--- title: Equivariant Tensor Network Potential url: https://www.emergentmind.com/topics/equivariant-tensor-network-potential type: topic --- # Equivariant Tensor Network Potential 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 [2304.08226]. Later work has positioned ETNP both as a compact local potential for multicomponent systems [2304.08226] and as a flexible short-range backbone that can be hybridized with explicit long-range physics such as dispersion [2504.15760] and Coulomb interactions with fixed charges [2509.15907]. ## 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 [2509.15907]. 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 \(T\) 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 [2304.08226]. 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 \[ \alpha(v^1,\ldots,v^d) = T^1 \Big( \,\ldots\, T^{d-1} \Big( T^d v^d \Big) v^{d-1} \,\ldots\, \Big) v^1, \] 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 \((\ell,m,n)\) [2304.08226, 2504.15760]. 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 \(SO(3)\), indexed by angular momentum \(\ell\), magnetic index \(m\), and channel index \(n\): \[ v_{(\ell m n)}, \qquad \ell=0,\ldots,L,\quad m=-\ell,\ldots,\ell,\quad n=1,\ldots,N(\ell). \] These features transform covariantly under rotation, while the final fully contracted scalar is invariant [2304.08226]. In the original formalism, real spherical harmonics are used. The paper defines a real basis \(\dot Y_{\ell m}\), 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 [2304.08226]. This choice enables real-valued equivariant features and undirected tensor-network contractions. A general order-3 equivariant tensor takes the factorized form \[ T_{(\ell_1 m_1 n_1) (\ell_2 m_2 n_2) (\ell_3 m_3 n_3)} = C_{(\ell_1 n_1) (\ell_2 n_2) (\ell_3 n_3)} \begin{Bmatrix} \ell_1 & \ell_2 & \ell_3 \\ m_1 & m_2 & m_3 \end{Bmatrix}, \] where the learnable reduced tensor \(C\) depends only on channel indices and angular degrees, while the Wigner-coupling term enforces the representation-theoretic selection rules [2304.08226]. In the later ETNP formulation used in practical applications, the same principle is expressed through Clebsch–Gordan factorization: \[ T_{(\ell_1 m_1 n_1)(\ell_2 m_2 n_2)(\ell_3 m_3 n_3)} = \theta_{(\ell_1 n_1)(\ell_2 n_2)(\ell_3 n_3)} C_{(\ell_1 m_1)(\ell_2 m_2)(\ell_3 m_3)}. \] This is explicitly the mechanism by which ETNP remains equivariant while still producing rotationally invariant scalar energies [2504.15760]. The original ETN paper develops the formalism primarily for \(SO(3)\). It further notes that full \(O(3)\) 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 [2304.08226]. Later ETNP papers inherit this rotational construction and use it in chemistry and materials settings without rederiving the group theory in full [2504.15760, 2509.15907]. ## 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 \[ v_{(\ell m n)}(\{r_{ij}\}) = \sum_j \dot Y_{\ell m}(\widehat r_{ij}) f_n, \] with \(f_n\) 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 \[ v_{(\ell m\mu\beta\gamma)}(\{ r_{ij} \}, z^i, \{ z^j \}) = \sum_j \dot Y_{\ell m}(\widehat r_{ij}) Q_\mu(r_{ij}) z^i_\beta z^j_\gamma, \] where \(Q_\mu(r_{ij})\) is a radial basis and \(z^i_\beta,z^j_\gamma\) are central and neighbor species indicators [2304.08226]. A low-rank front-end compression is then performed before the tensor train. In the original notation, \[ v_{(\ell m n)} = B_{\ell n\alpha\lambda} A_{\ell\lambda(\beta\gamma)} F_{\ell m\alpha(\beta\gamma)}, \] with \(A_\ell\) compressing species-pair information and \(B_\ell\) mixing radial and species channels into the final reduced channel index \(n\) [2304.08226]. The ETNP form used in subsequent application papers writes the same idea more explicitly as \[ v_{(\ell m n)} = \sum_j \Big( B_{\ell n\alpha\lambda} \, Q_\alpha(|r_{ij}|) \, \big(A_{\ell\lambda\beta\gamma} z_\beta^i z_\gamma^j\big) \Big) Y_{\ell m}\!\left(\frac{r_{ij}}{|r_{ij}|}\right). \] This makes clear that the ETNP environment descriptor is built by summing over neighbors, multiplying a radial basis \(Q_\alpha\), species compression tensors \(A\) and \(B\), and spherical harmonics \(Y_{\ell m}\) [2504.15760]. 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 \((\ell,m,n)\): \[ \begin{multline} V^{\rm ETN}(\mathfrak n_i) = \left(T^1_{(\ell_1' m_1' n_1') (\ell_1 m_1 n_1)} \, v_{(\ell_1' m_1' n_1')}(\mathfrak n_i)\right) \left(T^2_{(\ell_1 m_1 n_1) (\ell_2' m_2' n_2') (\ell_2 m_2 n_2)} \, v_{(\ell_2' m_2' n_2')}(\mathfrak n_i)\right) \\ \ldots \left(T^d_{(\ell_{d-1} m_{d-1} n_{d-1}) (\ell_d' m_d' n_d')} \, v_{(\ell_d' m_d' n_d')}(\mathfrak n_i)\right). \end{multline} \] The internal channel dimensions function as tensor-network ranks [2504.15760]. 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 \[ V^{\rm MTP}(\mathfrak n_i)=\sum_\alpha \xi_\alpha B_\alpha(\mathfrak n_i), \] where \(B_\alpha\) are invariant basis functions constructed from moment tensors \[ M_{\mu,\nu}(\mathfrak n_i)=\sum_{j=1}^{N_{\rm nbh}} f_\mu(|r_{ij}|,z_i,z_j)\, r_{ij}^{\otimes \nu}. \] The radial part is expanded as \[ f_{\mu}(|r_{ij}|,z_i,z_j) = \sum_{\beta} c^{(\beta)}_{\mu,z_i,z_j} T^{(\beta)}(|r_{ij}|)\, (R_{\rm cut}-|r_{ij}|)^2. \] Basis functions are restricted by a level measure \[ {\rm lev}\, M_{\mu,\nu}=2+4\mu+\nu, \qquad {\rm lev}\, B_\alpha = \sum_{p=1}^P (2+4\mu_p+\nu_p), \] and all \(B_\alpha\) with \({\rm lev}\,B_\alpha \le {\rm lev}_{\max}\) are included [2504.15760]. 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 [2509.15907]. 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 [2509.15907]. 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 [2304.08226]. 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” [2304.08226]. For order-2 tensors, QR and SVD can be performed blockwise over \(\ell\). 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 [2304.08226]. The same paper further notes that alternating least squares, backpropagation, and DMRG-like two-site optimization all extend naturally to the equivariant setting [2304.08226]. 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 [2304.08226]. Later ETNP application papers retained BFGS training with losses over energies and forces, typically using \[ w_{\rm e}=1,\qquad w_{\rm f}=0.01 \] in their least-squares objectives [2504.15760]. 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 [2304.08226]. On the MoNbTaW medium-entropy alloy and MoNbTaVW high-entropy alloy datasets, ETNs required about \(2\)–\(3\times\) fewer coefficients than MTPs at similar error levels [2304.08226]. In the BA10 binary-alloy benchmark, the advantage was smaller but remained significant, around \(1.5\)–\(1.8\times\) fewer coefficients [2304.08226]. 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 \(0.084/0.56\), \(0.025/0.30\), and \(0.018/0.24\), respectively, while GAP, cGAP, and trGAP reported \(0.032/0.41\), \(0.019/0.33\), and \(0.017/0.33\) [2304.08226]. ETN3 thus matched the best energy error and improved the force error in that comparison [2304.08226]. 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 [2304.08226]. 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 [2504.15760, 2509.15907]. 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 [2504.15760]. The total energy therefore takes the form \[ E^{\rm ETN+D2}=E^{\rm ETN}+E^{D2}, \qquad E^{\rm ETN+D3}=E^{\rm ETN}+E^{D3}. \] For D2, the paper uses \[ E^{D2} = - s_6 \sum_{ij} \frac{C_6^{ij}}{r_{ij}^6} f_{\rm damp}(r_{ij}), \] and for D3 \[ E^{D3} = - \sum_{ij}\sum_{n=6,8} s_n \frac{C_n^{ij}}{r_{ij}^n} f_{\rm damp}^{(n)}(r_{ij}). \] In the implementation, both D2 and D3 were used with Becke–Johnson damping rather than the original D2 Fermi damping [2504.15760]. The practical conclusion of that work is system dependent. For carbon tetrachloride and methane, increasing the ETNP cutoff to \(7.5\) 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 [2504.15760]. 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 [2504.15760]. 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 [2504.15760]. The Coulomb-augmentation paper makes the same conceptual move for charged molecular dimers. It defines the total energy as \[ E^{\rm total}(\bm x,\bm\theta,\bm a) = E^{\rm short}(\bm x,\bm\theta) + E^{\rm long}(\bm x,\bm a), \] with the ETN-specific form \[ E^{\rm total}_{\rm ETN+QRd}(\bm x,\bm\theta,\bm a) = \sum_{i=1}^N V^{\rm ETN}(\mathbf{\mathfrak n}_i) + \sum_{i