We present a new approach for thermal rendering in telepresence which improves transparency; it aims at reaching, as closely as possible, what is experienced in similar direct touch conditions. Our method is based on a neural networks learning classifier that allows generating appropriate thermal values (i.e., time trajectories) used as desired inputs of two independent controllers: the one controlling a bio-inspired remote thermal sensing device (i.e., an artificial finger), and the other one controlling the user's thermal display. To do so, two databases are built from real measurements recorded during direct contact between the operator's finger and different materials. One database is used for training a classifier to be used in online identification of the material being remotely explored; the other is used to generate desired thermal trajectories for the previously evoked control loops. The learning bloc is based on principal component analysis and a feed-forward neural network. Experimental tests validating our method in different scenarios have been carried out; the obtained results are discussed.