We are interested in detecting the cosmological imprint on properties of present dark matter halos by using Machine Learning methods. We analyse the halos formed in Dark Energy Universe Simulations using several dark energy models (ΛCDM, Quintessence Ratra Peebles), whose parameters were chosen in agreement with both CMB and SN Ia data. Their resulting halos are thus extremely close from one cosmological model to another. However, we have shown that machine learning techniques can be implemented to determine the cosmological model in which each halo was formed: we associate to each present day halos from ΛCDM and RP CDM ellipsoidal mass and shape profiles, defined to efficiently keep track of the matter distribution anisotropies and, then, we experimentally show that those quantities allow a properly trained learning device to find the dark energy model of the Universe within which these halos have grown. Training our device on 40,000 halos of 10 13 and 10 14 solar masses, we can correctly classify more than 70% of the halos in the test set. We also study the misleading ML methodological biases, "Clever Hans effects", and the way to fix them.