A machine-learning interatomic potential (often abbreviated MLIP) is a model that predicts the energy of a group of atoms, and the forces on each one, as a function of their positions. These quantities are what drive molecular-dynamics simulations, the workhorse method for studying how materials and molecules behave over time.
The reason MLIPs exist is a speed-versus-accuracy tradeoff. The most accurate way to compute atomic energies is quantum mechanics, typically density functional theory, but it is far too slow to simulate large systems or long timescales. Cheap hand-crafted formulas are fast but inaccurate. A machine-learning potential is trained on quantum-mechanical calculations and then reproduces that accuracy at a tiny fraction of the cost, letting researchers simulate thousands or millions of atoms over meaningful durations. The foundational example is the 2007 neural-network potential of Behler and Parrinello; the field has since moved to graph neural networks and equivariant architectures that respect the symmetries of physics.
For a general reader, MLIPs are a clear case of the broader pattern of replacing an expensive simulation with a learned surrogate. They are central to modern materials discovery, providing the fast, accurate stability checks behind efforts such as GNoME and helping turn AI-proposed materials into candidates worth synthesizing.