Attention-assisted deep neural network for nonlinear equalization and denoising in time-stretched photonic ADC

We present a photonic time-stretch (PTS) front-end architecture whose signal recovery methodology is agnostic to both sampling rate and input signal power, thereby eliminating the need for repeated model retraining. This is implemented using a deep neural architecture that seamlessly integrates an attention-enhanced long short-term memory (LSTM) recurrent neural network and a feedforward neural network (FFNN) for suppression of noise and nonlinear distortions induced by the front-end. The two neural network models function in parallel: the attention-assisted LSTM predicts the normalized signal waveform, while the FFNN estimates the root-mean-square (RMS) signal power. The recovered waveform and power are combined for accurate reconstruction of the original signal. Unlike the previous photonic analog-to-digital converter (ADC) system implementations, the proposed LSTM-driven framework is capable of processing input signals across arbitrary sampling rates and sample sizes without the need for model retraining. Further, the attention mechanism enables effective modeling of long-term temporal dependencies by prioritizing the more relevant time steps, ensuring accurate signal recovery even for longer input sequences without introducing any signal attenuation. The proposed system achieves a signal-to-noise-and-distortion ratio (SNDR) of 63 dB and an effective number of bits (ENOB) of 10.18 using a 1 GSa/s electronic ADC for input RF frequencies in the range of 0.6 GHz to 1.4 GHz. The proposed PTS front-end reduces the ADC bandwidth requirement by threefold. This approach enables high-frequency signal digitization using low-bandwidth ADCs, resulting in improved SNDR and ENOB.