Tensor Field Network

TFNs achieve SE(3) equivariance by design, meaning the network's output transforms predictably under any 3D rotation and translation of the input point cloud. This is accomplished by restricting learned filters to be functions of relative position and using spherical harmonic tensor products, ensuring that physical predictions like molecular forces are independent of the molecule's orientation in space.

Unlike scalar-only networks, TFNs propagate geometric tensors of arbitrary order (scalars, vectors, matrices, etc.) through the network layers. This allows the model to capture complex directional information:

• Scalars (order-0): Rotation-invariant features like energy or charge
• Vectors (order-1): Atomic forces and dipole moments
• Tensors (order-2+): Quadrupole moments and anisotropic properties

TFNs replace discrete graph message functions with continuous, learnable radial filters conditioned on interatomic distances. These filters are expanded in a spherical harmonic basis and combined via the tensor product operation, enabling the network to learn complex angular interactions between atoms without requiring hand-crafted angular features or explicit bond-angle binning.

TFNs operate directly on unstructured 3D point clouds—sets of atomic coordinates with associated features—without requiring a pre-defined molecular graph or bonding topology. This makes them inherently suitable for:

• Molecular dynamics trajectories where bonds break and form
• Non-covalent interactions like hydrogen bonding and pi-stacking
• Materials with periodic boundary conditions

The core mathematical operation in TFNs is the Clebsch-Gordan tensor product, which combines two irreducible representations of the rotation group to produce a new representation. This operation preserves angular momentum coupling rules from quantum mechanics, ensuring that the network's internal representations remain physically meaningful and that equivariance is strictly maintained throughout all layers.

TFNs laid the groundwork for modern equivariant interatomic potentials. NequIP extends TFNs with E(3) equivariance and achieves state-of-the-art data efficiency for force prediction. MACE further improves upon this by incorporating many-body expansions via higher-order tensor products, enabling highly accurate potential energy surfaces while maintaining the strict equivariance guarantees pioneered by TFNs.

The core symmetry principle behind TFNs. A function is SE(3) equivariant if rotating and translating the input point cloud produces an identically transformed output field. This ensures the model's predictions are independent of a molecule's orientation in space, a critical inductive bias for learning physical properties. TFNs achieve this by constraining their learnable filters to be products of radial functions and spherical harmonics, guaranteeing that feature tensors transform according to irreducible representations of the rotation group.

Multi-Atomic Cluster Expansion is a highly efficient equivariant message-passing architecture that builds on the body-ordered expansion principles of TFNs. It constructs messages using higher-order tensor products to systematically capture many-body interactions. MACE achieves remarkable computational efficiency by decoupling the correlation order from the computational scaling, enabling the use of high-order features without prohibitive cost. It is currently a leading architecture for materials and molecular simulation.

A generative framework that uses an SE(3)-equivariant denoising network (often built with TFN-like layers) to reverse a diffusion process on 3D coordinates. This allows for the generation of stable, physically valid molecular conformers. Key aspects:

• The network predicts the noise added to atomic coordinates while maintaining equivariance.
• Generates entire 3D geometries from scratch, not just torsional angles.
• Used for conformer generation, structure-based drug design, and molecular docking.

A systematic and complete basis set expansion of atomic environments that provides the mathematical foundation for many TFN-like models. ACE constructs a hierarchical, body-ordered expansion of the local atomic density using spherical harmonics and radial basis functions. This formalism yields highly efficient, invariant features for linear models and forms the theoretical backbone for the MACE architecture, unifying accuracy and speed.