High-Throughput Screening of Piezoelectric Hybrid Organic–Inorganic Perovskites via Density-Functional Theory and Machine Learning

Abstract

Hybrid organic–inorganic perovskites (HOIPs) are promising candidates for piezoelectric devices owing to their structural tunability, low density, and intrinsic mechanical compliance. However, rational design principles linking molecular composition to orientation-independent piezoelectric performance remain underdeveloped, and the vast ABX₃ compositional space precludes exhaustive first-principles exploration. Here, we establish a “distribution-aware” density functional theory–machine learning (DFT–ML) workflow for scalable discovery of high-performance and flexible piezoelectric HOIPs. “Distribution-aware” refers to an explicit strategy to mitigate compositional distribution shift between the DFT training set and the much larger virtual screening library, by augmenting training with representative structures sampled from the candidate space. We first construct a uniform DFT dataset of 1,346 structures spanning 192 ABX₃ compositions and compute elastic tensors, piezoelectric tensors, band gaps, densities, and formation energies. To enable symmetry-independent screening, a rotation-invariant spectral descriptor, σ_max, defined as the largest singular value of the Kelvin-scaled piezoelectric strain tensor, is introduced. Feedforward neural networks trained on periodic structural descriptors achieve near-unity accuracy for thermodynamic and electronic properties, and moderate but quantitatively reasonable predictive performance for mechanical moduli and σ_max. To mitigate compositional distribution shift between the DFT dataset and the larger candidate space, we augmented the training set with representative structures. This strategy improves training diversity. An external validation subset was further reserved to explicitly evaluate model generalization in unseen compositional regions. The resulting ML models demonstrate robust transferability beyond the original DFT dataset. Leveraging the retrained models, we screened 2.7 million enumerated ABX₃ formulations, yielding 5,573 stable and geometrically optimized candidates. Multi-objective optimization identified two distinct material families: mechanically robust HXA-based mixed-halide frameworks for high-performance piezoelectric applications, and compliant EDA/MS-based systems for flexible device concepts. The workflow delivers experimentally testable targets and practical guidance for lead-free hybrid piezoelectrics. In addition, this workflow also facilitates the experimental exploration of the broader material systems and illuminates the structure–property relationships that govern piezoelectricity of HOIPs.