First–Principles Spin–Lattice Coupling from Downfolded Electron–Phonon Interaction

The Heisenberg model and its associated spin dynamics have long been fundamental tools for understanding the magnetic properties of materials, enabling researchers to bridge the gap between microscopic electronic interactions and macroscopic magnetic behavior. However, as the field of materials science pushes toward the limits of efficiency and speed, the precise treatment of the coupling between spin and lattice degrees of freedom has become increasingly critical. This spin-phonon coupling is not merely a correction to the magnetic Hamiltonian but is the physical origin of many interesting phenomena, including strong modulation of magnon and phonon lifetimes [1], the emergence of new excitations [2] and magnon-to-phonon [3; 4] conversion, which could have a strong impact on spintronics, ultrafast demagnetization, magnetocalorics, spincaloritronics, and multiferroics [5].

Traditionally, the dynamics of magnetic structures are described using the Heisenberg Hamiltonian, which captures the quadratic interactions between localized spins. While such a model provides a robust framework for simulating magnetic properties via spin dynamics, it typically assumes a rigid lattice where atomic positions remain fixed. In real materials, thermal and quantum fluctuations of the atomic nuclei excite phonons which dynamically modulate the crystal field and intersite electronic hopping. This modulation in turn alters the exchange interactions between the spins, giving rise to an intrinsic coupling between the magnetic and lattice degrees of freedom [6; 7].

Several related terms are used to describe the coupling between spin and lattice degrees of freedom, including spin–lattice coupling (SLC), spin–phonon coupling (SPC), magnon–phonon coupling (MPC), and magnon–lattice coupling (MLC). These terms emphasize different representations: one may start from localized spin rotations or from collective magnon excitations, and similarly from local atomic displacements or collective phonons. The corresponding parameters can often be transformed into one another, sometimes with additional approximations. For example, the magnon spectrum can be obtained from a spin Heisenberg model through spin dynamics or linear spin-wave theory, while the phonon spectrum can be obtained from a lattice model through molecular dynamics or the harmonic approximation. Although these terms are often used interchangeably in the literature, they are not strictly identical. In this work, we use the term SLC because our formulation is based on localized real-space spin and lattice perturbations.

The spin-lattice coupled dynamics has been formulated in several works [8; 9; 10; 11; 12; 13; 14; 15; 16], and its application to specific materials [8; 17; 18] has shown that the quantitative analysis of some phenomena necessitates the inclusion of both spin and lattice degrees of freedom. The total Hamiltonian can be written as the sum of the spin, lattice and coupling terms: $H=H^{\mathrm{spin}}+H^{\mathrm{latt}}+H^{\mathrm{SLC}}$.

Both the lattice and the spin parts of the Hamiltonian can be constructed from the results of first-principles calculations. The spin part can be written in a Heisenberg-like form: $H^{\mathrm{spin}}=-\sum_{IJ}(J^{\mathrm{iso}}_{IJ}\mathbf{S}_{I}\cdot\mathbf{S}_{J}+\mathbf{S}_{I}\bar{\bar{\mathbf{J}}}^{\mathrm{ani}}_{{IJ}}\mathbf{S}_{J}+\mathbf{D}_{IJ}\cdot(\mathbf{S}_{I}\crossproduct\mathbf{S}_{J}))$, considering the isotropic exchange interaction, where $I$ and $J$ denote the indices of the atoms. With this sign convention, a positive $J_{IJ}^{\mathrm{iso}}$ favors ferromagnetic alignment and a negative $J_{IJ}^{\mathrm{iso}}$ favors antiferromagnetic alignment. Other terms such as the Dzyaloshinskii-Moriya interaction $\mathbf{D}_{IJ}$ (DMI), the anisotropic exchange $\bar{\bar{\mathbf{J}}}^{\mathrm{ani}}$, the single ion anisotropy, or the magnetic dipole-dipole interaction can also be included, and they would also be coupled with the lattice displacements. In this work, we focus on the coupling between the isotropic exchange interaction and lattice distortions, which to first order is parametrized with:

where $\tau_{u}$ is the displacement of an atom along one direction, and $u$ contains the atom and direction indices. Higher-order polynomial couplings can be generalized but are more complex to compute from first principles.

Many methods can be used to model the lattice vibrations, from those based on the simple harmonic approximation to more sophisticated models which include anharmonic interactions [19; 20], and more recently machine learning interatomic potentials (MLIPs) [21]. The harmonic phonons can be modeled with interatomic force constants, which can be calculated with either density functional perturbation theory (DFPT) [22; 23], or finite differences (FD) [24] from first-principles.

In the case of the spin part of the Hamiltonian, there are several methods for computing the parameters in spin models. One of the most commonly used methods is based on first-principles total-energy calculations and energy mapping [25]. Density Functional Perturbation Theory with respect to external magnetic fields was developed by Savrasov [26]. Alternatively, the Generalized Bloch Theorem (GBT) [27; 28] can be used to calculate the total energy of different spin-spiral states, from which the exchange parameters can be extracted. Finally, the magnetic force theorem (MFT) method makes use of Green’s functions, taking the local spin rotation as a perturbation, as proposed in the Liechtenstein, Katsnelson, Antropov and Gubanov formalism (LKAG) [29]. The MFT method has been widely used thanks to its simplicity (only unperturbed ground state calculations are required as input), efficiency (the Green’s function expressions are direct and closed-form expressions), and accuracy (for many purposes the perturbative approximation is sufficient). This method was extended to deal with correlated materials [30], magneto-crystalline-anisotropy [31; 32], DMI [33; 34; 35; 32] in fully relativistic cases with non-collinear spin ground states [36], and higher-order magnetic interactions [37; 38]

In the literature, several methods for estimating SLC parameters have been developed, including supercell-based finite-difference schemes [39; 10; 25; 40] and response-function approaches such as Berry-curvature calculations [41], two-particle Bethe–Salpeter equation [42], or time-dependent DFPT [43; 44]. While these methods provide comprehensive insights, they typically involve heavy supercell calculations or complex dynamical spectral frameworks.

We propose instead a perturbative downfolding approach that extracts real-space SLC parameters directly from the primitive-cell electron–phonon coupling (EPC). By extending the MFT to include atomic displacements as a simultaneous perturbation, our method computes derivatives of exchange parameters from Wannier-interpolated EPC and avoids the need for displaced magnetic supercells. The method provides an efficient and orbital-resolved framework that is closely related to the work of Mankovsky et al. [16]. This approach bridges the gap between simple force fitting and advanced response theories, offering a practical way to simulate complex spin-lattice dynamics.

We describe the MFT method for the calculation of the spin interaction parameters, and the Wannier representation of the EPC, in Section II.3. In Section III we demonstrate the method with the example of SrMnO₃. Then Section IV presents potential applications and extensions, and discusses limitations. An advantage of the present Wannier-based perturbative scheme is that it is built on already established workflows for computing phonons and magnetic exchange parameters, and can therefore be integrated into existing first-principles calculations. A standard ground-state DFT calculation first provides the reference electronic structure. A DFPT phonon calculation then yields the phonon frequencies and the variation of the self-consistent potential $\partial V_{\text{SCF}}/\partial\tau_{u}$, evaluated in the reference state. Maximally localized Wannier functions are constructed to transform the Kohn–Sham Hamiltonian $H(\mathbf{k})$ into its real-space representation $H(\mathbf{R})$ [45]. The EPC is then interpolated in the Wannier basis to obtain $g_{ij,u}(\mathbf{R}_{e},\mathbf{R}_{\mathrm{P}})$ [46]. Finally, magnetic post-processing codes [47] use the Wannier Hamiltonian to build the magnetic Green’s functions and evaluate the MFT expressions, so that the four-point vertex can be integrated to obtain the SLCs $\partial J/\partial\tau_{u}$. The workflow is presented in Fig. 1.

To describe SLC, we first need (i) the electronic Hamiltonian of the reference state and (ii) the coupling between lattice displacements and the electron density. The remaining task is to combine these two ingredients to obtain the SLC parameters. The notation and conventions are summarized in Table 1.

We validate the method with SrMnO₃ (SMO) in the cubic perovskite structure, where the magnetic exchange between Mn $3d$ electrons is strongly coupled to the phonons [65; 66]. The Mn ions are in a $3d^{3}$ electronic configuration, and interact with their first Mn neighbors through a superexchange mechanism, transmitted through the Mn-O orbital hybridization, and modulated by the Mn-O-Mn bonds.

The DFT calculations were performed with the Quantum ESPRESSO package [62]. We make use of the PBEsol Generalized Gradient Approximation (GGA) [67] exchange-correlation functional, and norm-conserving pseudopotentials from PseudoDojo [68; 69]. The phonons and EPC matrices are calculated with DFPT, with a $5\times 5\times 5$ Monkhorst-Pack grid of $\mathbf{q}_{\mathrm{P}}$-points, and a $10\times 10\times 10$ grid of $\bf k$-points. The SMO cubic structure in a ferromagnetic configuration is taken as the reference state. The Wannier90 [45] package allows us to build maximally localized Wannier functions [52], which include the Mn $3d$ and O $2p$ states. We then carry out a DFPT calculation to compute the phonons and EPC. Based on these results, the Wannier representation of the EPC is constructed with the EPW code [48; 46; 64]. In the current implementation of EPW, the Wannier representation of the EPC can be calculated at the collinear level, obtaining a spin-resolved EPC $g^{\uparrow}$ or $g^{\downarrow}$ [48]. We did not include the Hubbard $U$ correction in this work for simplicity, and it is left for future work since it has recently been made available in EPW [70; 71]. For the calculation of the exchange coupling parameters, we use the magnetic force theorem [29] implemented in the TB2J [47] code based on the Wannier function Hamiltonian.

To benchmark the method, we use a finite-difference method (FD) to compute the SLC parameters for comparison. We displace the atoms by $\pm 0.02$ Å along the $x$, $y$, and $z$ directions, and compute the exchange parameters in a $3\times 3\times 3$ supercell. Due to finite displacement effects, we find that the corresponding 3$\times$3$\times$3 $\mathbf{q}_{\mathrm{P}}$-point grid in PT is not sufficient and at least a 4$\times$4$\times$4 $\mathbf{q}_{\mathrm{P}}$-point grid is needed, see Fig. 7. The PT approach also offers a significant computational advantage over the FD method.

For SrMnO₃, each FD calculation in the $3\times 3\times 3$ supercell requires about 230 CPU-hours (AMD EPYC Milan 2.45 GHz), and two calculations (positive and negative displacements) are needed for each perturbation. This yields a total FD cost of $t_{\rm FD}\approx 460\times n$ CPU-hours, where $n$ is the number of independent perturbations. In contrast, the DFPT calculation with a $10\times 10\times 10$ $\mathbf{k}$-point mesh and a $5\times 5\times 5$ $\mathbf{q}_{\mathrm{P}}$-point mesh requires about 800 CPU-hours as a one-time cost.

The PT calculation corresponds to an effective $5\times 5\times 5$ supercell (from the $\mathbf{q}_{\mathrm{P}}$-point grid), which would be even more CPU-time and memory demanding if done with the FD approach.

A more important advantage of the PT calculation is the memory usage. Because the DFPT calculation is performed independently for each $\mathbf{q}_{\mathrm{P}}$ point, the memory footprint does not grow directly with the number of $\mathbf{q}_{\mathrm{P}}$ points, apart from small post-processing overheads. Some $\mathbf{q}_{\mathrm{P}}$ points reduce the number of symmetries and can therefore increase the maximum memory usage among the individual calculations. In contrast, for the supercell calculation, the storage of the wave functions scales as $O(N^{2})$, where $N$ is the supercell volume. In the present case, the $3\times 3\times 3$ calculation requires about 40 GB of memory, while a $5\times 5\times 5$ supercell would require about 21 times more memory, or roughly 880 GB.

We validate the computation of the SLC parameters with four atomic distortions: one Sr displacement, one Mn displacement, and two O atom displacements (along and perpendicular to the Mn-O bond in the $a$ direction as shown in Fig. 3), labeled by Sr-$x$, Mn-$z$, O-$x$, and O-$z$, respectively. The exchange and the SLC parameters connecting the Mn atoms are plotted as a function of the distance. The first-neighbor $J$ is significantly larger than the longer-ranged ones. The derivatives of $J$ with respect to atomic displacements also decay with distance as shown in Fig. 4. These exchange parameters can be used to compute the magnon band structure via linear spin-wave theory (LSWT) [72] as implemented in TB2J. The resulting magnon dispersion, shown in Fig. 5, shows a typical antiferromagnetic (AFM) behavior which is linear close to the zone center. We note that although our DFT calculations are done within a unit cell of FM structure, the computation of the magnon band for the reference AFM ground-state structure can be done following the method in Ref. [72].

The exchange values in Fig. 4 are much larger than those reported in previous works [73]. This difference is mainly due to the absence of a Hubbard $U$ correction in the calculations used for the present demonstration. For comparison, a DFT+$U$ calculation with $U(\mathrm{Mn})=3$ eV and $J(\mathrm{Mn})=0$ eV gives a first-neighbor Mn–Mn exchange of about $-7.04$ meV, whereas the calculation without $U$ used here gives $-21.45$ meV. The Hubbard correction pushes the occupied and unoccupied Mn $d$ states further apart, reducing Mn $d$–O $p$ hybridization and thus the superexchange interaction. The SLC parameters are expected to be overestimated for the same reason. As we mentioned before, the inclusion of the $U$ is left for future studies and we mainly focus on demonstrating the method to compute the SLC parameters regardless of the value of $J$. The corresponding values of the derivative of the isotropic exchange $J^{\mathrm{iso}}$ with respect to the displacements are shown in Table 2.

For the Sr-$x$ distortion [Fig. 6(a)], displacements 1 and $1^{\prime}$ are related by the mirror plane $m_{yz}$. They therefore give the same change in $J_{x-}$, so the first derivative of $J_{x-}$ with respect to the Sr-$x$ displacement vanishes. The couplings $J_{y+}$ and $J_{y-}$ are not symmetry equivalent for this distortion, whereas $J_{y-}$ and $J_{z-}$ are related by a four-fold rotation around the $x$ axis through the displaced Sr atom (not shown), so their derivatives are equal.

For the Mn-$z$ distortion [Fig. 6(b)], displacements 2 and $2^{\prime}$ are related by the mirror plane $m_{xy}$ and change the in-plane exchange couplings in the same way. As a result, the derivatives of $J_{x+}$, $J_{x-}$, $J_{y+}$, and $J_{y-}$ vanish, while the derivatives of $J_{z+}$ and $J_{z-}$ are antisymmetric. For the distortion O-$x$ [Fig. 6(c)], the first derivative of $J_{x-}$ should be 0, as the two displacements 3 and $3^{\prime}$ are related by a mirror plane $m_{yz}$ and thus give the same effect on $J_{x-}$. The other SLCs for the 1NN are non-zero. The $J_{y-}$, $J_{y+}$, $J_{z-}$, $J_{z+}$ should be equivalent by a four-fold rotation symmetry, and indeed we obtain the same values.

For the distortion O-$z$ [Fig. 6(d)], the O atom is displaced along the $z+$ direction. The $J_{x+}$ and $J_{x-}$ should be 0, as the two displacements 4 and $4^{\prime}$ are related by a mirror plane $m_{xy}$ and thus give the same effect on $J_{x-}$. The same holds for $J_{y+}$ and $J_{y-}$. The $J_{z+}$ and $J_{z-}$ are anti-symmetric.

The results in Table 2 confirm all the symmetry relations discussed above, validating the method, except for small numerical noise. We note that the results from the FD method have larger numerical noise when checking the symmetry relations. This is likely due to inaccuracies in the Wannier functions constructed for the distorted structures, which can break the symmetry.

As an additional numerical check, we calculate the exchange parameters in supercells with atomic As an additional numerical check, we calculate the exchange parameters in supercells with atomic displacements and then obtain the derivatives of the $J$ values by FD. The qualitative agreement between PT and FD is good, but some quantitative differences remain (e.g., for O1-$x$, PT gives $-7.38$ meV/Å while FD gives $-10.34$ meV/Å). The discrepancies can be attributed to the following reasons: (i) the Wannier functions do not preserve their shape under finite displacements, (ii) numerical noise in the Wannier Hamiltonian, and (iii) wrapping of the contributions from periodic images in the FD supercell calculation. Regarding point (3), the FD calculation was done in a $3\times 3\times 3$ supercell due to memory limitations, while the PT calculation corresponds to an effective $5\times 5\times 5$ supercell (from the $\mathbf{q}_{\mathrm{P}}$-point grid). The periodic images in the smaller FD supercell introduce spurious interactions that affect the SLC values. This suggests that the PT method is not only faster, but can also be more accurate than the FD approach, as it effectively bypasses the finite-size supercell errors by working directly in the primitive cell with a dense $\mathbf{q}_{\mathrm{P}}$-point grid.

Equations (4) and (14) can be evaluated for each pair of orbitals rather than only for each pair of atoms. This decomposes both the total exchange parameter and the SLC into orbital-pair contributions, providing a direct way to identify the electronic channels responsible for the coupling. We illustrate this orbital resolution for the Mn-$z$ displacement. For each Mn atom, the active $3d$ orbitals are $d_{z^{2}}$, $d_{xz}$, $d_{yz}$, $d_{x^{2}-y^{2}}$, and $d_{xy}$. The orbital-resolved isotropic exchange for $J_{z+}=-21.45$ meV can therefore be written as the following $5\times 5$ matrix:

The exchange is mainly due to the $d_{z^{2}}-d_{z^{2}}$, $d_{xz}-d_{xz}$, and $d_{yz}-d_{yz}$ pairs. The latter two show a superexchange picture, as the $d_{xz}$ and $d_{yz}$ orbitals are singly occupied and hybridize with the bridging oxygen $2p$ orbitals. As the unoccupied $d_{z^{2}}$ orbital hybridizes with the oxygen $p_{z}$ orbitals, the hopping favors a ferromagnetic alignment. Thus the antiferromagnetic interaction is the net result of these two competing mechanisms. The contributions from other orbital pairs are either zero due to symmetry or much smaller.

With the Mn-$z$ displacement, the distance between the two Mn atoms decreases. We obtain the following derivative of $J_{z+}$:

Both the ferromagnetic interaction between the $d_{z^{2}}-d_{z^{2}}$ pair, and the antiferromagnetic interaction from the $d_{xz}-d_{xz}$, and $d_{yz}-d_{yz}$ pairs are enhanced as the Mn-O-Mn bond is shortened. In total, $J_{z+}$ becomes more antiferromagnetic. The contributions from other orbital pairs are much smaller, or 0 due to the symmetry.

Another example is the $J$ between a Mn-Mn pair along the $y+$ direction with the O1-$x$ displacement (displacement 3 in Fig. 3 ). The isotropic exchange $J_{y+}$ is $J^{\mathrm{iso}}=-21.45$ meV at a distance of 3.810 Å. The orbital-resolved contribution to $J_{y+}$ is:

The derivative with respect to O1-$x$ displacement is $dJ_{y+}/d\tau_{u}=9.20$ meV/Å, with orbital contributions:

Interestingly, the $J_{y+}$ has significant contributions from the $d_{xy}-d_{xy}$ and $d_{yz}-d_{yz}$ pairs, both contributing -11.10 meV to $J_{y+}$. However, their derivatives with respect to the O1-$x$ displacement are very different: the $d_{yz}-d_{yz}$ pair enhances the antiferromagnetic interaction by -2.51 meV/Å, while the $d_{xy}-d_{xy}$ pair reduces the antiferromagnetic interaction by 13.09 meV/Å.

This can be understood as follows: with the O1 atom moving along the $x+$ direction, it gains stronger hybridization with the Mn atom, especially with the Mn $d_{xy}$ orbital. This competes with the hopping along the Mn $d_{xy}$–O $p$–Mn₂ $d_{xy}$ superexchange path in the $y$ direction, thus reducing the antiferromagnetic interaction. On the other hand, the O1 displacement does not significantly change the hybridization along the Mn $d_{yz}$–O $p$–Mn $d_{yz}$ hopping path. There are also contributions from other orbital pairs, notably the symmetric contribution from the $d_{x^{2}-y^{2}}-d_{z^{2}}$ and $d_{z^{2}}-d_{x^{2}-y^{2}}$ pairs, which are affected by the O1 displacement but largely cancel out (see the off-diagonal terms in the derivative matrix).

The analysis above shows that the orbital-resolved SLC parameters can provide deep insight into the mechanism of the SLC, and clearly illustrates the impact of ligand atoms in modulating the superexchange interaction.

This new method allows SLC parameters to be computed by post-processing standard DFT and DFPT calculations. Its inputs are the EPC matrix, the WF $U_{jn,\mathbf{k}}$ matrix, and the Wannier Hamiltonian. These quantities are already generated in established workflows for phonons, electron–phonon coupling, and magnetic exchange parameters. The present approach therefore adds the SLC calculation as a post-processing layer on top of these well-established methods and software packages.

The algorithm itself is simple and only involves a few matrix multiplications and Fourier transforms, which are easy to implement and efficient. There is no need for supercells. The cutoff ranges of the interactions are controlled by the $\mathbf{k}$ and $\mathbf{q}_{\mathrm{P}}$ grids. This real-space truncation is supported by the “nearsightedness” [74; 75] of electron–electron and electron–phonon interactions in the Wannier representation. The corresponding real-space couplings decay with distance and can therefore be truncated beyond a finite cutoff without strongly affecting the results. In practice, this approximation is well satisfied for localized magnetic exchange and for the short-range component of the EPC [63; 64].

In polar insulators, however, longitudinal optical phonons generate a long-range dipolar Fröhlich electron–phonon interaction [76] that decays slowly in real space and is divergent at $\mathbf{q}_{\mathrm{P}}\to 0$. Modern EPC implementations usually handle this by separating the EPC into a short-range part, which is Wannier-interpolated, and a long-range analytical part that describes the macroscopic dipole field in reciprocal space [77; 46]. In the mixed-space approach implemented in EPW, the long-range part is subtracted on the coarse grid and added back analytically on the fine grid. Recent developments have further extended this separation to include dynamical quadrupoles and higher-order multipoles [78; 79], as well as specific treatments for two-dimensional systems where screening is qualitatively different [80; 81; 82].

We compute the SLC with a series of $\mathbf{k}$-point and $\mathbf{q}_{\mathrm{P}}$-point meshes in the EPC calculation. With $n_{\mathbf{k}}=3$ and $n_{\mathbf{q}_{\mathrm{P}}}=3$, the values are far from the finite-difference results and also from PT results with denser $\mathbf{k}$ and $\mathbf{q}_{\mathrm{P}}$ meshes. For $n_{\mathbf{q}_{\mathrm{P}}}=5$, the $n_{\mathbf{k}}=10$ and $n_{\mathbf{k}}=15$ values converge and are also close to the FD results. Our present SLC formalism implicitly inherits this short-/long-range partition: the localized Wannier basis captures the short-range contribution, while any explicit long-range dipolar or multipolar correction must be added consistently at the EPC level. As a result, the locality assumption and the chosen real-space cutoffs should be carefully validated when applying the method to polar materials or when long-wavelength phonons are important.

The proposed method generates the exchange parameters $J_{IJ}^{\mathrm{iso}}$ and their gradients $dJ_{IJ}^{\mathrm{iso}}/d\tau_{u}$ in real space, corresponding to local spin configurations and atomic displacements. These quantities are sufficient to build the full momentum-dependent coupling between collective excitations, namely magnons and phonons. One route is analogous to Wannier interpolation for EPC: just as the electronic Hamiltonian is transformed from the localized Wannier basis to Bloch states to reconstruct band structures, the magnetic Hamiltonian can be transformed to momentum space. Because the $J$ parameters decay rapidly with distance, the full magnon band structure $\omega(\mathbf{k})$ can be computed efficiently using LSWT. The local derivatives with respect to $\tau_{u}$ can then be projected onto phonon eigenvectors, which constructs the MPC at any point in the Brillouin zone.

The orbital resolution illustrated above is another useful feature of the method, because it connects the exchange and SLC parameters to the electronic channels that generate them.

While we demonstrate the method for a collinear spin system without SOC, the generalization of the SLC to the DMI and anisotropic exchange is also proposed here. Once the DFPT method and WF interpolation methods are implemented for these cases, the additional terms can be calculated, which will enable the simulation of many interesting properties. For example, the Gilbert damping in atomistic spin dynamics is partly due to thermal noise from the phonons. With the coupled model this phenomenon can be simulated explicitly. The spectra of magnons and phonons can also be calculated at finite temperature by considering their mutual interaction.

The approximations used in the derivation also define the limitations of the method. The first limitation is the locality assumption discussed above, especially for polar materials where long-range dipolar or multipolar EPC terms may be important. The second limitation is that the spin–spin interaction is described by a classical Heisenberg model, which neglects possible non-Heisenberg terms and relies on a rigid-spin picture. In this approximation, the magnitude of each local moment is fixed and only its orientation changes, while the onsite response $\delta\bar{\bar{\mathbf{p}}}_{I}$ in Eq. (14) is treated as a small correction. In more itinerant systems, or close to a magnetic instability, lattice distortions can modify not only the exchange couplings between sites but also the size of the local moments themselves, corresponding to longitudinal spin fluctuations. In such cases, the rigid-spin approximation may become inaccurate, and a more general spin–lattice model in which both the direction and magnitude of the spins are dynamical variables would be more appropriate.

A related limitation is the neglect of higher-order spin interactions. The LKAG method has recently been extended to include higher-order terms, such as biquadratic, three-spin, and four-spin interactions [37; 38]. These extensions could be adapted to the present formalism fairly straightforwardly. Finally, the Heisenberg model also fails for strongly itinerant magnetic moments, which are beyond the scope of the current method.

In some systems, second- or higher-order derivatives of the exchange with respect to atomic displacements may be important. The current method can be extended by including terms with two lattice perturbations, as illustrated for higher-order SLC in Appendix A.

To summarize, we have presented a method that combines perturbations to the spin and lattice degrees of freedom to compute the SLC parameters. Our approach uses the localized Wannier-function representation of the electronic Hamiltonian and the electron–phonon coupling matrix as input, both of which can be obtained directly from DFPT primitive-cell calculations with minimal extra computational cost.

The method is computationally efficient: for SrMnO₃, it reduces the cost by two orders of magnitude compared to the finite-difference approach, and this advantage grows with the number of displacement perturbations. Moreover, by working directly in the primitive cell with a dense $\bf q$-point grid, the perturbative approach effectively bypasses finite-size supercell errors, making it potentially more accurate than supercell-based finite-difference methods. The method can be extended to other types of spin interactions and higher-order phonon terms. Furthermore, it can provide detailed information on the electronic and orbital origin of the SLC. We illustrated this method with the example of SrMnO₃. The resulting real-space parameters provide a direct route to parameterizing spin–lattice Hamiltonians for finite-temperature spin–lattice dynamics using standard perturbative workflows.

The TB2J code, including the spin–lattice coupling feature developed in this work, is available at https://github.com/mailhexu/TB2J. The Quantum ESPRESSO and EPW codes are available at https://www.quantum-espresso.org/ and https://epw-code.org/, respectively. The SrMnO₃ example data for computing the SLC parameters are openly available on Materials Cloud [83].