Abstract: Machine learning models are proving to be extremely effective in predicting the properties of atomistic configurations of matter,circumventing the need for time consuming electronic structure calculations. The most successful schemes achieve transferabilityby means of a local representation of structures, in which the problem of predicting a property is broken down into the predictionof local, atom-centred contributions. This approach is howevernot efficient in describing long-range interatomic forces, such asthose arising due to electrostatics, polarization, or dispersive interactions.I will discuss a few possible solutions to this problem, based on either the explicit description of the physical phenomena leading tonon-locality, or on the definition of a class of long-distance equivariant (LODE) features, that combine a localdescription of matter with the appropriate, long-range asymptoticbehaviour of interactions.