Deep Learning-Enhanced Microwave Photonic Sensing with Inverse-Design Assisted Fabry-Pérot Cavity

We present an integrated Fabry-Pérot (FP) cavity-based sensor that leverages deep learning (DL)-assisted microwave photonic (MWP) interrogation to achieve high sensing accuracy, even when utilizing low quality factor (Q) resonator and in the presence of significant noise. The reflective FP cavity, constructed with inverse-designed inline reflectors, functions as a one-port sensing probe. By conducting MWP interrogation with multiple modulator bias voltages, even in the case where the interrogation bandwidth - is much smaller than the optical resonance bandwidth, shifts induced by environmental factors in a low-Q resonance are able to be transformed into distinct radio frequency transmission spectrum patterns, which are then fed into a DL model for feature extraction. After training, the DL model is capable of making accurate predictions of the target measurand, even in the presence of intense noise. As a proof of concept, we demonstrate temperature sensing using resonances with distinct phase transmission profiles and a Q factor of ∼2000, with an interrogation bandwidth of less than one-fifth of the resonance spectral notch width. By training a DL model with sensor outputs from six different modulator bias conditions, we achieve a temperature sensing accuracy of up to 0.01°C despite intense noise, representing up to 3.6-fold improvement over the use of a fixed modulation bias.