In this paper
AbstractIntroductionSU(2)-Equivariant Spinor Message PassingSpinor initializationAddressing eigenbasis gauge freedomGauge-covariant message passingTensor readoutZero-field splitting in cobalt complexesEffective spin HamiltonianPrediction accuracy and OOD ligand generalizationCrystal-field learning and spin dynamics in a dysprosium complexHamiltonian and spin propagationCrystal-field predictionElectron-spin correlationsNuclear relaxation ratesTransfer from water to methanolDiscussionOOD ligand generalizationFrom coefficients to magnetic observablesSU(2) inductive biasConclusionsAcknowledgmentsData AvailabilityRotation equivariance and eigenbasis independenceFrom spin propagation to nuclear relaxationReferencesSupplemental MaterialImplementation and symmetry detailsHamiltonian initializationFeature updates and tensor outputsFactorized tensor updatesNumerical equivariance and implementation testsParallel transport for multinuclear systemsData preprocessing and trainingDatasets and averagingModels and optimizationCobalt tensor errors and anisotropy screeningDysprosium dynamics and additional resultsPropagation and coordinate conventionsPrediction errors and propagation diagnosticsRelaxation calculations and sensitivityTransfer from water to methanolReferencesWorksPhysics and Astronomy
SU(2)-Equivariant Spinor Message Passing for Molecular Magnetism
Abstract
Predicting molecular magnetism often requires electronic-structure calculations with spin-orbit coupling (SOC), which are costly to repeat across many molecules. The underlying electronic states can transform as spinors, yet equivariant surrogates typically use only integer-rank features. Inspired by the spinor nature of these electronic states, we develop eSpinor, an SU(2)-equivariant graph neural network that propagates spinor features. We initialize spinor features from eigenvectors of an approximate SOC Hamiltonian and use gauge-covariant message passing with a gauge-invariant readout to make predictions independent of the arbitrary eigenbasis choice. On cobalt single-molecule magnet candidates, eSpinor improves zero-field-splitting predictions relative to NequIP, including for unseen ligands. For a dysprosium magnetic resonance imaging probe, eSpinor improves real electron-spin correlations over a Hamiltonian-informed baseline despite similar static crystal-field errors and fewer parameters, while accurately predicting low-field nuclear relaxation in water and after transfer to methanol without retraining. Overall, these results demonstrate that SU(2)-equivariant spinor message passing is a useful inductive bias for molecular magnetism. We also find that similar static tensor errors can yield different dynamical errors, motivating model selection based on accuracy for the intended physical observables.
Introduction
Molecular magnetism underlies single-molecule magnets (SMMs) [1, 2], molecular qubits for quantum information [3], and magnetic resonance imaging (MRI) probes [4]. Designing these systems requires predicting how ligand chemistry and geometry affect magnetic energy levels and dynamics. Relevant quantities include zero-field-splitting (ZFS) parameters in transition-metal complexes and crystal-field parameters in lanthanide ions [5, 6]. The ligand field and spin–orbit coupling (SOC), the relativistic interaction between an electron’s orbital and spin angular momenta, shape the magnetic anisotropy of these complexes [7]. Changes in molecular geometry can therefore influence both their magnetic energy levels and spin dynamics [8].
A common computational approach first uses complete active space self-consistent field (CASSCF) calculations to describe the electronic structure without SOC [9], due to the inherent multi-reference character in these systems. SOC is then included by mixing the resulting electronic states in a state-interaction calculation [10]. Unlike density functional theory (DFT), which is single-reference and scales polynomially with system size, CASSCF captures the near-degenerate multiplets of open-shell metal ions by including all electronic configurations within an active space. The number of these configurations grows combinatorially with the size of the active space, which restricts conventional calculations to small active spaces [11]. Averaging over many low-lying states adds further cost, so a single geometry can require hours of computation (Sec. V B), making repeated calculations for screening or dynamics expensive. Machine learning surrogates that predict magnetic parameters from geometry offer a cheaper alternative. For example, Gaussian-moment neural networks (GM-NN) have been extended to predict magnetic anisotropy tensors [12].
Other models incorporate spin explicitly. For magnetic materials, SpinGNN uses atomic positions and magnetic moment directions to learn spin–lattice potential-energy surfaces [13]. SpinGNN++ extends this approach to SOC-dependent interactions while enforcing spatial and time-reversal symmetries [14]. For open-shell molecules, AIMNet2-NSE predicts energies using neural spin-charge equilibration constrained by total charge and spin multiplicity [15]. OrbitAll learns from spin-polarized orbital matrices computed with semiempirical methods [16]. DeepH-E3 and Uni-HamGNN instead make spin part of the target, predicting electronic Hamiltonian matrices with SOC from atomic structure [17, 18].
Several of these models impose equivariance under , the group of proper rotations in three-dimensional space, so that predicted tensors rotate consistently with the molecular geometry [19]. This inductive bias can improve data efficiency and accuracy. In machine-learning interatomic potentials (MLIPs), nonscalar -equivariant features help even for rotation-invariant energies [20], showing that internal features can usefully carry richer rotation laws than the outputs. For magnetic-property prediction, eqSNAP provides an -equivariant extension of the spectral neighbor analysis potential (SNAP) method. It has been used to predict ZFS tensors and crystal-field parameters using local atomic descriptors and linear regression [21, 22].
SOC couples orbital motion and electron spin, motivating representations that describe these degrees of freedom jointly [7]. Although magnetic observables follow tensor rotation laws, the underlying electronic states can transform as spinors under [23] [24]. By analogy with MLIPs, such features could serve as internal features and benefit the predictions of integer-rank magnetic targets, yet the models discussed above use only integer-rank internal features. This motivates our research question, illustrated in Figure 1: Can SU(2)-equivariant spinor message passing provide a useful inductive bias for predicting SOC-dependent molecular properties?
Figure 1. Representation choices for predicting SOC-dependent molecular magnetic properties. eSpinor propagates half-integer spinor features alongside integer-rank tensor features. The ZFS levels are schematic, and the electron-spin correlation shows the 3 T CASSCF+SOC reference [8].
Inspired by the spinor nature of spin–orbit-coupled electronic states, we develop eSpinor, an -equivariant graph neural network that propagates spinor features alongside integer-rank tensors. We initialize spinors from eigenvectors of an approximate SOC Hamiltonian and classify the allowed interaction types to guarantee rotation equivariance and eigenbasis independence. To assess the proposed architecture, we compare eSpinor with the widely used SO(3)-equivariant architecture NequIP [20] and its Hamiltonian-informed variant NequIP+H, which serves as an ablation. On cobalt SMM candidates [25], eSpinor generalizes better to unseen ligands. On a dysprosium MRI probe [8], it predicts real electron-spin correlations more accurately at 8k training geometries, even though it matches NequIP+H on static crystal-field tensors at individual geometries. The dynamical advantage also extends from water to methanol without retraining. These results indicate that a model with SU(2)-equivariant spinor features improves magnetic-property prediction in the tested systems, even for SO(3)-equivariant targets. They also show that similar static tensor errors can lead to different dynamical errors. Model selection should therefore be guided by accuracy on the physical observables relevant to the intended application.
SU(2)-Equivariant Spinor Message Passing
Our first task is to construct spinor features that transform consistently when the molecular geometry rotates (§ 0.2 A). Their initialization introduces a freedom in choosing the eigenbasis, which constrains the network design (§ 0.2 B). We then design message-passing layers that respect these constraints (§ 0.2 C) and a readout that produces magnetic tensors independent of the eigenbasis choice (§ 0.2 D).
Figure 2(a) summarizes this construction. We represent the molecule as a graph with atomic numbers and positions . Edges connect atoms within a radial cutoff. Every atom carries integer-rank tensor features , initialized from atomic-species embeddings. We consider SOC only at the magnetic metal sites, where we initialize spinor features from an approximate Hamiltonian that combines SOC with the ligand field (§ 0.2 A). Thus and are two feature types attached to the same graph. Within each interaction layer, the convolution in Figure 2(b) updates both feature types. The coupling in Figure 2(c) then contracts spinors to supply additional integer tensors while retaining the spinors for the next layer. A final readout produces magnetic tensors independent of the basis chosen for the initial eigenvectors.
Figure 2. eSpinor architecture. (a) Atomic embeddings and Hamiltonian eigenvectors initialize integer tensors and spinors for repeated message passing and tensor readout. (b) Convolution uses Clebsch–Gordan tensor products (CGTP). (c) Spinor contractions update integer tensors. Here even/odd denotes integer/half-integer rank and . Only and are benchmark targets. The preliminary tensor convolution used by the learned AOM initializer is omitted.
Spinor initialization
To specify the two feature types, we use irreducible representations (irreps). We label integer-rank tensor features by and spinor features by , which transform under half-integer irreps of SU(2), the double cover of SO(3) [23] [26].
Integer-rank geometric features cannot generate spinors through tensor products [27]. We therefore introduce electron spin explicitly by using the single-electron spin–orbital space of a metal shell. The orbital and spin angular momenta and co-rotate under joint spatial rotations generated by . Combining the shell’s integer orbital angular momentum with electron spin gives the half-integer irreps [24]. To obtain spinor features that depend on the ligand environment and SOC, we construct an approximate Hamiltonian at each metal site ,
Here acts on orbital states and on spin. The SOC strength is , with and in units of [7] [28]. This zero-field, single-electron Hamiltonian initializes features for targets obtained from many-electron calculations.
Specifically, we choose to reflect the metal–ligand bonding in each benchmark. For Dy(III), the relatively shielded shell motivates a point-charge model based on ligand positions and effective charges [7] [29]. For Co(II), we compare point-charge initialization with a learned angular-overlap model (AOM) to account for metal–ligand covalency [30]. The initializer formulas and numerical parameters are given in the Supplemental Material [31] (Sec. S1 A).
Diagonalizing the Hamiltonian gives approximate electronic energy levels and their corresponding eigenvectors. We collect the orthonormal eigenvectors as columns of and their energy levels in , so the eigenproblem is
We express these eigenvectors in the coupled basis , where is the total angular momentum projection. At each , their components form spinors with Wigner rotation matrices , as derived in Appendix A. As shown in Figure 2(a), the energy levels supply invariant scalar features and the eigenvectors initialize spinor features. Using these eigenvectors requires accounting for their eigenbasis freedom, as explained in § 0.2 B.
Addressing eigenbasis gauge freedom
Diagonalization in Eq. (2) determines eigenstates up to unitary transformations within each eigenspace, giving an eigenbasis gauge freedom. In our zero-field, single-electron initializer, time-reversal symmetry produces degenerate Kramers pairs [7] [32, 33]. The eigenbasis of each energetically isolated pair is therefore defined up to , which includes individual phase choices and unitary mixing of the 2 eigenvectors. Predictions must be independent of this choice, a requirement also encountered in spectral graph learning [34, 35]. In Appendix A, we give the conditions for larger degenerate eigenspaces.
This gauge freedom also enters the transformation under molecular rotations. For a rotation and either lift , the eigenvectors obey
Here the spin–orbital rotation acts through on the angular index at each , while the gauge transformation mixes eigenvector columns within each pair. Following gauge-equivariant learning [36, 37], we preserve gauge covariance in the spinor features alongside rotation equivariance. The spinor features remain basis dependent, but contracting their eigenbasis index gives the gauge-invariant product . We project this product onto integer-rank tensors.
For initialization, we use Boltzmann-form weights , where is the energy averaged within the Kramers pair containing eigenstate . The learned energy scale controls how strongly lower-energy states are favored. This energy dependence allows contractions of the eigenvectors to retain information about the ligand environment. To preserve gauge covariance, the weights and invariant gates act identically on both columns of each Kramers pair and therefore commute with . Equivariant channel maps then give initial spinor features at layer that satisfy
In Appendix A, we derive this transformation law from the initialization map. The weighting formulas and gate inputs are given in the Supplemental Material [31]. Each subsequent layer must preserve Eq. (4), which guides the message passing in § 0.2 C.
Gauge-covariant message passing
Each interaction layer updates the features using information from neighboring atoms [38]. We choose NequIP as the backbone because it uses spherical harmonics to encode interatomic directions and explicit Clebsch–Gordan tensor products, which can be directly extended to half-integer representations [20]. The spinor updates must also preserve the eigenbasis gauge law in Eq. (4).
We classify the allowed real-bilinear couplings into 3 interaction types,
Here denotes elementwise complex conjugation, and each product includes Clebsch–Gordan projection onto allowed angular ranks. The product couples integer-rank features to produce integer-rank tensors. The mixed product updates spinor features without changing the eigenbasis index , so its output carries the same . In the conjugated product , contraction over cancels the unitary while retaining angular information in gauge-invariant integer tensors. We derive this classification and its covariance properties in Appendix A.
We use these operations to assemble the convolution in Figure 2(b). In the equations below, square brackets denote Clebsch–Gordan projection onto the output irreps, with angular-component and feature-channel indices omitted for clarity. Let collect the spherical harmonics of the direction from atom to atom . The sum below runs over neighbors . In our mononuclear benchmarks, only the metal carries spinor features, so ligand neighbors condition an update of that metal’s own . We combine these neighbor contributions with an on-site coupling to in a gated residual update,
Here the equivariant channel maps are absorbed into the neighbor weights and the on-site coupling . The neighbor weights depend on distance and invariant scalar features at both atoms and include neighbor normalization. The gate uses invariant pair norms and scalar features. Its multiplier is shared over the angular components at each and both columns of each Kramers pair, so it commutes with rotations and .
The integer-rank features are fed into an equivariant tensor convolution, giving . The updated spinors then supply a tensor correction through conjugated contractions and angular projection, as shown in Figure 2(c). With an invariant channel gate , the integer update is
Here denotes channelwise multiplication. Atoms without spinors receive only . Both and enter the next layer, so contraction informs the integer features without discarding the spinor features. The contraction formulas, gate inputs, and construction of are detailed in the Supplemental Material [31] (Sec. S1 B).
The channel maps act identically on all eigenbasis columns. Together with the angular projections and invariant gates shared within each Kramers pair, this guarantees the spinor transformation law in Eq. (4) at every layer. The contraction over in Eq. (5) removes the eigenbasis dependence, and angular projection gives gauge-invariant integer-rank tensors with the required rotation laws, as proved in Appendix A.
Tensor readout
To predict magnetic tensors, the readout combines the final integer-rank features with gauge-invariant spinor contractions. After interaction layers, we omit the layer index from and and write the metal-site readout in Figure 2(a) as
The equivariant map combines channels of rank , while uses conjugated spinor contractions projected onto the same rank. Under the initialization assumptions in § 0.2 A and II B, including separated Kramers pairs, both contributions are gauge-invariant tensors of rank . This guarantees that the prediction is independent of the eigenbasis and obeys the required SO(3) rotation law,
We prove this result in Appendix A and report numerical equivariance tests in the Supplemental Material [31] (Sec. S1 D).
For the Co(II) benchmark, we convert the rank-2 output to the symmetric, traceless Cartesian ZFS tensor . For Dy(III), we convert outputs of ranks 2, 4, and 6 to Stevens coefficients using the fixed basis maps given in the Supplemental Material [31]. The physical definitions of these targets follow in Secs. III and IV.
Propagating spinors before contraction also introduces coefficient sharing. In a product , the learned update of determines both factors, linking the coefficients of the resulting tensor interactions, as derived in the Supplemental Material [31] (Sec. S1 C). To assess how the architecture uses electronic information, we compare eSpinor with NequIP and its Hamiltonian-informed ablation NequIP+H. NequIP constructs integer-rank features from atomic species and geometry. NequIP+H also injects gauge-invariant contractions of Hamiltonian eigenvectors at each layer and propagates the resulting integer features. eSpinor additionally updates and propagates spinor features before contraction. The Hamiltonian-informed models use comparable parameter budgets. All models are trained and selected using static tensor errors. Their implementation and training procedures are specified in the Supplemental Material [31] (§ 3.2).
Zero-field splitting in cobalt complexes
Zero-field splitting (ZFS) is the separation of spin energy levels without an applied magnetic field. In Co(II) complexes, SOC couples the spin to the ligand-dependent orbital structure, producing magnetic anisotropy [5] [25]. Easy-axis anisotropy favors opposite orientations of the magnetic moment along a molecular axis and can hinder transitions between them, supporting the slow magnetic relaxation sought in SMMs [1]. Changing the ligands can alter both the magnitude and character of this anisotropy. Predicting ZFS across ligand environments therefore helps identify promising molecular magnets.
COMPASS was developed to identify Co(II) complexes with large magnetic anisotropy for SMM design [25, 39]. We use its 4-coordinate complexes with CASSCF+SOC reference targets to test ZFS prediction and generalization to new ligands. All models share 6892 training, 862 validation, 862 in-distribution test, and 473 out-of-distribution (OOD) test structures. The OOD complexes contain only ligands absent from training. Data preprocessing and split construction are specified in the Supplemental Material [31] (§ 3.2 A).
| Model | Parameters | D tensor | D tensor | ||||
| eqSNAP | 1,870 | 3.674 | 16.832 | 1.729 | 2.989 | 12.870 | 1.670 |
| GM-NN | 943,874 | 2.380 | 12.090 | 1.417 | 2.988 | 14.710 | 1.666 |
| NequIP | 122,120 | 1.193 | 6.645 | 0.835 | 1.651 | 7.611 | 1.134 |
| NequIP+H | 144,923 | 1.085 | 5.925 | 0.825 | 1.626 | 8.619 | 1.211 |
| eSpinor (pc) | 149,332 | 1.004 | 5.505 | 0.709 | 1.370 | 6.893 | 1.061 |
| eSpinor (AOM) | 150,679 | 0.849 | 4.838 | 0.641 | 1.264 | 6.365 | 0.875 |
Table 1. Cobalt prediction MAEs (cm) on the splits in Sec. III. The tensor error averages all 9 Cartesian entries. Parameter counts are ridge coefficients for eqSNAP or trainable neural parameters. The eSpinor variants use point-charge (pc) or angular-overlap (AOM) initialization. In distribution () and unseen ligands (). *Lowest errors are bold and second lowest are underlined throughout the result tables.
NequIP+H and eSpinor use AOM initialization, with point-charge eSpinor as an additional comparison. We also evaluate eqSNAP, which uses linear regression on equivariant neighbor-density descriptors [21], and GM-NN, which learns invariant coefficients of geometric tensor features [12].
Effective spin Hamiltonian
The reference calculations represent the lowest 2 Kramers pairs by an effective spin [25, 40]. We predict the real, symmetric, traceless tensor in the ZFS Hamiltonian
where are dimensionless spin operators. The tensor’s 5 independent components specify the strength and orientation of the anisotropy.
In the principal-axis frame of , describes axial anisotropy and measures the difference in anisotropy between the and directions [5]. For , negative gives easy-axis anisotropy, whereas positive gives easy-plane anisotropy. Both and matter for assessing SMM candidates [25], so we evaluate their errors alongside the full tensor error, which is also sensitive to the orientation of the principal axes. We specify the principal-axis convention and MAE definitions in the Supplemental Material [31] (Sec. S3).
Prediction accuracy and OOD ligand generalization
eSpinor (AOM) gives the lowest tensor, , and errors on both test sets (Table I). Relative to NequIP, it reduces the tensor MAE by 29% in distribution and 23% for unseen ligands. Relative to NequIP+H, the reduction is approximately 22% on both tests. NequIP+H lowers the in-distribution tensor error relative to NequIP but offers little improvement for unseen ligands, where its and errors exceed NequIP’s. Thus, NequIP+H’s use of Hamiltonian-derived integer features does not consistently improve ligand generalization over NequIP in this comparison.
Within eSpinor, AOM lowers all reported errors compared with point-charge initialization. This gain is consistent with AOM’s description of directional metal–ligand orbital interactions, which change ligand-field splittings and hence the anisotropy produced by SOC [25, 30]. However, point-charge eSpinor also gives lower errors than AOM-initialized NequIP+H for all 3 metrics on both test sets. The advantage therefore cannot be attributed to AOM alone.
To assess the relevance for screening, we compare candidate rankings and individual prediction errors for NequIP, NequIP+H, and eSpinor (AOM). On unseen ligands, eSpinor improves the ordering of complexes by , raising the Spearman rank correlation between predicted and reference values from 0.757 for NequIP and 0.675 for NequIP+H to 0.787. It also lowers the 95th percentile of the per-complex tensor error from 4.73 and 5.88 to 3.41 cm, respectively. The improvement extends to overall candidate ordering and larger errors. For shorts ranked by most negative , the model with the largest overlap with the reference shortlist depends on how many complexes are selected. These diagnostics and the comparisons in Figure S2 are presented in the Supplemental Material [31].
Figure S2. Cobalt screening for the neural models. eSpinor uses AOM. (a) OOD fraction exceeding tensor-error threshold , on an axis linear up to and logarithmic above. (b), (c) Shortlist recovery for 862 in-distribution and 473 OOD complexes.
These comparisons support ligand generalization within 4-coordinate Co(II) complexes. Transfer to other ions or coordination numbers and prediction of SMM relaxation times require further testing.
Crystal-field learning and spin dynamics in a dysprosium complex
We next consider the Dy(III) complex [DyL] [1], a parametrically shifted (PARASHIFT) MRI probe [8]. In this type of probe, magnetic anisotropy shifts the nuclear magnetic resonances of protons on the ligands away from the water and lipid background, allowing the probe’s signal to be detected separately [4]. Nuclear spin relaxation governs the recovery of magnetization and the width of these resonances, linking probe sensitivity to electron-spin dynamics and molecular motion [8].
In solution, molecular motion changes the ligand field and drives fluctuations of the electronic magnetic moment that contribute to ligand-proton relaxation. The original PARASHIFT study [8] combined molecular dynamics with CASSCF+SOC calculations along a trajectory to resolve non-exponential electron-spin dynamics and predict longitudinal nuclear relaxation rates in agreement with experiment. This system therefore provides a test of whether magnetic tensors learned at individual geometries also reproduce spin dynamics and nuclear relaxation.
The released PARASHIFT data provide reference crystal-field coefficients at 1 fs intervals [41]. We use nested training sets of 500, 1k, 2k, 4k, and 8k geometries, with a common 1k-frame validation set and 39k-frame test trajectory. These sets occupy separate time intervals to avoid interleaving adjacent frames. We compare NequIP, NequIP+H, eSpinor, and eqSNAP, using point-charge initialization for the Hamiltonian-informed models. Preprocessing and exact splits are given in the Supplemental Material [31] (Sec. S2 A). We evaluate crystal-field coefficients, electron-spin correlations, and nuclear relaxation rates against the CASSCF+SOC reference, first in water and then after transfer to methanol without retraining.
Hamiltonian and spin propagation
The crystal field splits and mixes the 16 states of the Dy(III) ground multiplet with total angular momentum . Within this space, the Stevens-operator convention [6] gives
The prediction targets are the 27 real coefficients , expressed in cm with Stevens factors absorbed. The operators are dimensionless.
Along the molecular trajectory, the changing coefficients drive spin evolution under , where is the Zeeman interaction with the applied field. Replacing the reference coefficients by predictions changes both the electronic energy levels and their evolution. We derive the propagation and relaxation equations in Appendix B and specify their numerical evaluation in the Supplemental Material [31] (Sec. S4 A).
Crystal-field prediction
Figure S3 of the Supplemental Material [31] compares the coefficient MAEs, averaged over components within each rank and over test frames. The neural models outperform the tested eqSNAP configuration, with similar errors at 8k and no consistent neural-model ranking across ranks and training-set sizes.
Similar static tensor errors can nevertheless produce different dynamics. The coefficient errors enter through different Stevens operators and act throughout the propagation described in Appendix B, where their signs and temporal ordering matter. Rankwise MAEs average away this information. We therefore compare the propagated observables directly.
Electron-spin correlations
The longitudinal electron correlation measures how much angular momentum along the field remains correlated after a lag . With and propagator , we compute
Here is the initial thermal state at 300 K, and the brackets average starting times and field orientations. We compare the real curves through relative errors over 0–499 fs, without fitted tails. Errors are averaged equally over and 20 T (see the Supplemental Material [31], Sec. S4 B).
eSpinor gives the lowest mean real-correlation error from 1k through 8k, whereas NequIP+H leads at 500 samples Figure 4(a). At 8k, eSpinor reduces the error from NequIP+H’s 1.140% to 0.553% (Table II), although neither model is trained on dynamics. Figure 3 resolves this difference through signed residuals. The advantage persists after normalizing each curve by its initial value and across separate trajectory segments, so it extends to the decay shape (see the Supplemental Material [31]).
| Model | Parameters | Re 1–20 T | Complex 1–20 T | 1–3 T | 1–3 T |
| eqSNAP | 1,650 | 10.935 | 11.735 | 22.687 | 22.386 |
| NequIP | 122,080 | 1.882 | 2.121 | 3.448 | 3.361 |
| NequIP+H | 159,116 | 1.140 | 1.246 | 2.294 | 2.220 |
| eSpinor | 134,459 | 0.553 | 0.748 | 0.809 | 0.774 |
Table II. Dy downstream errors (%) relative to CASSCF+SOC for 8k training samples and 39k test frames. Metrics are defined in Secs. IV C and IV D.
Figure 3. Electron-spin correlations from neural models trained on 8k samples. (a) Raw Re at 3 T, with CASSCF+SOC as reference. (b)–(d) Signed model-minus-reference residuals at 3, 10, and 20 T on a common scale. Dotted lines mark zero error, and shaded bands show pointwise standard deviations. Curves retain lags 0–499 fs without a fitted tail. Results for eqSNAP are included in Table II and in Figure S3 of the Supplemental Material [31].
Figure 4. Downstream learning curves for NequIP, NequIP+H, and eSpinor. (a) Real electron-correlation relative error averaged over all 6 fields. (b), (c) Total nuclear and errors over H3/H4/H5 and 1, 1.5, 2, and 3 T. All errors are relative to CASSCF+SOC on the common 39k test frames. Shaded bands show standard deviations. Figure S3 of the Supplemental Material [31] includes eqSNAP.
The advantage also survives using the reference initial thermal state for every model at 3 and 20 T. Individual-window and orientation errors are comparable to NequIP+H’s, but eSpinor has a smaller ensemble-mean signed error, indicating less systematic error across the sampled windows and orientations. We define this diagnostic in the Supplemental Material [31] (Table S2).
The mean real-correlation error increases from 4k to 8k for all 3 neural models, even as coefficient MAEs decrease. Training and checkpoint selection optimize tensor accuracy, while propagated correlations also depend on how coefficient errors combine across tensor components and time. Identifying which changes cause the observed increase requires further investigation. Imaginary correlations and fitted decay parameters also give different rankings (see the Supplemental Material [31]). These results reinforce the need to evaluate the intended physical observable directly.
Nuclear relaxation rates
Nuclear relaxation connects spin dynamics to MRI signal recovery and linewidth. The longitudinal rate describes recovery of magnetization along the field, while describes transverse signal decay [8].
The dipolar contribution combines the full electronic correlation tensor with the changing Dy–proton geometry through the correlation for proton . Molecular tumbling also modulates susceptibility-dependent shielding, producing Curie relaxation. Following PARASHIFT [8], we combine these mechanisms as
We derive the correlation integrals in Appendix B. The fitted tails and ground-multiplet approximation used for model susceptibilities are detailed in the Supplemental Material [31] (Sec. S4 C).
We report mean absolute relative errors in the proton-group rates for H3/H4/H5 at 1, 1.5, 2, and 3 T using the definitions in the Supplemental Material [31]. We also compare
eSpinor gives the lowest mean real-correlation error from 1k through 8k, whereas NequIP+H leads at 500 samples Figure 4(a). At 8k, eSpinor reduces the error from NequIP+H’s 1.140% to 0.553% (Table II), although neither model is trained on dynamics. Figure 3 resolves this difference through signed residuals. The advantage persists after normalizing each curve by its initial value and across separate trajectory segments, so it extends to the decay shape (see the Supplemental Material [31]). The advantage also survives using the reference initial thermal state for every model at 3 and 20 T. Individual-window and orientation errors are comparable to NequIP+H’s, but eSpinor has a smaller ensemble-mean signed error, indicating less systematic error across the sampled windows and orientations. We define this diagnostic in the Supplemental Material [31] (Table S2). The mean real-correlation error increases from 4k to 8k for all 3 neural models, even as coefficient MAEs decrease. Training and checkpoint selection optimize tensor accuracy, while propagated correlations also depend on how coefficient errors combine across tensor components and time. Identifying which changes cause the observed increase requires further investigation. Imaginary correlations and fitted decay parameters also give different rankings (see the Supplemental Material [31]). These results reinforce the need to evaluate the intended physical observable directly.
before fitting and integration through relative errors averaged over these groups and all 6 fields.
eSpinor gives the lowest mean and errors at 4k and 8k (Table II). At 8k, the absolute rate MAEs are 0.51 and 0.49 s, compared with 1.40 and 1.38 s for NequIP+H. Its advantage also holds for and the dipolar rates, connecting the improvement in spin dynamics to the relaxation mechanism that dominates in this field range. The ranking changes at smaller training sets, where eqSNAP leads at 500 and NequIP+H at 1k and 2k (Figure S3 and Figure S4 in the Supplemental Material [31]).
Figure S3. Water learning curves on 39k test frames. (a)–(c) Crystal-field MAEs. (d), (e) Real electron and complex proton correlations over all 6 fields. (f) Total errors over 1–3 T. Proton quantities average H3/H4/H5. gives the same model ordering at every budget.
Figure S4. Field and mechanism comparisons at 8k. (a) Total errors over H3/H4/H5 at each field. (b) Dipolar and Curie errors over 1–3 T, normalized by their own reference mechanism. (c) Susceptibility errors from Eq. (50).
At 20 T, the susceptibility-dependent Curie mechanism supplies about 93% of the reference , and NequIP gives the lowest total-rate error, as shown in Figure S4 of the Supplemental Material [31]. The change in ranking follows the shift from dipolar to Curie relaxation, which probes a different electronic response.
Correlation and rate comparisons remain complementary because integral and fitted-tail errors can compensate. In the Supplemental Material [31], we examine this cancellation and show that varying the tail cutoff and reference susceptibility treatment preserves the 8k low-field ranking.
Transfer from water to methanol
We evaluate the 8k water-trained models on the methanol (MeOH) trajectory without retraining [8]. Inputs include nearby solvent atoms, so this probes transfer to the geometries and local environments sampled in another solvent (see the Supplemental Material [31], Sec. S4D).
Table III indicates that eSpinor achieves the lowest electron-spin and proton-correlation errors, while both nuclear-rate errors stay under 1%. This shows that eSpinor’s dynamical benefit carries over to a different solvent trajectory, despite being trained only on data from water. In contrast, susceptibility errors for are comparable across the neural models.
| Model | 1–20 T | Complex 1–20 T | 1–3 T | 1–3 T | |
| eqSNAP | 9.289 | 8.402 | 11.304 | 11.219 | 19.851 |
| NequIP | 2.427 | 2.427 | 3.209 | 3.186 | 5.423 |
| NequIP+H | 2.110 | 1.982 | 2.296 | 2.281 | 5.221 |
| eSpinor | 1.240 | 0.979 | 0.959 | 0.954 | 5.263 |
Table 3. Transfer errors (%) relative to the MeOH CASSCF+SOC reference using models trained on 8k water geometries. The trajectory and metrics are specified in the Supplemental Material [31].
For 8k eSpinor models, exporting 39k predictions averaged about 12 min on 1 NVIDIA H200 GPU, or 18 ms per geometry including setup and I/O. For context, 2 approximate CASSCF+SOC timing jobs took 1.4–2.3 h per geometry on 72 NVIDIA Grace CPU cores on DeltaAI. The Supplemental Material [31] specifies their different electronic-structure settings and the timing scope. This low inference cost, together with the demonstrated water-to-MeOH transfer, could enable longer and more densely sampled trajectories for spin-dynamics calculations using a limited reference set. Broader chemical benchmarks and experimental validation are needed to establish how this approach can guide MRI-probe design.
Discussion
OOD ligand generalization
The results in Sec. III show that eSpinor’s improvement extends to unseen ligands at fixed metal ion and coordination number. Better ordering of axial anisotropy and a lower 95th-percentile tensor error connect this gain to candidate screening. The model could prioritize unfamiliar 4-coordinate Co(II) complexes for electronic-structure calculations in searches for molecular magnets [42]. Prospective screening and transfer to other ions and coordination numbers require further testing.
The initializer comparison also has a physical interpretation. AOM describes directional metal–ligand interactions that alter orbital energies and the anisotropy produced by SOC [25, 30]. Its improvement over point-charge initialization is consistent with this role. Yet point-charge eSpinor also outperforms AOM-initialized NequIP+H. A suitable ligand-field description helps, and eSpinor’s advantage over NequIP+H persists with either initializer.
From coefficients to magnetic observables
For Dy, the clearest distinction appears in the dynamical consequences of the predicted tensors. At 8k, similar crystal-field MAEs lead to different electron-spin correlations and nuclear relaxation rates (Sec. IV). Propagation weights errors by their action on electronic states, and the correlation averages over molecular motion and field orientations. The diagnostic with a common initial thermal state identifies a smaller systematic residual for eSpinor in the sampled ensemble, despite comparable individual-sample errors. Similar static tensor errors can therefore yield different dynamical errors. Models should be assessed on the observables they are intended to predict.
The relevant accuracy criterion also changes with the relaxation mechanism. Dipolar rates depend on spin correlations and proton motion, whereas Curie rates depend quadratically on anisotropic shielding under molecular tumbling, as detailed in Appendix B and the Supplemental Material [31]. This explains why improved dipolar dynamics can coexist with a different high-field ranking, and why the full susceptibility error alone does not determine Curie-rate accuracy. Evaluating both correlations and integrated rates therefore gives a more complete assessment of a surrogate for relaxation calculations.
The MeOH test extends the comparison beyond the water trajectory without additional electronic-structure training data. The inputs include nearby solvent atoms, so this result supports transferring the learned mapping to geometries and local environments sampled under another solvent condition.
SU(2) inductive bias
NequIP+H and eSpinor both propagate Hamiltonian-derived information. NequIP+H carries it through integer-rank features, while eSpinor additionally propagates spinor features. Their comparison tests how the architectures use that information. On the water trajectory, the lower electron-correlation and low-field relaxation errors at 4k and 8k are achieved with 15.5% fewer parameters than NequIP+H (Table II). The gain does not require a larger model.
Spinor propagation constrains how tensor interactions are learned. When updated spinor features are contracted with their complex conjugates, the same learned transformation enters both factors, linking the coefficients of the resulting integer-rank interactions. Keeping the complete set of these products for each Kramers pair allows eSpinor’s updates to be rewritten entirely in integer-rank tensors, with identical predictions, as shown in the Supplemental Material [31]. This exact reformulation preserves eSpinor’s coefficient sharing, whereas NequIP+H uses different tensor updates. Our comparisons support eSpinor’s usefulness for ligand generalization and selected ensemble magnetic observables. A future comparison with integer-tensor updates that retain each Kramers pair’s information but learn the interaction coefficients independently would further test the contribution of coefficient sharing.
Conclusions
We investigated whether SU(2)-equivariant spinor message passing can provide a useful inductive bias for predicting molecular magnetic properties. To this end, we developed eSpinor to propagate Hamiltonian-derived spinors alongside integer-rank features, with gauge-covariant updates and a gauge-invariant tensor readout. For Co(II) complexes, eSpinor (AOM) lowers the zero-field-splitting tensor MAE relative to NequIP by 29% in distribution and 23% for unseen ligands. For the Dy(III) complex, eSpinor uses 15.5% fewer parameters than NequIP+H. At 4k and 8k water training geometries, it gives the lowest mean errors among the 4 models in real electron-spin correlations over the tested fields and total nuclear relaxation rates at 1–3 T. At 8k, this advantage persists despite coefficient accuracy similar to NequIP+H and extends to methanol without retraining. Together, these results support SU(2)-equivariant spinor message passing as a useful inductive bias for molecular magnetism in the evaluated systems.
However, the benefit depends on the observable, since susceptibility and some high-field relaxation rates favor integer-rank models. Similar static tensor errors can yield different dynami cal errors, so surrogates should be assessed on their intended physical observables. Our comparisons measure accuracy relative to the electronic-structure reference. Establishing the practical value of these gains for SMM screening or MRI-probe design requires further validation.
Our current benchmarks are limited to mononuclear complexes because we have not identified a suitable labeled multinuclear dataset. In fact, generating accurate reference data for these systems remains challenging, since electron correlation, SOC, and magnetic exchange between ions must be treated together [43]. Nevertheless, we have formulated parallel transport between metal sites that preserves covariance under independent local eigenbasis changes, as detailed in the Supplemental Material [31] (Sec. S1 E). Future work should test whether this extension improves predictions of magnetic interactions between ions.
Acknowledgments
A.A. acknowledges support in part from the Bren Endowed Chair, the Office of Naval Research through MURI grant N00014-18-1-2624, DARPA ExpMath (HR0011), and the AI2050 Senior Fellow program at Schmidt Sciences. E.W.S. acknowledges support through the Caltech Summer Undergraduate Research Fellowship. B.S.K. acknowledges support from the Pritzker AI+Science Fund, DARPA Biological Technologies (HR0011), and the Eddleman Graduate Fellowship. Y.L. acknowledges support from the Pritzker AI+Science Fund and the Wally Baer and Jeri Weiss Postdoctoral Fellowship. D.K. acknowledges support from the Pritzker AI+Science Fund and DARPA Biological Technologies (HR0011). Authors acknowledge the National Center for High-performance Computing (NCHC), Taiwan, for providing computing resources on Taiwania 2 and Nano 4. This work also used the Delta system at the National Center for Supercomputing Applications through allocation DMR160114 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.
Claude (Anthropic, Opus 5 and Fable 5) and ChatGPT (OpenAI, GPT-6 Astra) assisted with literature review and manuscript revision. Codex (OpenAI, GPT-5.6 Sol) assisted with writing and debugging data-processing, plotting, and unit-test scripts. The authors directed this assistance, reviewed the text and code, and take full responsibility for this work.
Data Availability
The original COMPASS and PARASHIFT datasets are publicly available from Zenodo [39] and the University of Manchester Figshare [41], respectively. The source code and preprocessing scripts will be made publicly available at AI-QChem/eSpinor upon publication.
Rotation equivariance and eigenbasis independence
We prove the rotation and eigenbasis transformation laws for the mononuclear architecture in Sec. II, assuming separated Kramers pairs. Linear maps and Clebsch–Gordan couplings are equivariant. Weights and gate values are invariant under rotations and eigenbasis changes. Gate multipliers are shared over all angular components within each irrep and both columns within each Kramers pair. Other column operations must also commute with every allowed . Below, denotes a molecular rotation and either lift of .
We first establish the Hamiltonian covariance needed for spinor initialization in Sec. II A. Both the point-charge and AOM initializers use the angular expansion
At a given metal site, points toward ligand , and is a rank- tensor operator acting on the orbital states of the shell [27]. The sums include and . The real scalar strengths come from the point-charge or AOM construction and are unchanged by molecular rotations. The spherical harmonics and orbital tensor operators transform with the same rank- rotation matrices. Their matching components in Eq. (14) therefore give . The strengths and operator conventions are specified in the Supplemental Material [31].
In the coupled basis , with , the joint spin–orbital rotation is [24]. Since is unchanged under this joint rotation, the full Hamiltonian obeys
For this fixed rotation, combining Eq. (15) with Eq. (2) gives
Thus is an eigenbasis at the rotated geometry with the same energy levels. A separate diagonalization with the same energy ordering can return another orthonormal basis, related by a unitary satisfying . The ligand field and are unchanged under time reversal, while the single-electron time-reversal operator obeys . These properties guarantee the Kramers pairs [32, 33]. When distinct pairs are separated in energy, with . Thus , giving the combined rotation and eigenbasis change in Eq. (3).
If eigenstates have exactly the same energy, instead permits arbitrary mixing within that eigenspace. For 2 coincident Kramers pairs, the freedom is therefore rather than . The proof extends only when all gate inputs are invariant under this larger group and column operations commute with it. In particular, weights and diagonal gate multipliers must be equal across all columns. Small nonzero gaps do not enlarge the gauge group, but can make eigenvectors sensitive to changes in geometry.
We next construct the weighted initialization used in Sec. II B. With uniform weights, the complete eigenbasis gives , which contains no geometry-dependent multipoles. Energy-dependent weights allow this contraction to retain information about the ligand environment. Using the weights defined in the Supplemental Material [31], write and . Rotations preserve the energy levels and therefore leave unchanged. Equal weights within each pair imply .
The invariant gate also commutes with . Initialization is . The channel map acts at fixed , identically on angular indices and eigenbasis columns . It therefore commutes with right multiplication by and obeys , where rotates all output channels. Consequently,
The prime denotes features at the rotated geometry, including their possible change of eigenbasis. For each rank , this is the feature law in Eq. (4). Gauge covariance requires equal weights within each pair, but does not require the Boltzmann form.
The message-passing operations in § 0.2 C must preserve Eq. (4). The brackets in that section denote Clebsch–Gordan projection. For the mixed product in Eq. (5), let project a rank- spinor and a rank- integer tensor onto output rank . For one pair of feature channels, its components are
Here is a Clebsch–Gordan coefficient, and in steps of 1 [27]. The projection couples the angular indices and while leaving the eigenbasis index unchanged. Taking gives the neighbor coupling in Eq. (6a). For and , its intertwining property gives
Invariant scalar weights, equivariant channel maps, gates shared within each Kramers pair, and residual addition preserve this law. Integer-tensor products follow the same angular rule without an eigenbasis index.
To form gauge-invariant integer tensors as in Eq. (5), let and be spinor features whose eigenbasis columns transform with the same . Write and for their rotation matrices. Their transformations and give
The gauge cancels because is unitary. Products of half-integer representations and their conjugates decompose into integer ranks, so angular projections of give gauge-invariant integer tensors.
For these feature transformation laws, every real-bilinear equivariant map is a linear combination of Clebsch–Gordan projections with channel coefficients [27]. Angular momentum addition gives integer outputs for equal integer or half-integer types and half-integer outputs for mixed types. Under a common gauge phase, and , whereas is unchanged. Thus a mixed product with the same gauge law as uses , and a gauge-invariant spinor contraction must be conjugated. For a bilinear column contraction between Kramers pairs and , with , gauge invariance further requires
Independent pair phases force for . For , Schur’s lemma gives . The allowed contractions are therefore angular projections of , including scalar weights for each pair. Together with integer-tensor and mixed products, these give the 3 interaction types in Eq. (5).
Starting from Eq. (17), induction over the interaction layers establishes spinor gauge covariance and rotation equivariance, together with gauge invariance and rotation equivariance of the integer features. The readout in § 0.2 D uses these integer tensors and equivariant channel maps, so its predictions obey Eq. (9). Changing the lift of a molecular rotation negates each half-integer factor, and the 2 signs cancel in every contraction. Predictions are consequently independent of both the initial eigenbasis and the choice of lift.
From spin propagation to nuclear relaxation
We follow the time-domain treatment of PARASHIFT [8]. The Hamiltonian in Sec. IV A includes the Zeeman term , with and in . This determines the propagator through
Here is in and time is in seconds because the Hamiltonian is expressed in . Each window starts from its Hamiltonian’s thermal state . The electronic correlation before averaging is , where . Its longitudinal projection gives Eq. (12).
To obtain nuclear relaxation, we include the dipolar interaction between the electronic magnetic moment and proton . For a Dy–proton vector , its geometry tensor is . Contracting it with the electronic correlation gives
All 3 indices run over Cartesian components. The contraction precedes the trajectory and orientation average, retaining the joint fluctuations of spin and proton geometry. With distances in metres, has units . No product of mean moments is subtracted, following the source calculation.
The explicit correlation is extended by a real fitted tail, denoted . Rotational damping with solvent-dependent correlation time then gives the spectral density
At the proton Larmor frequency , the dipolar rates are
where , is the vacuum permeability, and is the proton gyromagnetic ratio. Thus rate accuracy depends on an integrated correlation, including its extrapolated tail, rather than on the explicit curve error alone.
The Curie contribution in Eq. (13) instead uses the mean shielding tensor in the molecular frame. Here is the volume susceptibility per molecule. The shielding anisotropy determines the Curie rates under molecular tumbling. The susceptibility and Curie formulas, tail fitting procedure, and sensitivity checks are given in the Supplemental Material [31].
References
Supplemental Material
Implementation and symmetry details
Hamiltonian initialization
For Co(II), , and for Dy(III), . With metal–ligand vectors , the Condon–Shortley convention [1] gives
Here are orbital projections. The rotation laws are derived in Appendix A of the main text.
Point-charge strengths are , with . We use non-hydrogen neighbors within and Gasteiger charges [2]. Failed or zero assignments use for (N, O, S, F, Cl, Br, I, P, C, Se), in units of elementary charge. The PyCrystalField-based constants [3] are
| Co | not used | ||||
| Dy |
Table S2.
Both Hamiltonian-informed models use these prefactors in and distances in . The Dy values omit Bohr-to-\AA{} conversion and shielding, so they define the feature map rather than calibrated radial moments.
The angular-overlap model learns ligand-axis orbital energies [4]. With and , its strengths satisfy
For Co, denotes bonding. The MLP uses 8 element-embedding components, a width-32 SiLU layer, and first-convolution scalars. Zero final weights and inverse-softplus biases initialize , with a floor.
After adding SOC and diagonalizing in the coupled basis of Appendix A, we use
The learned scale is not a physical temperature. We subtract the lowest energy and floor at and at without renormalizing. The gauge conditions are derived in Appendix A.
Eigenvector backpropagation replaces inverse energy gaps by
The resulting Hamiltonian gradient is Hermitized. Only gradients are regularized. The eigensolver-failure fallback raises precision and adds a diagonal ramp. The covariance proof and numerical checks use unperturbed Hamiltonians.
Feature updates and tensor outputs
A fixed unitary map converts real features to the complex spherical basis, , with [5]. We use the spinor messages derived in Appendix A, divided by the square root of the receiving site’s neighbor count. Integer convolutions use a self-connection and Table S1’s normalization. The conjugated contraction is
| Setting | Co | Dy |
| Convolutions / maximum integer rank / channels per rank | 3 / 6 / 4 | |
| Radial basis size / cutoff order | 8 / 6 | |
| NequIP+H initializer | AOM | Point charge |
| eSpinor initializer | AOM or point charge | Point charge |
| Graph cutoff (Å) / average-neighbor normalization | 6 / 10 | 4 / 6 |
| Integer radial MLP depth / width | 3 / 64 | 1 / 128 |
| eSpinor radial hidden width | 64 | 4 |
| Batch size | 32 | 64 |
| Initial (cm) | 200 | 7500 (eSpinor), 200 (NequIP+H) |
| Peak / final learning rate | ||
| Warmup / maximum epochs / stopping patience | 10 / 300 / 100 | |
| Weight decay / gradient norm limit / EMA decay |
Table S1. Settings for NequIP, NequIP+H, and eSpinor, fixed across Dy training budgets.
Its real and imaginary parts in coordinates supply separate real tensor channels.
All gates use SiLU MLPs of hidden width 16, outputs clamped to , and initial multipliers of 1. For pair energies and , initialization uses
Here is the pair mean, (Co) or (Dy), and std centers and scales by the population standard deviation. Norms and denominators are floored at . Gating precedes a near-identity channel map. A width-16 MLP maps and to integer scalars. Increment gates add the invariants in Sec. S1 C, with pair-shared multipliers at fixed .
Before contraction, angular expansion increases tensor-channel multiplicities,
The embedding preserves the original spinors, which alone propagate between layers. The gate depends on integer scalars and is independent of the eigenbasis column. Each interaction contracts 4 learned pairs of , with and output rank . Readout uses all allowed products with and the contraction .
Co outputs use the orthonormal symmetric, traceless Cartesian basis of e3nn [5], . For Dy, let , so and . Equal-norm spherical operators use and the Condon–Shortley convention [1]. For , fixing the phase by , with , for unit-norm harmonics gives
All Dy neural models use this map and raw Stevens labels, including Stevens factors.
Factorized tensor updates
In exact arithmetic, before learned gates and numerical floors, Eq. (28) gives
This has the form of a thermal density matrix, with its factor and the auxiliary index of a purification [6]. The analogy is formal because is approximate and is learned. The later feature Gram matrix has no imposed thermal form.
For a Kramers pair, collect angular components and channels into and define . These matrices determine up to and decompose into integer tensors. Gate inputs follow from and , where projects onto angular momentum . At fixed geometry and , write and let contain the pair-shared gates. The residual update gives
Increment gates use in these identities, so is input dependent. Writing expansion as gives and . Learned channel maps give contractions . Thus complete pair matrices, including angular and channel cross terms, reproduce the predictions. At fixed gates, links tensor-interaction coefficients as . NequIP+H uses different integer-feature updates (§ 3.2 B).
Numerical equivariance and implementation tests
Residuals use . Each shell uses 3 distorted 6-ligand geometries with synthetic Kramers-paired eigenvectors, 3 rotation angles, 3 within-pair gauge changes, and a sign flip. Networks have 3 layers, 4 channels through rank 6, and 4 contraction pairs, with nonzero spinor contributions verified. Maximum float32/complex64 residuals are for rotation, for gauge, and for sign. Outputs cover ranks for and for shells. Across 9 float64/complex128 initializer cases, covariance, paired-energy, and time-reversal residuals are at most . Spinor and matrix updates agree within . Output-conversion checks at and 4 use tolerances (float32) and (float64).
We also examine the trained Co initializers on 862 in-distribution and 473 ligand-OOD geometries. For pair energies , the minimum inter-pair gap is . Its smallest values are for AOM and for point-charge initialization. Smaller gaps show no systematic increase in tensor error in Figure S1.
Figure S1. Kramers-pair gaps and prediction continuity for the current Co eSpinor checkpoints with AOM (top) and point-charge (bottom) initialization. (a), (d) Cumulative gap distributions. (b), (e) Cartesian tensor MAE per complex. (c), (f) Median output changes over 24 geometries and 2 directions, taking the larger change from the 2 displacement signs. Shading spans the double-precision minimum and maximum.
For each initializer and split, we perturb the 6 smallest-gap geometries and 6 near the median in 2 random directions, keeping the metal fixed. Nonmetal RMS displacements range from to , with neighbors and ligand masks recomputed. At small steps with unchanged connectivity, double-precision output changes are proportional to displacement. Single precision reaches a numerical floor at the smallest steps. These checks concern local continuity at the sampled finite gaps.
Parallel transport for multinuclear systems
For same-shell metals, transport between independent eigenbases [7] uses
maps receiver to sender bases ( in common axes), and is invariant. For , and give covariance at the receiving site.
Data preprocessing and training
Datasets and averaging
COMPASS filtering retains 14573 of 15559 Set-1 records with complete ab initio ligand-field entries, including VLFT, and 206 ligand identifiers [8, 9]. These entries check completeness only. NumPy’s default_rng(42) shuffles sorted identifiers, reserving the first 41. Among 4-coordinate complexes, 8616 use only retained ligands and 473 use only reserved ligands (40 identities), forming the OOD set. Mixed complexes are excluded. Shuffling the 8616 structures from filtered dataset order with default_rng(43) gives 862 validation, 862 in-distribution test, and 6892 training structures, in that order.
PARASHIFT provides Dy CASSCF+SOC labels along the trajectory of the deuterated Gd analogue in DO [10, 11]. At 1 fs spacing, the zero-based frame intervals are
Preprocessing selects periodic images nearest the metal and retains atoms within 6 Å, including solvent. Coordinates and targets share laboratory axes. Dy neural fits use seeds 0, 1, and 42 on the fixed split, with unchanged settings across budgets. eqSNAP uses deterministic ridge regression. We average errors evaluated separately for each fit. Bands show sample standard deviations, faint points show individual fits, and larger markers show means. This spread measures training variability on the same trajectory.
Models and optimization
eqSNAP uses MolForge bispectra through and a 4 Å cutoff [12, 13]. For Dy, species-pooled features and rankwise ridge fits are
Weights are shared over . With 5 species and 55 bispectra, ranks (2, 4, 6) use 825 real weights, stored as 1650 real/imaginary entries. Dy fits raw Stevens units with at every budget, so effective regularization per sample decreases with . Co uses metal-centered descriptors and validation tensor RMSE to select from {0.01, 0.1, 1, 10, 100}.
Co GM-NN [14] uses widths (360, 512, 512, 1), 5 radial functions with 7 Gaussians each, an 8 Å cutoff, batch size 32, and 1000 maximum epochs. Learning rates decay linearly from 0.02. Training uses the implementation’s 5-component tensor loss, with validation tensor MAE selecting the checkpoint.
Graphs include hydrogens, with cutoffs in Table S1. NequIP+H and eSpinor fit their initializers separately, using reference magnetic parameters only as targets. At each layer, NequIP+H gates the weighted initial eigenvectors using current integer scalars and Eq. (31). The invariant diagonal gate has equal values within each Kramers pair. Its gauge-invariant projector sum and normalized channel contractions are
Maps , multiply each subspace by a learned complex scalar. Projection gives real/imaginary tensor channels without expansion. Each layer recomputes these tensors from and couples them to , which propagates the electronic information. Dy readout multiplicities at ranks (2, 4, 6) are (3, 3, 3) for NequIP+H and (43, 64, 73) for eSpinor. The point-charge eSpinor initializer precedes all 3 convolutions, while Co AOM follows the first convolution and leaves 2 spinor updates.
NequIP, NequIP+H, and eSpinor minimize the normalized tensor error
Here labels prediction heads, and are reference and predicted component vectors, and is the batch mean. The component count is for Co and 5, 9, and 13 for Dy ranks 2, 4, and 6. The scale is the training-target standard deviation over samples and components, floored at in target units. Dy uses fixed Bessel functions and a polynomial cutoff [15]. Linear warmup starts at 0.01 times the peak rate, followed by cosine decay. AdamW [16] excludes biases and 1-dimensional parameters from weight decay. We select the checkpoint with the lowest validation tensor loss, evaluated using exponential-moving-average weights.
Learned features use float32/complex64 in PyTorch [17], with higher-precision angular tables, e3nn [5], and cuEquivariance tensor products [18]. Archived Co models used a 32 GB NVIDIA V100 on Taiwania 2 through TWCC. Co eSpinor and NequIP+H used PyTorch 2.9.1+cu126, NequIP used 2.8.0+cu128, and eSpinor used cuEquivariance 0.8.0. Dy inference used an H200 and naive tensor-product kernels.
Cobalt tensor errors and anisotropy screening
We remove the trace from the released matrices, . This preserves level splittings. In principal axes, follows the largest absolute eigenvalue, , and [19]. We evaluate because exchanging reverses its sign. Per-complex tensor errors are . Table I averages , , and over complexes.
Spearman correlation compares predicted and reference ranks of . Shortlists contain their most negative values, with recovery for and . Figure S2 compares candidate rankings by axial anisotropy. Tensor errors avoid the axis-reassignment sensitivity of scalar near the rhombic limit.
Dysprosium dynamics and additional results
Propagation and coordinate conventions
Following Appendix B and PARASHIFT [10], water propagation uses 78 nonoverlapping 500-frame test windows, , 100 equally weighted Fibonacci field orientations, , and . Fields are , with low-field averages over . Each model uses
Here , , and time in the exponential is in seconds. Hermitian diagonalization uses float64/complex128, subtracting the lowest energy in the thermal exponential. Electron uses laboratory axes and a fixed field, with 499 steps and 500 lags.
Nuclear correlations and susceptibility use minimum-image, metal-centered coordinates. Orient donor-plane unit normals from the axial-N centroid toward the O centroid and define the molecular frame by
Hats normalize vectors. Water donor indices are O (1, 3, 5), axial N (721, 722, 723), and equatorial N (724, 725, 726), counted from 1. The same rotates coefficients and Dy–proton vectors, with field .
Exporting 39k predictions averaged 709.6 s for 8k eSpinor models on 1 H200 (batch size 8, naive tensor products), including setup and I/O but excluding training and propagation. This corresponds to 18.2 ms per geometry averaged over the export. For context, 2 approximate OpenMolcas CASSCF+SOC jobs [20] took 804 s per geometry and 5137 s per geometry, including input preparation and parameter extraction, on 72 NVIDIA Grace CPU cores on DeltaAI [21]. Each used 1 MPI process and 72 OpenMP threads, with no GPU computation. The setup used CAS(9,7), 18 equally weighted sextet roots, RASSI SOC, and ANO-RCC bases with VTZP on Dy, VDZP on O/N, and VDZ elsewhere. Compared with the benchmark reference protocol, it used placeholder solvent charges ( on O and on H) and VDZP on all O/N atoms instead of only donors. These jobs provide a cost scale for electronic-structure evaluation, not timings for generating the benchmark labels or a controlled speedup ratio.
Prediction errors and propagation diagnostics
For geometries and the 500 explicit correlation samples , without fitted tails, we use
Here selects real, imaginary, or complex correlations. Shape errors normalize each curve by its own . Errors average 6 fields equally and, for , H3/H4/H5. Figure S3 includes all budgets and models. The normalized real-correlation advantage persists at 1k–8k and in each 26-window segment starting at 51k, 64k, and 77k at 8k. At 8k, imaginary-correlation errors favor NequIP+H over eSpinor (1.264% versus 1.534%).
To isolate propagation from initial-state error, we combine the reference thermal state with each model’s propagator,
| Model | 3 T | 20 T | ||||
| NequIP | 0.7270 | 0.4184 | 0.1962 | 0.7009 | 0.3104 | 0.1859 |
| NequIP+H | 0.6929 | 0.3706 | 0.1129 | 0.6718 | 0.2701 | 0.1143 |
| eSpinor | 0.7195 | 0.3795 | 0.0559 | 0.6882 | 0.2419 | 0.0572 |
Table S2 (PDF p. 18). Dimensionless propagation RMS errors at 8k with the reference initial state. Columns follow Eq. (45).
Here , denote reference and model, with identical Zeeman terms. For windows , orientations , and arithmetic means,
Tests use all windows and orientations at 8k and 3/20 T. Unmixed runs reproduce saved correlations within . Table S2 shows comparable individual errors but a smaller ensemble-mean signed error for eSpinor than for NequIP+H. Here the mean residual measures the systematic error in the sampled ensemble, and .
Relaxation calculations and sensitivity
Each proton’s real is fitted separately with the PARASHIFT form [10],
For each , , we anchor the late branch at the observed . SciPy [22] minimizes , with and . We choose the converged fit with the smallest objective. Fitting characterizes electron decay only. At 8k, eSpinor leads accuracy and NequIP leads .
The explicit complex is retained through 499 fs and followed by the real fitted tail from 500 fs. We include the first tail sample below , with , then set the tail to zero. This cutoff follows our reconstruction of the archive. The spectral integral is evaluated with equal endpoint weights, , with time in seconds. This integral gives the dipolar rates derived in Appendix B. Proton-group correlations are used only for curve comparisons, so averaging does not precede the proton-resolved fits.
For the mean shielding defined in that appendix, let and be its symmetric and antisymmetric parts. The invariants and give
No empirical scale factor is applied. Reference Curie rates use archived CASSCF+SOC susceptibilities. Model susceptibilities use the zero-field Van Vleck expression within [23]. For energies , thermal weights , , and ,
Here gives cgs molar susceptibility, and is Avogadro’s constant. Errors against the archive include the ground-multiplet approximation.
Total rates are averaged within each proton environment before taking errors. For and ,
Component errors use their own reference mechanism in the denominator.
Susceptibility errors compare the trajectory and its mean,
Figure S4 compares field-dependent rates, their mechanisms, and susceptibility. To check cancellation, split each rate into explicit-integral, fitted-tail, and Curie contributions, . The signed errors then add exactly. For eqSNAP, explicit-integral and tail errors compensate at 500 samples but reinforce at 8k. With fits fixed, varying from to and using matched reference susceptibility gives 4 combinations. Mean rate errors change by less than 0.10 percentage points for every model, with unchanged 8k low-field ranking. Longer trajectories are needed to validate the extrapolated tails.
Transfer from water to methanol
Water-trained models transfer without retraining or selection on MeOH. Propagation uses the first 82,000 of 82,300 matched frames in 164 windows. Susceptibility uses all frames, with archived values for reference Curie rates. The water protocol is unchanged except ps and one-based donor indices O (1, 3, 5), axial N (697, 698, 699), and equatorial N (700, 701, 702).
Figure S5. Water-to-MeOH transfer across training budgets. (a), (b) Real electron-spin and complex proton-correlation errors over all 6 fields. (c) Total errors over H3/H4/H5 and 1–3 T. gives the same model ordering at every budget. Table III includes susceptibility and both rates at 8k.
References
References
- [1]D. A. Varshalovich, A. N. Moskalev, and V. K. Khersonskii, Quantum Theory of Angular Momentum (World Scientific, Singapore, 1988).
- [2]J. Gasteiger and M. Marsili, Tetrahedron 36, 3219 (1980).
- [3]A. Scheie, Journal of Applied Crystallography 54, 356 (2021).
- [4]C. E. Schäffer and C. K. Jørgensen, Molecular Physics 9, 401 (1965).
- [5]M. Geiger and T. Smidt, arXiv preprint arXiv:2207.09453 10.48550/arXiv.2207.09453 (2022).
- [6]T. Barthel, New Journal of Physics 15, 073010 (2013).
- [7]T. S. Cohen, M. Weiler, B. Kicanaoglu, and M. Welling, in International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, Vol. 97 (PMLR, 2019) pp. 1321–1330.
- [8]L. A. Mariano, V. H. A. Nguyen, V. Briganti, and A. Lunghi, Journal of the American Chemical Society 146, 34158 (2024).
- [9]A. L. Mariano, A. Lunghi, and V. H. A. Nguyen, COMPASS, Zenodo (2024), data set, version 1.
- [10]T. R. C. Thompson, B. Alnami, J. M. Skelton, and N. F. Chilton, Journal of the American Chemical Society 148, 19536 (2026).
- [11]N. Chilton, Research data for: Electron and Nuclear Spin Dynamics of a Lanthanide Complex in Solution, University of Manchester Figshare (2025), dataset.
- [12]V. H. A. Nguyen and A. Lunghi, Physical Review B 105, 165131 (2022).
- [13]V. Briganti and A. Lunghi, npj Computational Materials 11, 62 (2025).
- [14]V. Zaverkin, J. Netz, F. Zills, A. Köhn, and J. Kästner, Journal of Chemical Theory and Computation 18, 1 (2022).
- [15]J. Gasteiger, J. Groß, and S. Günnemann, in International Conference on Learning Representations (ICLR) (2020).
- [16]I. Loshchilov and F. Hutter, in International Conference on Learning Representations (2019).
- [17]A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, in Advances in Neural Information Processing Systems, Vol. 32 (2019) pp. 8026–8037.
- [18]NVIDIA Corporation, cuEquivariance: NVIDIA library for high-performance equivariant neural networks, Software, version 0.8.0 (2025).
- [19]R. Boča, Coordination Chemistry Reviews 248, 757 (2004).
- [20]I. Fedz. Galván, M. Vacher, A. Alavi, C. Angeli, F. Aquilante, J. Autschbach, et al., Journal of Chemical Theory and Computation 15, 5925 (2019).
- [21]B. Bode, G. Bauer, L. Herriott, V. Kindratenko, and W. Gropp, in Practice and Experience in Advanced Research Computing 2025: The Power of Collaboration, PEARC ’25 (Association for Computing Machinery, New York, 2025) pp. 1–4, article 53.
- [22]P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, et al., Nature Methods 17, 261 (2020).
- [23]J. H. Van Vleck, The Theory of Electric and Magnetic Susceptibilities (Clarendon Press, Oxford, 1932).