Equivariant Tensor Network Potential

• Equivariant Tensor Network Potential is a symmetry-constrained, low-rank ML potential that represents local atomic energies using SO(3)-equivariant tensor networks.
• It compresses high-dimensional polynomial coefficients via tensor‐train contractions, preserving rotational symmetry with spherical harmonics and Wigner coupling.
• ETNP has been integrated with explicit dispersion and Coulomb corrections to accurately capture long-range interactions in multicomponent systems.

Searching arXiv for the core ETN/ETNP paper and closely related follow-ups. Equivariant Tensor Network Potential (ETNP) denotes a class of machine-learning interatomic potentials in which the local atomic energy is represented by an $SO(3)$-equivariant tensor network, typically an equivariant tensor train, acting on symmetry-adapted features of an atom’s neighborhood. In the formulation introduced by Hodapp and Shapeev, ETNP is a polynomial-based, low-rank alternative to conventional local polynomial MLIPs: it preserves rotational structure through irreducible $SO(3)$ channels and Wigner-coupling constraints, while compressing the coefficient tensors that would otherwise grow prohibitively with feature dimension and interaction order (Hodapp et al., 2023). Later work has positioned ETNP both as a compact local potential for multicomponent systems (Hodapp et al., 2023) and as a flexible short-range backbone that can be hybridized with explicit long-range physics such as dispersion (Chalykh et al., 22 Apr 2025) and Coulomb interactions with fixed charges (Korogod et al., 19 Sep 2025).

1. Formal definition and local energy decomposition

ETNP is a local atom-centered potential: the total energy is expressed as a sum of atomic or site energies, each depending only on a finite neighborhood $\mathfrak n_i$ of atom $i$. In the ETN-specific notation used in later applications, the short-range energy is written as

$$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$

and more generally

$$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$

This locality is the same extensivity principle used in many MLIPs, but ETNP differs in how the site energy $V^{\rm ETN}$ is parameterized (Korogod et al., 19 Sep 2025).

The starting point is the generic polynomial multilinear form

$$V(\mathfrak n_i)=T_{k_1\ldots k_d} v^1_{k_1}\cdots v^d_{k_d},$$

where $d$ controls effective body order and the $v^a$ are feature vectors. A direct representation of the coefficient tensor $SO(3)$0 is expensive because the number of parameters grows rapidly with both polynomial degree and feature dimension. ETNP replaces this full coefficient tensor by a low-rank tensor network, specifically a tensor-train-like sequence of equivariant contractions (Hodapp et al., 2023).

In the original ETN formulation, the local energy is a scalar obtained by contracting several copies of a covariant feature vector through equivariant order-2 and order-3 tensors. In one tensor-train expression, the scalar output is represented as

$SO(3)$1

while in the ETNP applications to materials and molecular systems the site energy is written in multi-index form as an equivariant tensor-train over channels $SO(3)$2 (Hodapp et al., 2023, Chalykh et al., 22 Apr 2025).

This suggests a useful characterization: ETNP is a compact high-body-order polynomial potential whose coefficient structure is factorized into symmetry-constrained low-rank cores rather than stored as a monolithic tensor.

2. Symmetry structure and equivariant feature algebra

The defining property of ETNP is that symmetry is imposed analytically rather than learned implicitly. Features are decomposed into irreducible representations of $SO(3)$3, indexed by angular momentum $SO(3)$4, magnetic index $SO(3)$5, and channel index $SO(3)$6: $SO(3)$7 These features transform covariantly under rotation, while the final fully contracted scalar is invariant (Hodapp et al., 2023).

In the original formalism, real spherical harmonics are used. The paper defines a real basis $SO(3)$8, obtained from complex spherical harmonics by a unitary transformation, and constructs real-basis Wigner 3-j symbols as the fixed symmetry tensors controlling admissible angular couplings (Hodapp et al., 2023). This choice enables real-valued equivariant features and undirected tensor-network contractions.

A general order-3 equivariant tensor takes the factorized form

$SO(3)$9

where the learnable reduced tensor $\mathfrak n_i$0 depends only on channel indices and angular degrees, while the Wigner-coupling term enforces the representation-theoretic selection rules (Hodapp et al., 2023). In the later ETNP formulation used in practical applications, the same principle is expressed through Clebsch–Gordan factorization: $\mathfrak n_i$1 This is explicitly the mechanism by which ETNP remains equivariant while still producing rotationally invariant scalar energies (Chalykh et al., 22 Apr 2025).

The original ETN paper develops the formalism primarily for $\mathfrak n_i$2. It further notes that full $\mathfrak n_i$3 invariance, including reflections, is obtained when the total scalar contraction has the appropriate parity structure; in the original discussion, this is tied to the parity of summed angular momenta and to whether the resulting scalar is real (Hodapp et al., 2023). Later ETNP papers inherit this rotational construction and use it in chemistry and materials settings without rederiving the group theory in full (Chalykh et al., 22 Apr 2025, Korogod et al., 19 Sep 2025).

3. Local environment representation and tensor-train architecture

ETNP represents each local atomic environment by combining radial information, chemical identity, and angular dependence. In the original ETN formalism, a covariant environment feature is introduced as

$\mathfrak n_i$4

with $\mathfrak n_i$5 collecting non-angular pair features. For multicomponent systems, species are one-hot encoded and combined with radial basis functions. The fully species- and radial-dependent uncompressed feature is written as

$\mathfrak n_i$6

where $\mathfrak n_i$7 is a radial basis and $\mathfrak n_i$8 are central and neighbor species indicators (Hodapp et al., 2023).

A low-rank front-end compression is then performed before the tensor train. In the original notation,

$\mathfrak n_i$9

with $i$0 compressing species-pair information and $i$1 mixing radial and species channels into the final reduced channel index $i$2 (Hodapp et al., 2023).

The ETNP form used in subsequent application papers writes the same idea more explicitly as

$i$3

This makes clear that the ETNP environment descriptor is built by summing over neighbors, multiplying a radial basis $i$4, species compression tensors $i$5 and $i$6, and spherical harmonics $i$7 (Chalykh et al., 22 Apr 2025).

The atomic energy is then obtained by an equivariant tensor train. In application-oriented notation, this is written as a sequence of contractions over multi-indices $i$8: $i$9 The internal channel dimensions function as tensor-network ranks (Chalykh et al., 22 Apr 2025).

This suggests the central architectural distinction from MTP and related polynomial MLIPs: ETNP keeps intermediate objects equivariant and organizes the many-body polynomial interaction through train cores, rather than enumerating invariant basis functions and fitting a linear coefficient vector.

4. Relation to Moment Tensor Potentials and other equivariant architectures

The comparison to Moment Tensor Potential (MTP) is foundational because ETNP emerged partly as a low-rank equivariant generalization of polynomial local MLIPs. MTP writes the site energy as

$$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$0

where $$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$1 are invariant basis functions constructed from moment tensors

$$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$2

The radial part is expanded as

$$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$3

Basis functions are restricted by a level measure

$$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$4

and all $$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$5 with $$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$6 are included (Chalykh et al., 22 Apr 2025).

Conceptually, MTP is a linear expansion over invariant polynomial basis functions, whereas ETNP is a nonlinear low-rank tensor-network contraction over equivariant spherical-harmonic features (Korogod et al., 19 Sep 2025). The later Coulomb-augmentation paper makes this contrast especially explicit, summarizing the two models as follows: MTP is a linear basis expansion over systematically constructed moment-tensor basis functions, while ETN is a nonlinear low-rank tensor-network contraction over equivariant spherical-harmonic features (Korogod et al., 19 Sep 2025).

The original ETN paper also situates its method relative to SNAP and ACE. SNAP is interpreted as a low-body-order spherical-harmonic polynomial with a single equivariant coupling stage, ACE as a more complete polynomial basis generated by chaining Clebsch–Gordan-type couplings, and ETN as a low-rank multilinear alternative in which learnable equivariant order-3 tensors replace the full basis coefficient structure (Hodapp et al., 2023).

A plausible implication is that ETNP occupies a middle ground between complete but potentially large polynomial basis models and more opaque neural equivariant GNNs: it retains analytic symmetry structure and polynomial interpretability while compressing the combinatorics through tensor-network ranks.

5. Tensor-network algorithms, parameter efficiency, and training

A distinctive feature of the original ETN work is that it does not merely specify an architecture; it develops a broader tensor-network formalism for equivariant cores. Because equivariant order-2 tensors are block diagonal in angular momentum sectors, and order-3 tensors can be reshaped into block matrices after coupling angular indices, standard tensor-network operations carry over. The paper states that “all the key algorithms of tensor networks like orthogonalization of cores and DMRG-based algorithms carry over to our equivariant case” (Hodapp et al., 2023).

For order-2 tensors, QR and SVD can be performed blockwise over $$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$7. For order-3 tensors, two legs can be fused and the tensor reshaped into a matrix, allowing orthogonalization and rank truncation in a manner analogous to ordinary tensor trains (Hodapp et al., 2023). The same paper further notes that alternating least squares, backpropagation, and DMRG-like two-site optimization all extend naturally to the equivariant setting (Hodapp et al., 2023).

In practice, however, the reported ETN MLIP experiments were not yet trained with a full production tensor-network optimizer. The original paper used Python BFGS, monitored validation loss, and stopped after a fixed iteration budget (Hodapp et al., 2023). Later ETNP application papers retained BFGS training with losses over energies and forces, typically using

$$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$8

in their least-squares objectives (Chalykh et al., 22 Apr 2025).

The principal empirical argument for ETNP is parameter efficiency. The original paper reports that ETN potentials require substantially fewer parameters than MTPs for comparable accuracy on several datasets, especially multicomponent alloys (Hodapp et al., 2023). On the MoNbTaW medium-entropy alloy and MoNbTaVW high-entropy alloy datasets, ETNs required about $$E^{\rm short}_{\rm ETN}(\bm x,\bm\theta)=\sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i),$$9–$$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$0 fewer coefficients than MTPs at similar error levels (Hodapp et al., 2023). In the BA10 binary-alloy benchmark, the advantage was smaller but remained significant, around $$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$1–$$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$2 fewer coefficients (Hodapp et al., 2023).

The same paper also compared ETNs to compressed GAP variants on a high-entropy alloy dataset. Three ETNs with 92, 278, and 577 coefficients achieved test RMSE/RMSF values of $$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$3, $$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$4, and $$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$5, respectively, while GAP, cGAP, and trGAP reported $$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$6, $$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$7, and $$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$8 (Hodapp et al., 2023). ETN3 thus matched the best energy error and improved the force error in that comparison (Hodapp et al., 2023).

The original authors also reported a limitation: on QM9 and in higher-complexity ETNs, BFGS often converged prematurely, so some of the potential representational advantage was not fully realized in training (Hodapp et al., 2023). This remains an important caveat in interpreting ETNP as a model family: the architecture and symmetry formalism appear more mature than the optimization practice reported so far.

6. Explicit long-range physics: dispersion and Coulomb augmentation

A major development after the original ETN paper is the use of ETNP as a short-range equivariant backbone within hybrid models that add explicit long-range interactions. These later papers are clear that the central contribution is not a redesign of ETNP itself, but a decomposition in which ETNP learns the short-range residual while a separate analytic term enforces the physically correct asymptotic behavior (Chalykh et al., 22 Apr 2025, Korogod et al., 19 Sep 2025).

In the dispersion study, local ETN potentials were combined with D2 or D3 corrections. The workflow was to subtract the chosen dispersion contribution from ab initio reference energies and forces, fit the ETN to the residual short-/medium-range target, and then add the explicit dispersion back at inference time (Chalykh et al., 22 Apr 2025). The total energy therefore takes the form

$$E^{\rm short} = \sum_i V^{\rm ETN}(\mathbf{\mathfrak n}_i).$$9

For D2, the paper uses

$V^{\rm ETN}$0

and for D3

$V^{\rm ETN}$1

In the implementation, both D2 and D3 were used with Becke–Johnson damping rather than the original D2 Fermi damping (Chalykh et al., 22 Apr 2025).

The practical conclusion of that work is system dependent. For carbon tetrachloride and methane, increasing the ETNP cutoff to $V^{\rm ETN}$2 Å allowed plain local ETN to recover much of the benefit of explicit dispersion, though explicit D2/D3 still improved binding-curve width and asymptotics (Chalykh et al., 22 Apr 2025). For toluene, by contrast, local ETN without explicit dispersion showed overbinding, oscillatory dimer curves, density errors, and RDF artifacts, while ETN+D2 substantially improved all of these (Chalykh et al., 22 Apr 2025). The authors therefore concluded that explicit dispersion remains important for more anisotropic, orientation-sensitive systems, and that D2 is nearly as effective as D3 while being less expensive (Chalykh et al., 22 Apr 2025).

The Coulomb-augmentation paper makes the same conceptual move for charged molecular dimers. It defines the total energy as

$V^{\rm ETN}$3

with the ETN-specific form

$V^{\rm ETN}$4

The fixed-charge model is

$V^{\rm ETN}$5

which preserves the total charge exactly (Korogod et al., 19 Sep 2025).

This ETN+QRd model was introduced to correct the fundamental long-range deficiency of strictly local MLIPs: once fragments separate beyond the cutoff, a purely local model predicts interaction energies that flatten rather than following the correct Coulomb asymptotics (Korogod et al., 19 Sep 2025). On charged dimers, ETN+QRd reduced ETN energy RMSEs from $V^{\rm ETN}$6 to $V^{\rm ETN}$7 meV/atom for $V^{\rm ETN}$8, and from $V^{\rm ETN}$9 to $$V(\mathfrak n_i)=T_{k_1\ldots k_d} v^1_{k_1}\cdots v^d_{k_d},$$0 meV/atom for $$V(\mathfrak n_i)=T_{k_1\ldots k_d} v^1_{k_1}\cdots v^d_{k_d},$$1, while substantially improving binding-curve agreement with DFT (Korogod et al., 19 Sep 2025).

These hybrid studies clarify a common misconception: ETNP is not, by itself, a universal answer to long-range interactions. In the reported formulations it remains a local short-range model; long-range electrostatics or dispersion must either be learned indirectly through a large cutoff or added explicitly through analytic corrections (Chalykh et al., 22 Apr 2025, Korogod et al., 19 Sep 2025).

7. Applications, misconceptions, and broader significance

ETNP has now been used in at least three distinct ways. First, as the original polynomial tensor-network MLIP for small molecules and multicomponent alloys (Hodapp et al., 2023). Second, as a local short-range backbone augmented with explicit dispersion for molecular liquids (Chalykh et al., 22 Apr 2025). Third, as a local short-range backbone augmented with explicit Coulomb interactions for charged molecular dimers (Korogod et al., 19 Sep 2025).

Several points are therefore essential to an accurate characterization.

ETNP is not a tensor network in the quantum many-body wavefunction sense. The phrase refers here to the low-rank factorization of coefficient tensors in an interatomic potential, not to matrix product states or PEPS. The original ETN paper makes this explicit by deriving tensor-train-like contractions for polynomial MLIPs (Hodapp et al., 2023).

ETNP is also not merely “an equivariant graph neural network.” It does use equivariant feature channels and can emulate message-passing-like operations, but the reported ETN potentials are fundamentally polynomial tensor-network models rather than end-to-end neural message-passing architectures (Hodapp et al., 2023). Later papers sometimes place ETN and MTP together as local MLIP backbones and describe ETN as “an equivariant tensor network in tensor-train form,” which is the most precise concise summary in the current literature (Korogod et al., 19 Sep 2025).

A second misconception is that equivariance alone solves all physically relevant symmetries. The later ETNP papers show that locality still matters crucially. Without explicit long-range terms, ETNP misses the asymptotic physics of charged dimers and some dispersion-dominated molecular systems (Chalykh et al., 22 Apr 2025, Korogod et al., 19 Sep 2025). This suggests that equivariance and low-rank expressivity address one part of the interatomic-potential problem—efficient and symmetry-respecting representation of local many-body structure—but not the entirety of range separation and electrostatics.

A third point is that ETNP’s advantages are strongest in multicomponent settings with latent low-rank structure. The original authors interpret the good alloy results as evidence that ETNs efficiently capture similarities among species, radial channels, and higher-order interactions through shared latent channels (Hodapp et al., 2023). This suggests that ETNP is especially well matched to chemically diverse local environments where a full polynomial basis would overparameterize.

8. Prospects and open directions

The original ETN paper closes with several directions that remain technically important. It proposes implementing more mature tensor-network training procedures such as ALS and DMRG-style optimization, exploring tensor-network topologies beyond tensor trains, and incorporating richer physical features such as magnetic moments, charges, dipoles, or electronic temperature (Hodapp et al., 2023). It also notes that message passing, pooling, and attention-like mechanisms can be represented within the ETN formalism in modified tensor-contraction form (Hodapp et al., 2023).

The later hybrid papers imply a parallel agenda. Because ETNP functions well as a compact local short-range model, it is naturally suited to modular decomposition strategies in which explicit analytic long-range physics is added separately. The Coulomb and dispersion studies both support the interpretation that the ETN backbone learns the short-range residual many-body interaction, while the analytic augmentation restores correct asymptotic behavior (Chalykh et al., 22 Apr 2025, Korogod et al., 19 Sep 2025).

A plausible implication is that future ETNP developments are likely to proceed along two coupled lines. One is internal: better optimization, richer tensor-network structures, and possibly tighter integration with modern equivariant ML tooling. The other is external: broader hybridization with explicit physics, including environment-dependent charge models or more general long-range electrostatics, since fixed charges were already identified as insufficient for some ion–neutral cases (Korogod et al., 19 Sep 2025).

In the current literature, the most accurate compact summary is therefore that an Equivariant Tensor Network Potential is a symmetry-constrained, low-rank, polynomial local interatomic potential whose atomic energy is obtained by contracting spherical-harmonic environment features through an equivariant tensor network, and whose recent applications show that this local backbone can be effectively combined with explicit long-range corrections when the target systems require physics beyond the finite cutoff (Hodapp et al., 2023, Chalykh et al., 22 Apr 2025, Korogod et al., 19 Sep 2025).