Deep learning assisted screening of light-responsive piezoelectric crystals for next-generation photoacoustic and photothermal spectroscopy

Abstract

Introduction High piezoelectric materials are promising alternative resonator materials for quartz-enhanced photoacoustic spectroscopy (QEPAS) and laser-induced thermoelastic spectroscopy (LITES), as enhanced piezoelectric response directly improves acoustic detection sensitivity. However, the discovery of materials combining high piezoelectricity with visible-light responsiveness remains a significant challenge. Methods We developed an improved crystal graph convolutional neural network (I-CGCNN) model for high-throughput screening of visible-light-responsive piezoelectric crystals. The model incorporates a specialized algorithm for learning bond feature vectors and synchronizes the updates of atomic and bond feature vectors during multilayer convolution. Using this model, we re-calculated the bandgap and piezoelectric moduli of all piezoelectric crystals in the Materials Project database. Results The I-CGCNN model significantly outperforms existing graph neural networks in predictive accuracy for both bandgap and piezoelectric modulus. The high-throughput screening identified several promising candidate crystals that exhibit both strong piezoelectric response and visible-light responsiveness. Discussion The screened candidates show considerable potential for piezo-phototronic applications, and the superior predictive performance of I-CGCNN suggests that it provides a powerful tool for accelerating the discovery of novel piezoelectric semiconductors for next-generation photoacoustic and photothermal spectroscopy.