Convolutional neural network prediction of the photocurrent–voltage curve directly from scanning electron microscopy images

Abstract

In the pursuit of efficient and sustainable energy conversion, high-performance photocatalytic devices show promise. A key characteristic of these devices is the photocurrent density vs. applied voltage (J–V) curve, providing crucial insights into their functionality. We demonstrate prediction of the J–V curve for BiVO₄ using a convolutional neural network (CNN) trained by scanning electron microscopy (SEM) images. Our methodology achieved a 98.9% curve match ratio. To optimize training, we varied magnification, SEM image types (backscattering electron and secondary electron images), and cut scale from a single SEM image. We built the model with a limited number of samples (28) by segmenting the original SEM image into smaller ones, totaling 840–26 656 data. We identified valuable structural features for predicting photocurrent using local interpretable model-agnostic explanation (LIME) activity images. This methodology can be extended to other photocatalytic materials, advancing our understanding of photocatalytic activity and facilitating the development of new materials and devices.