Prediction of the photoelectrochemical performance of hematite electrodes using analytical data

Abstract

Combination of analytical data could predict and specify the critical factors for the photoelectrode performance.

Identification of the Contributing Factors to the Photoelectric Conversion Efficiency for Hematite Photoanodes by Using Machine Learning

“…At first, we predicted the photocurrent values similarly as the method described in the previous paper. 20 Four analytical methods were used to extract feature values in the analytical data, as shown in Figures 1 and 2. Then the explanatory variables (Table X, with a size of 101 × 47) were prepared.…”

“…In this study, the performance of dozens of hematite samples was carefully examined, 20 and we noticed that about 50% of the samples did not or scarcely show any photocurrent. We classified these samples as "inactive samples" and sought to identify the primary reasons behind their inactivity.…”

“…3 In the previous work, we showed that hematite photoanodes prepared even under the same preparation condition had a large variation of the performance from 1.2 to 350% from their average photocurrent. 20 In the material development stage, researchers encounter this situation due to the lack of sufficient knowledge and experience, especially for those who are new to working on new materials. We need to have a good strategic resolution for these situations.…”

“…The analytical data for the samples are sophisticated descriptors to represent samples to predict the physical parameters even by using a small amount of data. 20,30 We prepared dozens of samples under the same preparation procedure, and they were analyzed with various analytical methods, such as X-ray diffraction (XRD), Raman spectroscopy, UV/vis absorption spectroscopy, and photoelectrochemical impedance spectroscopy (PEIS). The peaks/patterns in the analytical data were used as the descriptors of the samples, and they were used for the prediction of the photocurrent values representing their performance.…”

“…Therefore, the objective of this work was to investigate why half of the samples displayed extremely low photocurrent (<0.05 mA/cm 2 ), along with the prediction of the photocurrent values as was performed in the previous study. 20 We focused on the origin of inactive samples to identify dominant factors for "inactive". The ML models were constructed for the distinction of "active" or "inactive" (discrete values), instead the photocurrent values (continuous values) were predicted, and the dominant descriptors for determination of "inactive" samples were identified.…”