Multimodal learning enables instant ionizing radiation alerts on unmodified mobile phones for real-world emergency response

Mobile phone complementary metal–oxide–semiconductor (CMOS) image sensors exhibit intrinsic sensitivity to ionizing radiation and can serve as opportunistic detectors when dedicated instrumentation is unavailable. In this work, an ionizing radiation detection framework is developed using unmodified commercial mobile phones, in which sparse radiation-induced charge clusters are extracted from raw video streams through multimodal signal processing and deep learning. A hybrid 3D–2D spatio-temporal convolutional neural network is employed to identify transient radiation-caused events, and the resulting signal distribution is fused with brightness information to suppress visually induced false positives and to estimate dose rate. Monte Carlo simulations using Allpix 2 and controlled irradiation experiments were conducted to investigate radiation interaction behavior within the CMOS sensor and to validate the proposed detection mechanism. Experimental results demonstrate that hazardous dose rates of 0.25 mSv h −1 to 2.80 mSv h −1 can be identified within 6 s with an accuracy of 86%–96%, while extended measurements enable detection of low-level radiation down to approximately 6 µSv h −1 with an accuracy of 87%. The practical detection threshold of the proposed framework is substantially above natural background radiation levels ( < 0.2 µSv h −1 ). Therefore, the method is not intended for environmental radiation monitoring, but rather for opportunistic identification of unusually elevated radiation fields associated with lost radioactive sources, industrial incidents, or emergency response scenarios.