MODEL LEARNING PATH
NequIP
NequIP builds an invariant atomic energy from equivariant geometric features. Summing atomic contributions gives system energy, and differentiating that energy yields conservative forces.
Loading the interactive page…
Made E(3)-equivariant tensor message passing a compelling approach to data-efficient, conservative learned interatomic potentials, with subsequent architectures and simulation systems building on the approach.
Chapters
Describe atoms and neighbors
Species, positions and cell information determine neighbor geometry and initial scalar features.
Construct neighbor geometry with species and periodic-cell information.
Learn equivariant interactions
Radial features and spherical harmonics drive tensor-product messages and symmetry-compatible gated updates.
Distinguish radial scalar features from equivariant angular messages.
Read a system energy
Invariant scalar channels yield atom energies, which sum to the system’s total energy.
Sum invariant atomic energies into a system energy.
Differentiate for forces
Coordinate derivatives of the same total energy give the force prediction.
Obtain forces as the negative coordinate derivative of energy.
All exercises in this guide
- Periodic Radius Graph
- Masked Sum
- Masked Mean
- Stable vector norms, normalization, dot and cross features
- Masked pairwise distances and hard distance-bin targets
- Masked batched kNN graphs and neighbor gather
- Ragged segment sum and mean
- Typed radius, bond-prefix and sequence-neighbor graph assembly
- Affine Projection
- Affine projection and embedding lookup
- Origin-safe Bessel channels with a compact polynomial cutoff
- Real spherical harmonics from Cartesian polynomial tables
- Real-irrep Clebsch–Gordan paths and channel tensor products
- Radial-weighted irrep convolution with receiver aggregation
- Species-conditioned self connections and parity-aware gates
- Distance basis expansion with smooth finite support
- Feedforward Block
- One EGNN scalar-feature and coordinate update
- Differentiate energy into forces and train through forces