PaiNN (Polarizable Atom Interaction Neural Network) is an E(3)-equivariant message passing neural network that maintains both scalar (invariant) and vector (equivariant) features for each atom, passing directional messages that explicitly track the orientation of forces and dipole moments — achieving state-of-the-art accuracy for molecular property prediction and force field learning by combining the efficiency of EGNN-style coordinate processing with richer geometric information through first-order ($l=1$) equivariant features.

What Is PaiNN?

• Definition: PaiNN (Schütt et al., 2021) maintains two feature types per atom: scalar features $s_i in mathbb{R}^F$ (invariant under rotation) and vector features $vec{v}_i in mathbb{R}^{F imes 3}$ (transform as 3D vectors under rotation). Each message passing layer performs: (1) Message: compute scalar messages from distances and features; (2) Update scalars: aggregate scalar messages from neighbors; (3) Update vectors: aggregate directional messages $Deltavec{v}_{ij} = phi_v(s_j, d_{ij}) cdot hat{r}_{ij}$ where $hat{r}_{ij}$ is the unit direction vector from $j$ to $i$; (4) Mix: interchange information between scalar and vector channels through inner products $langle vec{v}_i, vec{v}_i angle$ and scaling $s_i cdot vec{v}_i$.
• Scalar-Vector Interaction: The key innovation is the equivariant mixing between scalar and vector features — the inner product $langle vec{v}_i, vec{v}_i angle$ creates rotation-invariant scalars from vectors (useful for energy prediction), while scalar multiplication $s_i cdot vec{v}_i$ modulates vector features with learned scalar gates (useful for force prediction). These operations are the only equivariant bilinear operations at order $l leq 1$.
• Radial Basis Expansion: Like SchNet, PaiNN expands interatomic distances using radial basis functions with a smooth cosine cutoff: $e_{RBF}(d) = sin(n pi d / d_{cut}) / d$, combined with a cutoff envelope that ensures messages smoothly vanish at the cutoff distance. This continuous distance encoding avoids discretization artifacts.

Why PaiNN Matters

• Directional Force Prediction: Predicting atomic forces for molecular dynamics requires equivariant vector outputs — the force on each atom has both magnitude and direction that must rotate with the molecule. PaiNN's vector features naturally produce equivariant force predictions without requiring energy-gradient computation (which requires backpropagation through the energy model), enabling 2–5× faster force evaluation.
• Dipole and Polarizability: Molecular dipole moments (vectors) and polarizability tensors require equivariant and second-order equivariant outputs respectively. PaiNN's vector features directly predict dipole moments, and outer products of vector features yield polarizability predictions — enabling prediction of spectroscopic properties that scalar-only models cannot represent.
• Efficiency-Accuracy Balance: PaiNN achieves accuracy comparable to DimeNet++ (which uses expensive angle computations) at significantly lower computational cost by using $l=1$ equivariant features instead of explicit angle calculations. This positions PaiNN in the "sweet spot" between minimal models (EGNN, distance-only) and high-order models (MACE, NequIP with $l geq 2$).
• Neural Force Fields: PaiNN is one of the most widely used architectures for training neural network interatomic potentials — learning to predict energies and forces from quantum mechanical training data (DFT calculations), then running molecular dynamics simulations 1000× faster than the original quantum calculations while maintaining near-DFT accuracy.

PaiNN Feature Types

PaiNN is vector-aware molecular messaging — maintaining explicit directional features alongside scalar features for each atom, providing the geometric resolution needed to predict forces, dipoles, and other directional molecular properties with an efficiency-accuracy balance that makes it a workhorse for neural molecular dynamics.