This study investigates non-orthogonal multiple access (NOMA) receivers based on deep learning (DL) employing the long-short term memory technique (LSTM) over frequency-flat Rayleigh distributed fading links. The fading links are independently and identically distributed (i.i.d.). When comparing the DL-based NOMA receiver’s system performance to that of the traditional NOMA technique, the DL-based receiver surpasses the conventional successive interference cancellation (SIC)-based NOMA receiver. The simulations are conducted for various values of cyclic prefix (CP) considering the clipping noise (CN) under real-time propagation characteristics. It has been discovered that neither minimum mean square error (MMSE) nor least square error (LSE) can provide precise information on fading channel coefficients. With a signal-to-noise ratio (SNR) value exceeding 14 dB, precision tends to be saturated. On the other hand, DL techniques continue to be effective in channel estimation and detection. Lower learning rates improve system performance, whereas a high learning rate generates rapid changes in the weights of the DL NOMA detector, leading to a very high validation error value.