The ultimate objective of materials science is to be able to adapt microstructures to reach desired properties. However, no consistent constitutive models were made to date essentially because of the need to statistically link the microscopic and macroscopic scales. In this project, we propose an original methodology where a crystalline plasticity code will be coupled to a supervised learning algorithm to obtain a system capable of suggesting the distribution of operating mechanisms in a polycrystal with its microstructural parameters in order to obtain desired macroscopic mechanical properties. This new model resulting from the learning process will be instructed from a large set of experimental data obtained by scanning electron microscopy and translated into an input-output system. This project will have a major impact in current societal issues by enabling energy savings and limited costs associated with the tuning of microstructures targeting specific mechanical performances.