Considerable attempts have been made on removing the crosstalk noise in a simultaneous source data using the popular K-means Singular Value Decomposition algorithm ( KSVD) ． Several hybrids of this method have been designed and successfully deployed，but the complex nature of blending noise makes it difficult to manipulate easily． One of the challenges of the K-means Singular Value Decomposition approach is the challenge to obtain an exact KSVD for each data patch which is believed to result in a better output． In this work，we propose a learnable architecture capable of data training while retaining the K-means Singular Value Decomposition essence to deblend simultaneous source data