Yuki Kusano1,Hiroyuki Yoshida1
Chiba University1
Yuki Kusano1,Hiroyuki Yoshida1
Chiba University1
<b>Introduction</b><br/>The edges of the valence and conduction bands are essential for the operation of semiconductor devices. The energy of the conduction band edge with respect to the vacuum level is the electron affinity (EA). The precise measurement of EA became possible by the development of <u>low-energy inverse photoelectron spectroscopy (LEIPS)</u> in 2012 [1]. In LEIPS spectra analysis, the spectral onset is usually determined from the intersection of fitted lines to the baseline and the rising edge of the spectra. However, this method requires expertise and is time-consuming. As the LEIPS instruments are commercialized and widely used, quick and reliable analysis by non-experts is highly demanded. This study demonstrates an automated analysis method for LEIPS spectra based on machine learning.<br/><br/><b>Method</b><br/>So far, machine learning analysis of photoemission yield spectroscopy (PYS) spectra has been reported [2]; to determine the ionization energy, the spectral onset is determined using <u>Random Forest (RF)</u> and <u>Gradient Boosting (GB)</u> models. We also use these models in this work. Many machine learning studies on spectral analysis use simulated data because large numbers of data are required. However, there is no theoretical model for the spectral shape of LEIPS. Moreover, the simulations do not include the variations in spectral line shape owing to different film preparation methods or material-specific variations. Therefore, we utilized a dataset of 253 experimental LEIPS spectra of 28 organic semiconductors measured in our laboratory.<br/>The intensity for each energy was used as the explanatory variable, and the EA values determined by a skilled analyst (referred to as the analysis values) as the objective variable. The data set was divided into 80% training data set and 20% test data set. The model was trained on the training data and then validated on the test data. The predicted EA was output by inputting the LEIPS spectra into the trained model. The EA values predicted by each trained model were compared to the analytical values as a validation method.<br/><br/><b>Results and Discussion</b><br/>The results show that both RF and GB satisfactorily predict EA. For example, the differences between the analytical and the predicted values were only 0.04 eV and 0.00 eV for the C60 spectra, respectively. The differences between the analytical values and the predicted values (residuals) of the two models are summarized for all the test data: for RF and GB, 61% and 51% of the test data were predicted within ±0.1 eV, and 78% and 73% when the predicted values were extended to ±0.2 eV. The R<sup>2</sup> values were 0.915 and 0.856 for RF and GB, respectively. For some spectra, machine learning shows large deviation from the analytical values. We analyze the reason and propose the preprocess method to improve the precision.<br/><br/><b>References</b><br/>[1] H. Yoshida, <i>Chem. Phys. Lett.,</i> <b>539-540,</b> 180 (2012)<br/>[2] S. Yagyu, <i>Hyomen to Kagaku (Surface and Science) </i><b>62</b>, 504 (2019) in Japanese