---
title: 'DetaNet: Deep Equivariant Tensor Attention Network'
url: https://www.emergentmind.com/topics/deep-equivariant-tensor-attention-network-detanet
type: topic
---

# DetaNet: Deep Equivariant Tensor Attention Network

Searching arXiv for the specified DetaNet papers to ground the article in current preprints.
arXiv search query: 2510.04227
Deep Equivariant Tensor Attention Network (DetaNet) denotes two distinct tensor-equivariant neural architectures introduced in 2025 in different research domains: molecular simulation and vibrational spectroscopy, and symbol-level precoding for wireless communications. In the molecular setting, DetaNet is an $E(3)$-equivariant, tensor-aware graph neural network trained to predict total energy $E$, forces $\mathbf{F}$, dipole moments $\boldsymbol{\mu}$, and polarizability tensors $\boldsymbol{\alpha}$, and then coupled to machine-learning molecular dynamics (MLMD) and ring-polymer molecular dynamics (RPMD) for infrared and Raman spectral prediction [2510.04227]. In wireless communications, DetaNet is an attention-based tensor-equivariant backbone that exploits permutation symmetries across users and symbols, as well as higher-order equivariance on user-user interaction tensors, to approximate optimal constructive-interference symbol-level precoding with low inference complexity [2510.02108]. The shared acronym reflects a common design principle—embedding tensor symmetry directly into network operators—but not a single canonical architecture.

## 1. Terminology, scope, and disambiguation

In current arXiv usage, the name DetaNet is not unique to one application area. One paper defines DetaNet as a “deep equivariant tensor attention network” for universal force-field learning and tensorial response prediction in chemistry and materials simulation, while another uses the same name for a tensor-equivariant attention network for symbol-level precoding under perfect and imperfect channel-state information [2510.04227] [2510.02108].

| Aspect | Molecular DetaNet | Precoding DetaNet |
|---|---|---|
| Domain | MLMD, IR/Raman spectra | Symbol-level precoding |
| Primary symmetry | $E(3)/SE(3)$ equivariance | User/symbol permutation TE |
| Main outputs | $E$, $\mathbf{F}$, $\boldsymbol{\mu}$, $\boldsymbol{\alpha}$ | $\delta_\mu$, $\delta_\nu$, or $\Psi$ |

This terminological overlap has a practical consequence for literature reading: “DetaNet” does not by itself specify a unique model family. A common misconception is to treat it as a single architecture transferable across fields. The published record instead shows two separately defined architectures that share tensor-aware attention and symmetry-preserving parameter sharing, but differ in symmetry group, input geometry, objective, and downstream task.

In the molecular work, the central requirement is explicit Euclidean equivariance and direct prediction of mixed-order quantities: scalar energy, vectorial forces and dipoles, and a rank-2 polarizability tensor. In the communications work, the central requirement is tensor equivariance under permutations of user and symbol axes, together with invariance on the antenna axis in the robust-CSI module. The two works therefore instantiate “equivariance” with different group actions and different operator families.

## 2. Symmetry formalism and tensor-attention principles

The conceptual core of both DetaNet variants is that the network should transform outputs consistently when structured transformations are applied to inputs. In the molecular formulation, rigid translations are removed by constructing geometry from relative displacements $\mathbf{r}_{ij} = \mathbf{r}_j - \mathbf{r}_i$, and rotations $R \in SO(3)$ act on outputs according to
$$
E' = E,\qquad \mathbf{F}'_i = R\mathbf{F}_i,\qquad \boldsymbol{\mu}' = R\boldsymbol{\mu},\qquad \boldsymbol{\alpha}' = R\boldsymbol{\alpha}R^\top.
$$
Permutation symmetry over atom labels is enforced by symmetric aggregation, such as sum pooling for global quantities [2510.04227].

The molecular architecture follows a typical $E(3)$-equivariant construction. Directional dependence is expanded in spherical harmonics $Y_{\ell m}(\hat{r}_{ij})$ with learned radial basis functions $\phi_\ell(r_{ij})$, producing equivariant edge features. Couplings between different tensor orders are handled through Clebsch–Gordan tensor products, and tensor attention is designed so that queries, keys, and values respect feature type. Scalar channels generate invariant weights, whereas vector and tensor channels modulate direction-dependent messaging through equivariant dot-products and Clebsch–Gordan coupling. Gating keeps nonlinearities type-consistent.

In the communications formulation, tensor equivariance (TE) is defined on multiple tensor modes rather than on Euclidean space. If permutation matrices $\Pi_u$ and $\Pi_b$ act on user and symbol modes, then the map is required to permute outputs correspondingly when those input modes are permuted. The paper emphasizes that this is not ordinary single-mode graph equivariance: it includes multidimensional equivariance (MDE) on user and symbol axes and higher-order equivariance (HOE) on user-user interaction tensors [2510.02108].

This difference matters analytically. In the molecular case, symmetry is physical, arising from invariance under rigid motions and relabeling of indistinguishable atoms. In the precoding case, symmetry is structural, arising from the fact that permuting users or symbols reorders an optimization problem without changing its mathematical form. A plausible implication is that the molecular DetaNet is constrained by physical covariance laws, whereas the precoding DetaNet is constrained by problem-index symmetries encoded in Karush–Kuhn–Tucker (KKT) information tensors.

## 3. Molecular DetaNet: architecture, training targets, and dynamical workflow

The molecular DetaNet takes atom types $Z$, atomic positions $R = \{r_i\}$, and optionally periodic cell parameters as input, forms a molecular graph with edges inside a cutoff, and propagates messages in typed channels corresponding to scalars, vectors, and rank-2 tensors. Separate equivariant readout heads produce global scalar energy, per-atom vector forces, a global dipole vector, and a global rank-2 polarizability tensor. Polarizability is predicted directly rather than reconstructed post hoc, and the isotropic component $\alpha_{\mathrm{iso}} = \frac{1}{3}\mathrm{Tr}\,\alpha$ follows naturally from the tensor output [2510.04227].

Training is multitask. The targets are total energy in eV, per-atom forces in eV/Å, dipole in Debye, and polarizability in Å$^3$. The loss is a weighted sum of mean-squared errors over these properties plus optional regularizers. The model is trained explicitly on forces rather than relying only on energy gradients, although energy-conserving forces can also be obtained by automatic differentiation,
$$
\mathbf{F}_i = - \nabla_{\mathbf{r}_i} E(\{\mathbf{r}\}).
$$
The training corpus is the QMe14S dataset, comprising 186,102 small organic molecules with equilibrium and non-equilibrium geometries sampled by ADMP, and generated with Gaussian 16; the split is 90% train, 5% validation, and 5% test. On the QMe14S test set, the reported errors are MAE/RMSE/$R^2$ of 0.0401/0.0532/0.99993 for $E$, 0.0348/0.0576/0.99856 for $\mathbf{F}$, 0.0253/0.0390/0.99943 for $\boldsymbol{\mu}$, and 0.2432/0.4825/0.99962 for $\boldsymbol{\alpha}$.

For dynamics, DetaNet is coupled to a velocity-Verlet integrator. Initial momenta are drawn from the Maxwell–Boltzmann distribution at the target temperature, accelerations are computed from DetaNet forces, and the thermostat is a Nosé–Hoover chain of length 3 with relaxation time 100 fs at 300 K. The MLMD workflow uses 50 ps total trajectories, discards the first 10 ps as equilibration, and analyzes the remaining 40 ps. The timestep is 0.2 fs for isolated molecules, polycyclic aromatic hydrocarbons, and ethanol aggregates, and 0.5 fs for 2-methylpyrazine aggregates, paracetamol and SiO$_2$ crystals, and polypeptides.

Spectra are obtained from time-correlation functions. Infrared intensity is computed from dipole autocorrelation or equivalently dipole-velocity autocorrelation, and Raman intensity from isotropic and anisotropic polarizability autocorrelations. DetaNet supplies $\mu(t)$ and $\alpha(t)$ on the fly along MLMD or RPMD trajectories; the autocorrelation functions from the last 40 ps are Fourier transformed to frequency-domain spectra, thereby capturing anharmonicity and thermal broadening without artificial line-shape postprocessing. Nuclear quantum effects are introduced with RPMD using $P=45$ beads for isolated molecules and $P=48$ for crystals, implemented through SchNetPack and i-PI with DetaNet providing energies and forces to all beads in parallel.

## 4. Precoding DetaNet: optimization structure, TE layers, and AMDE backbone

In wireless communications, DetaNet is formulated for downlink symbol-level precoding with constructive interference for $K$ users, $N_T$ transmit antennas, and $L$ symbols per coherence block. The received signal is modeled by
$$
y_k[l] = h_k^\top x_c[l] + n_k[l], \qquad n_k[l] \sim \mathcal{CN}(0,\sigma^2),
$$
and the constructive-interference region is parameterized through perturbations along two boundary directions $\mu_k$ and $\nu_k$. Two canonical optimization targets are used: CIZF, a CI-constrained SINR-balancing formulation, and CIMMSE, a CI-constrained MMSE formulation. Both reduce to power-minimization problems and, in real-valued form, to non-negative least squares (NNLS) problems [2510.02108].

The paper’s central observation is that the optimal perturbations can be expressed as a tensor-equivariant map from KKT-derived tensors. In complex form, these are $\mathfrak{B}_c[l]$ and $\mathcal{C}_c[l]$, stacked over $L$ to obtain $\mathfrak{B}_c \in \mathbb{C}^{K \times L \times 2}$, $\mathcal{C}_c \in \mathbb{C}^{K \times K \times L \times 4}$, and output perturbations $\mathfrak{D} \in \mathbb{R}^{K \times L \times 2}$. The proposition proved in the paper states that the mapping from $(\mathcal{C}_c,\mathfrak{B}_c)$ to $\mathfrak{D}^\star$ is equivariant to permutations over users and symbols, with HOE on the user-user interaction tensor.

The architecture consists of TE input projection, AMDE blocks, and a closed-form post-net scalar refinement. TE layers include MDE and HOE operators with parameter sharing over permutation orbits. The AMDE block combines residual TE layers with two attentional components: a feature attention equivariant module (FA-MDE) and an equivariant-dimension attention module (EA-MDE). For perfect CSI, the resulting SLPN takes
$$
\mathbf{X} \in \mathbb{R}^{K \times K \times L \times 8}, \qquad \mathbf{Y} \in \mathbb{R}^{K \times L \times 4},
$$
and outputs $\hat{\mathfrak{D}} \in \mathbb{R}^{K \times L \times 2}$, with non-negativity enforced by ReLU. A post-net one-dimensional refinement chooses $\rho^\star[l] \ge 0$ in closed form to minimize the power-minimization objective, after which the transmit vector is reconstructed using the known CIZF or CIMMSE closed forms.

Under imperfect CSI, the framework is split into two subnets. RSLPN-A estimates the auxiliary diagonal variable $\Psi[l]$ from a tensor input $\mathbf{T} \in \mathbb{R}^{K \times N_T \times L \times 8}$ while enforcing TE over users and symbols and invariance over antennas. RSLPN-B then uses robust KKT tensors constructed from $\Psi$ to estimate perturbations. The robust design is built on an a posteriori channel-aging model and a closed-form robust MMSE precoder.

Supervised training uses MSE loss on either perturbation tensors or auxiliary $\Psi$ labels. The data are generated with QuaDRiGa under 3GPP TR 38.901 v19.0.0 UMa NLOS at 3.5 GHz. The optimizer is Adam with batch size 400 and a two-stage learning rate schedule of $5\times10^{-3}$ followed by $5\times10^{-4}$. SLPN is trained for 800 epochs in CIZF and 400 epochs in CIMMSE; RSLPN-A and RSLPN-B are each trained for 300 epochs.

## 5. Empirical performance, transferability, and scaling

The molecular DetaNet is presented as a universal and transferable force field over small organics and as a fine-tunable backbone for crystals, aggregates, and biomolecules. Reported speedups relative to AIMD are substantial: more than $500\times$ for isolated molecules relative to Gaussian-based AIMD, and $750$–$1000\times$ for extended systems relative to CP2K AIMD. A cited example is a paracetamol crystal with 80 atoms for a 2 ps trajectory including dipole and polarizability, where DetaNet requires about 1,600 s versus about $1.2 \times 10^6$ s for CP2K. For isolated molecules, a 45-atom case is reported as more than 10,000 s for AIMD versus 8.08 s for DetaNet. The framework is also reported to support proteins up to 9,244 atoms on commodity CPU nodes with DFT-level accuracy in tensorial properties and spectra [2510.04227].

Spectrally, the molecular study reports that across molecules and PAHs, similarity to experiment improves in the order QC-Harmonic $<$ DetaNet-MLMD $<$ DetaNet-RPMD. Specific examples include ethanol IR, where cosine/Pearson/Spearman scores improve from 0.45/0.38/0.69 in QC to 0.66/0.60/0.70 in MLMD and 0.94/0.92/0.80 in RPMD; benzene Raman, where QC gives 0.20/0.15/0.42, MLMD 0.89/0.81/0.50, and RPMD 0.90/0.89/0.55; and paracetamol-crystal IR, where MLMD yields cosine 0.87 and RPMD yields Pearson 0.90 with experiment-alignment Pearson values of 0.85 and 0.88, respectively. Transfer learning from QMe14S to system-specific AIMD data is reported to reduce data requirements. For paracetamol crystal, force MAE/$R^2$ changes from about 0.04946 eV/Å and 0.98739 for direct training on 2,000 frames to about 0.01961 eV/Å and 0.99912 when fine-tuning from QMe14S with the same 2,000 frames.

The precoding DetaNet likewise emphasizes performance-speed tradeoffs. Inference complexity is stated to be linear in the sizes of equivariant axes, with parameter counts independent of $K$, $L$, and $N_T$. For perfect CSI, empirical runtime is about $80\times$ faster than conventional CIZF/CIMMSE; one reported operating point is $N_T=14$, $K=12$, $L=100$, SNR = 30 dB, where CIZF-DL uses about 3.62% of CIZF CPU runtime and about 1.22% of CIZF GPU runtime. For robust CSI at $N_T=14$, $K=12$, $L=50$, $\alpha=0.995$, and SNR = 30 dB, RCIMMSE-DL uses about 0.18% of RCIMMSE CPU runtime and about 0.07% of RCIMMSE GPU runtime [2510.02108].

Communication performance is reported in terms of SER, power, and MSE. Under perfect CSI with $N_T=K=12$ and 4-QAM, CIMMSE-DL requires approximately 6 dB less SNR than MMSE to reach SER $10^{-2}$; CIZF-DL exceeds MMSE by approximately 3 dB and CIZF-CF by more than 6 dB at medium-to-high SNR. For transmit power at $N_T=12$, $K=12$, and 4-QAM, CIZF-DL reduces required power by approximately 11.3 dB versus ZF and approximately 5.3 dB versus CIZF-CF. At $N_T=14$, $K=12$, it is reported to match CIZF nearly exactly. Under imperfect CSI with $\alpha=0.995$, $N_T=14$, $K=12$, and QPSK, RCIMMSE-DL closely tracks RCIMMSE in MSE and at 40 dB yields SER about $3.2 \times 10^{-4}$ versus about $2.4 \times 10^{-4}$ for RCIMMSE, compared with about $4.7 \times 10^{-3}$ for CIMMSE.

## 6. Limitations, misconceptions, and future directions

The two DetaNet lines of work expose different limitations. In the molecular setting, force-field fidelity and spectral quality depend on the reference electronic structure, including exchange-correlation functional and basis; the paper states that M06-2X performed better for O–H stretches than PBE or BP86 in tests. Long-range interactions remain sensitive to cutoff radius, especially in biomolecules, where increasing the cutoff from 5 Å to 8 Å markedly improves force accuracy and IR bands, including amide I/II/III. RPMD is computationally heavy for large biomolecules, so MLMD is used for big peptides and nuclear quantum effects remain difficult to include at scale. The model is restricted to ground-state potential-energy surfaces and linear response via $\mu$ and $\alpha$; excited-state properties and electronic-vibrational couplings are not modeled. The paper also notes that small deviations in relative Raman or IR band intensities, such as amide I versus II in the 20-mer peptide, suggest that further refinement of $\mu/\alpha$ surfaces and environment or polarization effects may be needed [2510.04227].

In the communications setting, the authors explicitly note that CIR-based modeling depends on the boundary-direction definitions $\mu$ and $\nu$, and that accuracy degrades in extreme modulations such as very high-order QAM or under severe channel correlation where the CIR may be tight or nonconvex. When $N_T \gg K$, the perturbations approach zero and SLP collapses toward linear precoding; the model remains consistent, but the incremental gain over LP becomes limited. Robust performance also depends on the quality of $\bar{H}$ and $m_k$ estimation in the a posteriori channel-aging model, and antenna-mode pooling may be inadequate if antenna-dependent asymmetries, such as per-antenna power constraints, are operationally important [2510.02108].

A recurring misconception is that “equivariance” in DetaNet has a uniform meaning across applications. The literature here indicates otherwise. In one case, equivariance enforces covariance under $SO(3)$ rotations and invariance under translations and atom relabeling; in the other, it enforces permutation equivariance over discrete tensor modes and invariance over an antenna axis. Another misconception is that tensor attention alone defines DetaNet. In both papers, attention is only one component of a broader symmetry-aware design that includes typed channels, tensor products, pooled invariants, residual structure, or closed-form post-processing.

The forward-looking agenda is similarly domain-specific. The molecular work points toward improved long-range electrostatics, better treatment of many-body polarization, higher-level reference-data fine-tuning, and broader property coverage. The communications work points toward joint SLP and receiver demodulation, mixed-integer scheduling with TE, per-antenna constraints, multi-carrier extensions, and integration with statistical CSI models. Taken together, these papers position DetaNet less as a single model than as a design pattern: symmetry-constrained tensor attention specialized to the transformation laws of a target scientific or engineering problem.

Source: https://www.emergentmind.com/topics/deep-equivariant-tensor-attention-network-detanet