Ultra-broadband communication in emerging spectrum, like Terahertz (THz) band, is the frontier to meet the data rate requirements of future wireless communication systems. Existing signal processing based methods are developed for sub-6 GHz band, which cannot capture the intricacies in ultra-broad THz bandwidth and non-linearities arising from hardware. To overcome these limitations, we develop neural network (NN) models for OFDM receiver, where expert knowledge of wireless communication is infused in different stages and parameters of the model to create a practical receiver that can adapt to different wireless environments. The parameters of the NN are derived from underlying theory and can be adapted to different wireless environments. Our model is designed to capture the correlation between real and imaginary components of wireless signals, that can be trained with limited data. The models are trained with over-the-air captured OFDM signals, transmitted in THz band with 10 GHz bandwidth. Our results show significant improvement in bit error rate (BER) for different modulation orders (upto 6 dB in BPSK and 1.2 dB in QAM 64) compared to existing signal processing based receiver designs.