing the electron affinity (EA) from the LEIPS spectrum of an organic semiconductor has been described, but it goes without saying that the present invention is applicable to automatically determining the energy levels of the electron spectroscopy spectrum measured by various electron spectroscopy methods for various solids. In addition, since the characteristics of the electron spectroscopy spectrum differ depending on the type of solid, for example, the energy level determination system may be provided with a plurality of trained machine learning models that have been trained with actual measurement data sets for each type of solid, and the machine learning model may be switched on the system to improve the prediction accuracy of the energy levels for various solids. In this case, the machine learning model may be manually switched by an analyst, or a machine learning model that has been trained to classify solids according to the characteristics of the electron spectroscopy spectrum may be separately provided, and a machine learning model that determines the energy level of the classified solid may be automatically selected.

[0069] As described above, by using a measured data set of electron spectroscopy spectra for learning, it is possible to widely apply the method to determine the energy levels of various solids, and the energy levels of solids can be determined easily and with high accuracy from the electron spectroscopy spectra using a trained machine learning model. It goes without saying that the energy levels of solids can also be determined easily and with high accuracy from the optical absorption spectra in a similar manner.

[0070] In addition, in LEIPS and IPES, the intensity of the spectra measured is weak in principle, and the spectrum is easily affected by noise components. However, smoothing processing can reduce noise without compromising the characteristics of the LEIPS spectrum, and is therefore particularly effective as a preprocessing method for measured data sets.

[0071] Although the embodiments of the present invention have been described above with reference to the drawings, the specific configuration is not limited to these embodiments, and the present invention also includes modifications and additions that do not deviate from the gist of the present invention. [Industrial Applicability]

[0072] The present invention can be widely applied to the determination of the energy levels of various solids by using a measured data set of electronic spectroscopy spectrum or optical absorption spectrum for learning, and the energy levels of solids can be determined easily and with high accuracy from electronic spectroscopy spectrum or optical absorption spectrum by a trained machine learning model, so that it is highly versatile and can perform a large amount of spectrum analysis in a short time without depending on the skill of the analyst, and has industrial applicability. In addition, by performing signal processing (smoothing processing) on ​​the measured data of electronic spectroscopy spectrum or optical absorption spectrum, it is possible to improve the accuracy of prediction for spectra with poor S / N ratios, and by performing duplication processing (shift processing) of training data, it is possible to improve the accuracy of prediction even for those with a small number of measured values, and the application is wide. In addition, by combining automatic detection of the number of peaks by Fourier transform for the curve of the first derivative dI / dV of the LEET spectrum obtained in the process of LEIPS measurement of a solid, automatic analysis of the LEET spectrum (onset analysis), and automatic analysis of the LEIPS spectrum, it is possible to realize fully automatic determination of the electron affinity.

Claims

1. A method for determining an energy level, such as an electron affinity or an ionization energy, of a solid from an electronic spectroscopy spectrum or a light absorption spectrum, comprising the steps of: A method for determining energy levels, comprising automatically determining a rise of the electronic spectroscopy spectrum or the optical absorption spectrum by a trained machine learning model trained on a data set of actual measurements of the electronic spectroscopy spectrum or the optical absorption spectrum.

2. 2. The method of claim 1, wherein the measured value data set includes a set of energy ranges of the electron spectroscopy spectrum or the optical absorption spectrum.

3. 2. The method for determining energy levels according to claim 1, wherein the number of data in the actual measurement data set is increased by shifting the electronic spectroscopy spectrum or the optical absorption spectrum in a kinetic energy axis direction and duplicating the spectrum.

4. 4. The method for determining an energy level according to claim 3, wherein a shift width of the electronic spectroscopy spectrum or the optical absorption spectrum is 0.01 to 0.5 eV.

5. 2. The method of claim 1, wherein the measured data set is obtained by performing a smoothing process on the electron spectroscopy spectrum or the optical absorption spectrum.

6. 6. The method of claim 5, wherein the smoothing process is performed a plurality of times.

7. 7. The method for determining an energy level according to claim 1, wherein an electron affinity is automatically determined from the electronic spectroscopy spectrum or the optical absorption spectrum.

8. The measured values ​​of the electron spectroscopy spectrum are converted based on the vacuum level of the solid, 2. The method of claim 1, wherein the vacuum level is determined by fitting the first derivative of the low energy electron transmission spectrum with a Gaussian function or a third order polynomial function.

9. 9. The method of claim 8, further comprising the step of Fourier transforming the first derivative of the low energy electron transmission spectrum.

10. An energy level determination system for determining an energy level such as an electron affinity or an ionization energy of a solid from an electronic spectroscopy spectrum or a light absorption spectrum, comprising: A trained machine learning model trained on a data set of actual measurements of the electron spectroscopy spectrum or the optical absorption spectrum, An energy level determination system, characterized in that the onset of the electronic spectroscopy spectrum or the optical absorption spectrum is automatically determined by the trained machine learning model.

Images