International Conference on Digital Innovations for Sustainable Solutions (ICDISS)
Hybrid Quantum Kernel Support Vector Regression for Bandgap Prediction in Halide Perovskites
Vaishnavi Bhardwaj
Abstract
Predicting the electronic bandgap ($E_{g}$) of hybrid organic-inorganic halide perovskites (HOIPs) remain a fundamental challenge in the design of high-efficiency solar cells and light-emitting diodes. While classical machine learning (ML) methods have achieved competitive accuracy, their reliance on fixed kernel spaces limits their ability to capture the intricate quantum correlations underlying$E_{g}$variations. To address this limitation, we propose a Hybrid Quantum Kernel Support Vector Regression (QKSVR) framework that integrates quantum feature encoding with classical regression. The model employs an optimized 3-qubit, single-layer quantum feature map constructed using RY rotations, T-gates for enhanced nonlinearity, and CZ entanglement for quantum correlation mapping. This design embeds compositional descriptors into a high-dimensional Hilbert space, enabling the model to learn electronic structure dependencies inaccessible to classical kernels. Feature selection was guided through the Kernel-Target Alignment (KTA) metric, a quantum-native criterion that identified the optimal descriptor subset (Cs, I,$t$) yielding a peak KTA of 0.8337. The resulting QKSVR achieved an outstanding predictive performance with$R^{2}=0.9806$and$\text{RMSE}=0.0618 \text{eV}$on the benchmark perovskite dataset (10.1007/s40033-023-00553-Z)[5]. Remarkably, this accuracy was achieved using only three chemical descriptors, underscoring the efficiency of quantum feature representations. The proposed KTA-driven hybrid modelling paradigm demonstrates the potential of quantum kernels in advancing materials informatics by revealing complex structure-property relationships beyond the reach of conventional ML models.