Automated Radiation Detection Using CsPbBr₃ Perovskite Sensors Using Machine Learning

Cesium lead bromide (CsPbBr₃ or CPB) is a promising all-inorganic perovskite for radiation detection; however, CsPbBr₃ single crystals exhibit unstable electrical behavior under low-dose irradiation. To address challenges in low-does detection, this study advances CsPbBr₃-based detection by evaluating two sensing architectures integrated with machine learning (ML): a direct electrical single-crystal detector and a strain-based CPB-polymethyl methacrylate (CPB-PMMA) coated Fiber Bragg Grating (FBG) sensor. This study enables a systematic comparison of the performance of both sensing modalities in low-dose regimes, which has not been previously reported. Gamma radiation experiments using low-dose Cs-137 and higher-dose Co-57 were performed, and signals were analyzed with ensemble-based ML models. The single crystal detector achieved 85.1% accuracy for Cs-137 and 95.8% for Co-57. Cross-source transfer analysis revealed asymmetric generalization behavior, with models trained on Cs-137 generalizing strongly to Co-57 (97.2% accuracy), whereas the reverse transfer performed more modestly (82.7% accuracy). Additionally, a multi-class model distinguished background and isotope classes with 85.5% accuracy. The CPB–PMMA FBG sensor demonstrated superior low-dose sensitivity, achieving 96.60% accuracy for Cs-137. These results demonstrate the feasibility of combining CsPbBr₃-based sensing with ML for automated radiation state monitoring and photonic strain-based sensing as a promising strategy for low-dose gamma radiation detection. • Evaluated the performance of CsPbBr₃ via direct electrical using single-crystal and strain-based FBG architectures • Applied ensemble machine learning combined with derivative feature engineering to enhance classification for low and high-dose radiation • CsPbBr₃ single crystal achieved 85.5% accuracy in multi-class identification using RUSBoosted ensemble models to distinguish specific isotopes from background signals • CPB-PMMA FBGs achieved 96.0% accuracy, overcoming inherent low-dose radiation instabilities in CsPbBr₃ single crystal • Demonstrated asymmetric transferability, where low-dose trained models generalize robustly to high-dose regimes