Abstract—Massive Multiple-Input Multiple-Output (MIMO) is a wireless access technology used to enable 5G and nextgeneration mobile communications. This 5G network operates in a Frequency Range1(FR1) band, which includes sub-6GHz bands, and a Frequency Range2(FR2) band, which determines bands in the mm-wave range. The sub- 6GHz/mm-wave 5G networks encounter a range of difficulties in terms of adaptability, latency, throughput, and improved signal to noise ratio (SNR), where it is desirable for the User Equipment (UE) to limit the transmit power and efficiently manage radio resources to improve battery life. Fading, Bit Error Rate (BER) and noise are significant features in wireless technologies that have an impact on the quality of data and signals, such effects can be efficiently suppressed by Sub-Optimal MIMO detection/equalizer algorithms. The performance of BER Versus SNR is interpreted between Zero Forcing (ZF), ZF-Successive Interference Cancellation (SIC), Minimum Mean Square Error (MMSE), and MMSE-SIC equalizers through different modulation schemes and observed that MMSE-SIC performs best coping with low BER compared to other algorithms, and with its low BER and interference followed by admiring SNR is strongly advised at Uplink (UL)/Downlink(DL) regions. A semi-blind UL/DL decoupling algorithm is also proposed in this context, where the processing unit collects measurements of the Rician K-factor reflecting the line of sight (LOS) condition of the UE and DL reference signal receive power (RSRP) for both 2.6 GHz and 28 GHz frequency bands, followed by the training of a machine learning algorithms. For these frequency bands, the trained algorithm is utilized to make blind predictions about the target frequencies and access points that may be used separately for the UL and DL. Hence, decoupling of UL and DL has been assessed by adapting MMSE-SIC equalizer and various Machine learning and Deep learning algorithms, it is found that Convolutional neural network (CNN) achieves the highest decoupling success rate, with an average accuracy of 98.93% but a maximum accuracy of 99.83% for a small number of training samples