Reliable design for AlGaN nanoarray photocathode based on three-dimensional physical model accelerated by machine learning

Abstract

The aluminum gallium nitride (AlGaN) nanoarray is becoming increasingly significant as a photocathode configuration for use in electronic devices and ultraviolet sensors. Traditional photocathode design techniques rely on physical models that utilize empirical data and trial-and-error methods, making it difficult to optimize all parameters simultaneously. This research explores how machine learning (ML) can expedite the photocathode design process by creating a comprehensive database derived from a full-parameter photoemission model. This database includes various characteristic parameters (such as structural, material, and system parameters) alongside target parameters like photocurrent. We compare the effectiveness of several ML algorithms (including linear regression, random forest, support vector machine, and bi-layer/tri-layer fully connected neural network) using the data generated from the physical model. The findings indicate that ML can significantly enhance the design workflow for nanoarray photocathodes and assess the significance of each parameter. The neural network effectively maintains an error rate in performance predictions of around 4.7% while achieving a prediction speed of 13 000 observations per second. Optimal ranges for several parameters were identified, including nanowire diameter (243–347 nm), nanoarray height (278–412 nm), incident light angle (40–60°), Al component (0.3–0.5), and surface emission rate (103–105 cm/s). These outcomes imply that combining ML with physical modeling presents a fresh perspective for the design and conceptualization of nanoarray photocathodes.