A double-layered ultraviolet flame sensor design for high-resolution spectral discrimination utilizing machine learning

Flame sensors are widely deployed in different fire monitoring scenarios. Identifying the origin of the flame, the fuel, often relies on the ultraviolet (UV) spectral signatures emitted by excited radicals during combustion. However, UV flame sensors capable of spectral discrimination typically require complex optical and material designs, resulting in high fabrication costs that hinder their application. To address this limitation, we propose a simulation-based multi-channel semiconductor UV flame sensor design based on an Al₂O₃–TiO₂ double-layered nanostructure. By tuning the geometric parameters of the nanostructure, a seven-channel sensor array was designed to exhibit distinct wavelength-dependent responses within the spectral band of 280–320 nm. Simulations showed that the proposed structure achieved high responsivities and high internal quantum efficiencies in the target band. On this basis, a convolutional neural network was trained to perform single-peak spectral discrimination, achieving a classification accuracy of 91% with a mean peak-position error of 0.32 nm. A U-Net-based network was further employed for dual-peak spectral reconstruction, yielding an average peak-intensity deviation ratio of 7.4% and high-fidelity recovery of the underlying spectra. These results suggest the feasibility of combining Al₂O₃–TiO₂ nanostructures with machine-learning-based spectral discrimination.