The classical targeted energy transfer involves a nonlinear resonance that enables efficient transfer from a donor to an acceptor molecule. In the quantum regime a number of bosons may be transferred similarly from a certain crystal site to an alternative one using the same nonlinear resonance. While the process is complex, the use of machine learning enables learning of the optimal quantum paths in the analytically solvable case of a dimer. This knowledge may be extended to more complex quantum walks in configurations that cannot be handled analytically. This application shows that the use of machine learning derived methods in quantum processes enables the efficient design of configurations that lead to efficient quantum devices.