Inverse design of high-performance piezoelectric semiconductors via advanced crystal representation and large language models

Abstract

The inverse design of solid-state materials with targeted properties represents a significant challenge in materials science, particularly for piezoelectric semiconductors where both structural symmetry and electronic properties must be carefully controlled. Here, we employ the simplified line-input crystal-encoding system representation combined with the MatterGPT framework for discovering potential piezoelectric semiconductors. By training on a curated dataset of 1556 piezoelectric materials from the Materials Project database, our model learns to generate crystal structures with targeted piezoelectric properties through an autoregressive sampling process. Starting from approximately 5000 generated structures, we implemented a comprehensive screening workflow incorporating structural validity, thermodynamic stability, and property verification. This approach identified several promising candidates from 4100 reconstructed structures, each representing compounds unrecorded in existing databases. Among these, the most notable material demonstrated a piezoelectric stress coefficient of 25.9 C/m2 in the e[1,6] direction. Additionally, these materials demonstrate suitable bandgaps ranging from 1.63 to 3.61 eV, suggesting potential applications in high-sensitivity sensors and high-temperature electronics. Our work demonstrates the effectiveness of combining crystal structure language encoding with generative models for accelerating the discovery of functional materials with targeted properties.