Using statistical learning based optimization technique, we provide rigorous modelling and optimal design of graded index photonic metalens. Our dielectric metalens consists of silicon (Si) strips of different widths and operating at the telecommunication wavelnegth. In this work, we tackle challenging Transverse Electric case (TE), where the electric field is polarized perpendicular to the strips. First, we show for the classical design with fixed strips lengths in which an efficient device of 80% efficiency can be realized. Second, we prove that the efficiency can reach 87% in the case of strips with different lengths. This finding opens the door for the realization of highly efficient integrated photonic components especially for the TE based applications.