NequIP

NequIP guarantees that its predictions for energy and forces transform correctly under rotation, translation, and inversion of the input atomic coordinates. This is achieved by operating on geometric tensors built from irreducible representations (irreps) of the 3D rotation group. Unlike models that rely on data augmentation to learn approximate invariance, NequIP's strict mathematical constraint ensures physical consistency and dramatically improves data efficiency.

Atomic interactions are modeled using tensor products of irreps, which act as a mathematically complete basis for expressing functions of relative position vectors. This mechanism allows NequIP to naturally capture complex directional information, such as the angular dependencies of covalent bonds, without needing hand-crafted angular features. The architecture systematically builds many-body correlations through successive convolutions over the atomic neighborhood.

By embedding physical symmetries as hard constraints, NequIP learns accurate potential energy surfaces from significantly smaller training datasets compared to invariant or non-equivariant models. This is critical for applications where high-fidelity reference data, such as Coupled Cluster (CCSD(T)) calculations, is computationally expensive to generate. The model focuses its capacity on learning physically meaningful representations rather than rediscovering basic symmetries.

NequIP has demonstrated leading performance on established benchmarks like the MD17 molecular dynamics dataset and the Open Catalyst Project (OC20). It achieves chemical accuracy (errors below 1 kcal/mol) for energy predictions and highly precise force predictions, often surpassing classical force fields and competing with or exceeding other deep learning potentials like SchNet and DimeNet++ while using fewer training examples.

The architecture is designed for efficient parallelization across multiple GPUs, enabling training on large-scale systems containing hundreds of thousands of atoms. NequIP leverages optimized tensor product kernels and distributed communication patterns to maintain high throughput, making it feasible to construct potentials for complex materials, alloys, and biological systems that require large periodic simulation cells.

NequIP's principled uncertainty quantification, derived from its deep ensemble or committee of models, makes it an ideal surrogate for active learning workflows. The model can identify configurations where its predictions are unreliable and request new first-principles calculations. This closes the loop between simulation and data generation, automating the construction of robust potentials that cover all relevant regions of configurational space.