DSpinGNN: A Physics-Informed GNN for CrI3 Dynamics

Updated 5 July 2026

The paper introduces DSpinGNN, a bifurcated architecture that accurately predicts local magnetic exchange coupling in dynamic, strained CrI3 with near-DFT accuracy using an E(3)-equivariant GNN and a physics-informed Δ-MLP.
It leverages an analytical inductive bias via the Goodenough–Kanamori relationship to stabilize predictions under large distortions and ensure transferability from small cells to mesoscale simulations.
The model achieves high accuracy in energy and force metrics while enabling mesoscale mapping of transient magnetic textures induced by propagating biaxial strain waves.

DSpinGNN is a physics-informed, bifurcated machine-learning architecture introduced for dynamic magnetic exchange prediction in strain-deformed monolayer CrI $_3$ . It was designed to resolve the instantaneous, position-dependent isotropic magnetic exchange coupling $J_{ij}$ across a dynamically deforming crystal lattice while simultaneously modeling structural forces at length scales inaccessible to first-principles methods. The model combines an $E(3)$ -equivariant graph neural network for classical Langevin structural dynamics with a physics-informed $\Delta$ -MLP that maps instantaneous local Cr–I–Cr bond geometry to exchange couplings, embedding the Goodenough–Kanamori superexchange relationship as an analytical inductive bias. Trained on 345 DFT+U configurations and evaluated on a strictly withheld 61-configuration test set, it reports an energy MAE of $1.1$ meV/atom, a force MAE of $6.5$ meV/Å, and an exchange coupling MAE of $0.18$ meV with $R^2=0.91$ ; it is then deployed at $400\times$ scale in a 3,200-atom supercell to map exchange textures driven by a propagating biaxial strain wave (Balghari et al., 10 Jun 2026).

1. Problem definition and scientific scope

The central problem addressed by DSpinGNN is the prediction of instantaneous, local magnetic exchange in a crystal whose geometry is evolving in time. In monolayer CrI $_3$ , the relevant target is the isotropic first-nearest-neighbor exchange coupling $J_{ij}$ 0, which depends on local bond geometry and therefore changes under dynamical strain. The paper frames this as a multiscale bottleneck: first-principles methods can establish the underlying structure–exchange relationship, but not at the mesoscopic length and time scales required to follow a propagating strain wave through a large supercell (Balghari et al., 10 Jun 2026).

DSpinGNN therefore couples two tasks that are usually handled separately: structural dynamics and exchange prediction. Its design assumes that structural evolution can be learned from DFT+U data with an equivariant interatomic potential, while local exchange can be inferred from geometry using a hybrid of analytical superexchange physics and a data-driven residual corrector. A plausible implication is that the model is not merely an interatomic potential with an auxiliary magnetic head, but a deliberately split architecture in which each branch is matched to a different physical subproblem.

The resulting workflow targets mesoscale exchange mapping in strain-driven 2D magnetic materials. In the reported application, the model is used to study how a propagating biaxial strain wave produces spatially heterogeneous exchange coupling textures, including transient antiferromagnetic-sign regions when local compressive strain crosses a threshold established from DFT.

2. Bifurcated architecture

The structural branch is an $J_{ij}$ 1-equivariant GNN based on the NequIP framework. It takes as input atomic species labels $J_{ij}$ 2 and equivariant edge features constructed from interatomic vectors $J_{ij}$ 3. Message passing at layer $J_{ij}$ 4 is written as

$J_{ij}$ 5

followed by the node update

$J_{ij}$ 6

After three interaction layers with a cutoff of 7 Å, scalar outputs are pooled to predict the total potential energy $J_{ij}$ 7, and vector outputs are differentiated to yield atomic forces,

$J_{ij}$ 8

The exchange branch is a physics-informed $J_{ij}$ 9-MLP operating on the local geometry of each first-nearest-neighbor Cr–Cr pair. The extracted descriptors are the bridging angle $E(3)$ 0, four Cr–I leg lengths $E(3)$ 1 summarized by the mean $E(3)$ 2, and the Cr–Cr distance $E(3)$ 3. Its analytical baseline is

$E(3)$ 4

with learnable parameters $E(3)$ 5. A shallow two-layer residual MLP then models corrections beyond the analytical ansatz, and the final prediction is

$E(3)$ 6

The paper explicitly identifies the embedded inductive bias as the Goodenough–Kanamori superexchange–direct-overlap competition. Equivariance under translations, rotations, and inversion is described as crucial for transfer from the 8-atom training cell to the 3,200-atom deployment supercell, because forces rotate correctly and no angular data augmentation is needed (Balghari et al., 10 Jun 2026).

3. Dataset construction and optimization protocol

The training data are generated on the 8-atom primitive cell of monolayer CrI $E(3)$ 7 using a DFT+U workflow specified as GGA-PBE + $E(3)$ 8 eV, DFT-D3, Quantum ESPRESSO $E(3)$ 9 Wannier90 $\Delta$ 0 TB2J. Three strain modes are sampled: biaxial, uniaxial along the $\Delta$ 1-axis, and shear in the $\Delta$ 2 channel. Each mode spans $\Delta$ 3 to $\Delta$ 4 in 11 steps, with each strained configuration relaxed and then “rattled” by Gaussian displacements with $\Delta$ 5 and $\Delta$ 6 Å, producing 3–5 snapshots per strain. This yields 467 total configurations, stratified by mode and strain magnitude into 345 training, 61 validation, and 61 test configurations (Balghari et al., 10 Jun 2026).

Optimization is also split by branch. The structural branch uses a combined energy-and-force loss with force weight $\Delta$ 7. The exchange branch uses an L1 loss on bondwise exchange MAE. The E-GNN is trained with Adam, initial $\Delta$ 8, reduce on plateau by $\Delta$ 9 with patience 200 epochs, batch size 64, for 40,000 epochs. The $1.1$0-MLP is trained with AdamW, $1.1$1, weight decay 0.4, batch size 64, and L1 loss. Featurization uses 32 basis functions for each distance through CosineSmearing and for each angle through ChebyshevAngleSmearing. Early stopping is based on validation MAE: combined energy/force MAE for the E-GNN and $1.1$2 MAE for the $1.1$3-MLP.

The stratified split and the one-time evaluation on a strictly withheld test set are methodologically significant. They support the paper’s claim that train/validation/test agreement reflects negligible overfitting rather than repeated tuning against the final test configurations.

4. Accuracy and ablation evidence

The reported test metrics are summarized below.

Quantity	Test result	Evaluation set
Energy MAE	$1.1$4 meV/atom	61 withheld configurations
Force MAE	$1.1$5 meV/Å	61 withheld configurations
Exchange coupling MAE	$1.1$6 meV, $1.1$7	61 withheld configurations

The paper states that Fig. 3 shows the parity plot $1.1$8 versus $1.1$9 on the test set, and that close train/validation/test agreement demonstrates negligible overfitting. It also emphasizes that no direct baseline on CrI$6.5$0 exists, so the main architectural evidence comes from internal ablations rather than benchmark ranking (Balghari et al., 10 Jun 2026).

Two ablations are highlighted. First, removing the Goodenough–Kanamori analytical block degrades robustness at large distortions: residual-only MLPs tend to over- or under-predict extreme $6.5$1 values under large-angle distortions. Second, removing $6.5$2 equivariance destroys length-scale transferability, with models failing to generalize from the 8-atom cell to the 3,200-atom supercell. This suggests that the analytical bias and the symmetry constraint are not interchangeable regularizers; each controls a distinct failure mode, namely extrapolation in exchange space and extrapolation in system size.

5. Mesoscale deployment under dynamic strain

The mesoscale application is carried out under a collinear Ising, adiabatic approximation. The simulation cell is a $6.5$3 primitive-cell supercell containing 3,200 atoms and having dimensions $6.5$4 Å. Structural dynamics are propagated with Langevin MD in ASE using a 5 fs time step, damping $6.5$5 ps$6.5$6, and $6.5$7 K. The initial displacement field is a sinusoidal biaxial perturbation,

$6.5$8

with $6.5$9 Å and $0.18$0, producing nominal strains of approximately $0.18$1–$0.18$2 (Balghari et al., 10 Jun 2026).

The dynamical picture reported in the paper is specific. The incident strain wave propagates freely while the local strain remains below the FM–AFM threshold of approximately $0.18$3. After reflection at the periodic boundaries, constructive interference locally pushes the compressive strain beyond $0.18$4, and transient AFM-sign exchange regions with $0.18$5 emerge in circular domains. The authors then extract two mesoscale observables that are inaccessible to direct DFT.

For the domain wall width $0.18$6, the largest AFM cluster is extracted at interference peaks in frames 36 and 89, a radial profile $0.18$7 is computed from the cluster center, and the profile is fit to a hyperbolic tangent. The fitted values are $0.18$8 nm with $0.18$9 at frame 36 and $R^2=0.91$ 0 nm with $R^2=0.91$ 1 at frame 89, giving a mean $R^2=0.91$ 2 nm. For the oscillation period $R^2=0.91$ 3, the AFM Cr fraction $R^2=0.91$ 4 is monitored in time; constructive-interference peaks at frames 36 and 89 are separated by $R^2=0.91$ 5 fs $R^2=0.91$ 6 fs, yielding $R^2=0.91$ 7 ps.

The local strain field is quantified through the deformation gradient $R^2=0.91$ 8 and its symmetric part,

$R^2=0.91$ 9

with the most compressive principal eigenvalue $400\times$ 0 tracking regions that exceed the $400\times$ 1 threshold. The paper identifies the domain wall width and oscillation period as observables lying within the resolution of cryogenic magnetic force microscopy, and therefore as testable predictions.

6. Reproducibility, transferability, and terminological boundaries

The reproducibility package described in the paper includes two components: a GitHub repository for the DFT + Wannier90 + TB2J pipeline and raw dataset, and a separate GitHub repository for the DSpinGNN model, training scripts, and ASE simulation workflows. The implementation stack is NequIP/e3nn for the structural branch, PyTorch for the $400\times$ 2-MLP, and ASE for molecular dynamics (Balghari et al., 10 Jun 2026).

The paper also provides a transfer recipe for other 2D magnets. The stated procedure is: generate a DFT + U or DFT + SOC dataset on an 8-atom or minimal cell, sample relevant strains and small thermal distortions, extract exchange couplings with TB2J or a similar tool, adopt the bifurcated architecture with an $400\times$ 3-equivariant structural GNN and a physics-informed $400\times$ 4-MLP, retrain jointly with combined losses for forces, energies, and exchange couplings, and validate on withheld test configurations before deploying to large-scale MD. Within its stated approximations—collinear Ising spins, SOC omission, adiabatic decoupling, and nearest-neighbor isotropic exchange—the model is presented as delivering near-DFT accuracy on structural and magnetic observables and as being transferable to mesoscopic simulations inaccessible to first principles.

A practical source of confusion is nomenclature. The 2026 DSpinGNN for monolayer CrI $400\times$ 5 is distinct from the rotation-invariant graph model based on spin convolutions in “Rotation Invariant Graph Neural Networks using Spin Convolutions” (Shuaibi et al., 2021), and it is also unrelated to Dy-SIGN, the “Dynamic Spiking Framework for Graph Neural Networks” (Yin et al., 2023). The shared presence of “spin,” “dynamic,” and “GNN” vocabulary does not imply a common architecture or application domain.