Reference site interaction site model (also known as 1D/3D-RISM) based on the statistical-mechanical molecular theory of solvation provides an exceptionally successive approach to study thermodynamics and nanomorphology of complex polymer blends as opposed to time-consuming and computationally expensive computer simulations. Moreover, combined with electronic structure theory methods (DFT, semi-empirical) it allows to successively predict the solvation structure around macro and biomolecules in a self-consistent way taking into account quantum effects. However, because the main equation (Ornstein-Zernike equation) of the theory relates two unknown functions, the total h(r) and direct c(r) correlations, another expression (closure) between these two functions are needed. The latter function cannot be obtained in a closed analytical form as it is composed of an infinite series of slowly converging diagrams and needs to be approximated. Because of this fact, thermodynamic quantities obtained with RISM theory sometimes deviate from experimental values and MD results. In principle, the infinite series starting from particular therm could be replaced by a small rough correction depending on an "atomic fingerprint" (a description of interacting atomic sites) which is typically employed in ML to represent the system in the way acceptable for the ML algorithm. In that work, we suggest to use the ML algorithms to improve closure relation accuracy based on the results of molecular dynamics simulations for reference systems. This will help us to reach the chemical accuracy in prediction of solvation properties and so to extend the applicability of the developed methodology to the various fields of bio and material science.