Interatomic Potential
A mathematical function that approximates the energy of a system of atoms given their positions — used in molecular dynamics and materials simulation to avoid the cost of full quantum mechanical calculations.
What it means
An interatomic potential (also called a force field or potential energy surface) is a mathematical function that maps the positions of atoms onto a potential energy value. In molecular simulation, rather than solving the full quantum mechanical Schrödinger equation at every step (prohibitively expensive), the simulation uses this approximate function to compute forces and propagate atomic motion.
Classical interatomic potentials encode physical knowledge through parameterized equations — terms for bond stretching, angle bending, dihedral rotations, van der Waals interactions, and electrostatics. They are fast to evaluate but limited to systems similar to their parameterization domain.
Machine learning interatomic potentials (MLIPs) — also called neural network potentials or machine learning force fields — replace these hand-crafted equations with AI models trained on quantum mechanical reference data. They learn the energy landscape directly from DFT calculations, achieving near-quantum accuracy at a fraction of the cost.
Why it matters for researchers
MLIPs are the key technology enabling AI-accelerated materials discovery and computational chemistry at scale. Tools and models in this space that appear in the research literature:
- DeePMD (Deep Potential Molecular Dynamics) — widely used framework for training neural network potentials
- NequIP and MACE — equivariant neural network potentials with high accuracy on small molecule and materials systems
- MLIP-3 / OpenMM — infrastructure for deploying MLIPs in production molecular dynamics workflows
- Universal MLIPs (MACE-MP-0, CHGNet) — pre-trained potentials covering broad regions of the periodic table, enabling immediate use without training on new data
In the context of materials discovery tools on this site: GNoME (Google DeepMind) and BayBE both sit in a pipeline where energy evaluation — whether from DFT or an MLIP — is a bottleneck. MLIPs reduce this bottleneck by orders of magnitude.