Quantum-accurate magneto-elastic predictions with

classical spin-lattice dynamics

Svetoslav Nikolov1 , Mitchell A. Wood1 , Attila Cangi2 , Jean-Bernard Maillet3,4 ,
Mihai-Cosmin Marinica5 , Aidan P. Thompson1 , Michael P. Desjarlais6 , and Julien
Tranchida1,*
1 Computational Multiscale Department, Sandia National Laboratories, P.O. Box 5800, MS 1322, Albuquerque, NM
87185
2 Center for Advanced Systems Understanding (CASUS), Helmholtz-Zentrum Dresden-Rossendorf, 02826 Görlitz,

Germany
arXiv:2101.07332v2 [cond-mat.mtrl-sci] 20 Jan 2021

3 CEA - DAM, DIF, Arpajon Cedex F-91297, France

4 Université Paris-Saclay, CEA, LMCE, 91680 Bruyères-le-Châtel, France
5 Université Paris-Saclay, CEA, Service de Recherches de Métallurgie Physique, Gif-sur-Yvette 91191, France
6 Sandia National Laboratories, P.O. Box 5800, MS 1322, Albuquerque, NM 87185
* jtranch@sandia.gov

ABSTRACT

A data-driven framework is presented for building magneto-elastic machine-learning interatomic potentials (ML-IAPs) for
large-scale spin-lattice dynamics simulations. The magneto-elastic ML-IAPs are constructed by coupling a collective atomic
spin model with an ML-IAP. Together they represent a potential energy surface from which the mechanical forces on the
atoms and the precession dynamics of the atomic spins are computed. Both the atomic spin model and the ML-IAP are
parametrized on data from first-principles calculations. We demonstrate the efficacy of our data-driven framework across
magneto-structural phase transitions by generating a magneto-elastic ML-IAP for α-iron. The combined potential energy
surface yields excellent agreement with first-principles magneto-elastic calculations and quantitative predictions of diverse
materials properties including bulk modulus, magnetization, and specific heat across the ferromagnetic-paramagnetic phase
transition.

Introduction a second-order phase transition19 . Without an efficiently scal-

able coupled spin-lattice dynamics simulation environment,
Magnetism strongly influences thermomechanical properties that properly accounts for thermal expansion and magnetic
in a large variety of materials, such as single-element mag- contribution to the pressure, reproducing the divergence of C p
netic metals1, 2 , steels3 , high-entropy alloys4, 5 , nuclear fuels at the critical point would not be possible.
such as uranium dioxide6 , magnetic oxides7, 8 , and numerous Accurate numerical simulations are critical for enabling
other classes of functional materials9 . Despite the critical technological advances, as they shape our fundamental un-
role of magnetism in the aforementioned materials classes, derstanding of the underlying solid state physics that dictates
modeling efforts to study the interplay between structural and material behavior. Developing high fidelity models however
magnetic properties have been notably lacking. Furthermore, is challenging, because it necessitates capturing physical phe-
there are unanswered scientific questions regarding the signif- nomena that occur across several length and time scales. This
icance of magnetism in matter that is shock-compressed10, 11 can only be achieved with sufficiently accurate multiscale
or exposed to strong electromagnetic fields such as in co- simulation tools22, 23 , which is the focus of this work.
herent lights sources12, 13 , pulsed power and high magnetic Classical molecular dynamics (MD) simulations24 provide
fields facilities14, 15 . These conditions are highly relevant for a useful framework for multiscale modeling by leveraging in-
improving our understanding of astrophysical objects such teratomic potentials (IAPs) to represent the dynamics of atoms
as the structure, dynamics, and formation processes of the on a Born-Oppenheimer potential energy surface (PES)25 . By
Earth’s core16 and other planets17, 18 . Properties of interest utilizing massively parallel algorithms26 and long time-scale
include phase transitions, thermal stability of spin defects, methodologies27 , MD enables bridging first-principles with
magneto-mechanical couplings, but many of these subjects continuum-scale simulations28 .
are challenging or prohibited by state of the art computational The absorption of machine learning (ML) techniques into
tools. A prime, but simple, example of the computational ad- the creation of interatomic potentials has lead to classical
vance made herein is the heat-capacity of α-iron displayed in MD simulations that approach the accuracy of first-principles
Figure 1. The experimental measurement of the heat capacity methods. A growing number of these highly accurate ML-
C p diverges at the magnetic Curie transition, characteristic of IAPs29–35 have been developed. In general, they are parame-

1
 125 from first-principles calculations46 . However, this remained
an isolated attempt as there is no general methodology for
100
Cp (J.mol 1.K 1)
generating a magneto-elastic PES in a classical context that
enables large-scale spin-lattice dynamics simulations for any
75 magnetic material.
50 In this work, we overcome this methodological obstacle by
providing a data-driven framework for generating magneto-
25 elastic ML-IAPs that (1) provide a consistent representation
of both mechanical and magnetic degrees of freedom and (2)
00 250 500 750 1000 1250 achieve near first-principles accuracy. Our framework couples
Lattice Temp. Tl (K) an atomic spin model (Heisenberg Hamiltonian) with an ML-
IAP and provides a unified magneto-elastic PES which yields
Figure 1. Constant pressure heat capacity of α-iron versus the correct mechanical forces on the atoms in the MD frame-
temperature. The black triangles denote experimental mea- work. The Heisenberg Hamiltonian is parameterized with
surements20, 21 , the red squares our simulation results, and data from DFT spin-spiral calculations at different degrees of
black dashed line indicates the experimental Curie transition lattice compression. In constructing the ML-IAP, we leverage
temperature. This illustrates the well-known ferromagnetic- the flexible and data-driven spectral neighbor analysis poten-
paramagnetic phase transition, where the heat capacity di- tial (SNAP) methodology35 which is trained on a database of
verges at the Curie temperature. magnetic configurations generated using DFT calculations.
We apply our framework to generate a magneto-elastic ML-
IAP for the α phase of iron within a temperature and pressure
terized on training data (configuration energy, atomic forces) range of 0 to 1200 K and 0 to 13 GPa (up to the α → γ
from first-principles methods like density functional theory and α → ε transitions, respectively). The Curie temperature,
(DFT)36 and utilize different flavors of ML model forms to which experimentally occurs at approximately 1045 K, lies
construct the PES. While they have proven to be useful for within this parameter space. We highlight that our framework
large-scale simulations of thermodynamic materials proper- yields quantitative agreement with first-principles calculations.
ties37, 38 , further progress in multiscale modeling is hampered We also stress that our simulations take into account both
by the limitation of ML-IAPs to non-magnetic materials phe- the thermal expansion of the lattice and magnetic pressure
nomena. Even with highly accurate ML-IAPs, state-of-the-art due to spin disorder. This enables us to maintain a constant
MD simulations cannot reproduce the divergent behavior of ambient pressure throughout all calculations of thermome-
C p near the critical point (Figure 1) because they fail to ac- chanical properties, consistent with conditions prevalent in
count for the magnetic degrees of freedom39 . experiments. As illustrated in Figure 1, our framework allows
Coupling atomic spin dynamics with classical MD has been us to perform the first pressure-controlled quantitative pre-
pioneered by Ma et al.40–42 . Herein, a classical magnetic spin diction of the critical behavior across a second-order phase
is assigned to each atom in addition to its position leading transition within a classical spin-lattice dynamics simulation.
to a 6N-dimensional PES (5N if the magnetic spin norms
are fixed), instead of the common 3N-dimensional PES in Results
classical MD:
In this section we outline our advancements in magnetic ma-
N
terials modeling. We first present our training workflow and
E = ∑ ε ({rr i j , s i }) , (1)
subsequently assess our results by comparing both static and
i=1
dynamic properties in α-iron against first-principles calcula-
where r i j = r i −rr j denotes the relative position between atoms tions and experiments.
i and j, s i the classical spin assigned to atom i, and N the num- Figure 2 displays our training workflow. Further details to
ber of atoms in the system. In most classical spin-lattice each box in this diagram are presented as a subsection in the
calculations, the 6N-dimensional PES is constructed by intro- "Methods" section. All atomic configurations in the training
ducing an atomic spin model on top of a mechanical IAP40 . set result from first-principles calculations performed with the
For example, a common approach is to combine a distance- same DFT setup (same pseudo-potential and energy cutoff,
dependent Heisenberg Hamiltonian with an embedded-atom- similar k-point densities) as detailed in the "Methods" sec-
method (EAM) potential43 . tion. In contrast to traditional force-matching approaches in
While these prior approaches recover experimental proper- the development of classical IAPs, we treat the magnetic and
ties on a qualitative level44, 45 , their combined representation phononic degrees of freedom in the PES in a consistent and
of phononic and magnetic degrees of freedom is not suffi- unified manner, as indicated by the exchange of information
ciently consistent for providing quantitative predictions at the between spin Hamiltonian and SNAP potential parametriza-
level of first-principles results. More recently, Ma et al. devel- tion steps. After parameterizing our atomic spin Hamiltonian,
oped a magneto-elastic IAP for magnetic iron based on data its energy, forces, and stress contributions are subtracted from

2/14
 functions are based on both experimental and DFT data, as
outlined in Table 2. Objective function evaluations are done
within LAMMPS26 .
Herein, the critical innovation that enables a leap forward
in predictive simulations of magnetic materials is this data-
driven workflow. Magnetic and phononic contributions to the
PES are taken into account explicitly and any miscounting is
avoided. The obtained magneto-elastic ML-IAP can directly
be used to run spin-lattice calculations in LAMMPS26, 43, 48 .

Magneto-Static Accuracy
We first assess the quantitative agreement of our magneto-
elastic ML-IAP by comparing with DFT results where mag-
netic order and elastic deformations are coupled. Equation-
of-state calculations (energy and pressure versus volume) are
performed at the Γ point (corresponding to the purely ferro-
magnetic state) and for spin-spirals corresponding to q-vectors
along the ΓH and ΓP high-symmetry lines. The calculations
at the Γ point represent the magnetic ground state and, hence,
serve as a point of reference for the spin spiral calculations.
The geometric orientation of the various computed spin spirals
is visualized in Figure 3. The first set (q = 0.01 along ΓH
and q = 0.07 along ΓP) represents "long" spirals, close to the
Γ point, the second set (q = 0.1 along ΓH and q = 0.14 in
ΓP) represents spirals with intermediate periodicity, and the
last set (q = 0.2 along ΓH and q = 0.21 along ΓP) is chosen
close to the borders of the magnetic training set (see red de-
marcation lines in Figure 5 in the "Methods" section). The
DFT results are obtained by leveraging the generalized Bloch
theorem, whereas our classical spin-lattice calculations were
performed by generating the corresponding supercells (details
given in the "Methods" section).
Figure 2. Magneto-elastic ML-IAP training workflow. A
Excellent agreement between our classical spin-lattice
training set of DFT calculations is partitioned into those that
model and DFT is achieved at the Γ point and for the two first
train the SNAP interatomic potential and those that train
q-vectors on each high-symmetry line (q = 0.01 and q = 0.1
the spin Hamiltonian, respectively. A non-magnetic inter-
along ΓH, q = 0.07 and q = 0.14 along ΓP) in the pressure
atomic potential is fit to configuration energies and atomic
range relevant for the α-phase of iron (up to 13 GPa which cor-
forces after the spin Hamiltonian contribution is subtracted
responds to the α → ε transition). At higher q-vector values,
and is validated against magneto-elastic properties computed
the energy and pressure predictions of our atomic spin-lattice
in LAMMPS. Optimization of the spin Hamiltonian and inter-
model still agree reasonably well with the DFT calculations.
atomic potential parameters is handled by DAKOTA.
The observed small deviation from the DFT results can be ex-
plained by the limitations of our atomic spin-lattice model: as
both the pressure and the relative angle between neighboring
each atomic configuration in the first-principles training set. spins increase, fluctuations of the atomic spin norms become
The ML-IAP is then trained to reproduce the non-magnetic more important. As discussed in the "Methods" and "Discus-
component of the first-principles data. Finally, both com- sion" sections, these are not included in the Hamiltonian of
ponents of the magneto-elastic PES are recombined to con- our atomic spin-lattice model.
struct a unified magneto-elastic ML-IAP that is consistently
trained on first-principles data. Optimization is handled by Magneto-Dynamic Accuracy
the DAKOTA software package47 in both fitting steps. For Turning now to spin-lattice dynamics calculations based on
the SNAP potential, DAKOTA varies the radial cutoff along our magneto-elastic ML-IAP (as detailed in the "Methods"
with the weights of each training data set to generate different section), we assess the quantitative accuracy with respect to
phononic candidate potentials which are recombined with the experimental measurements of changes in magnetic and ther-
spin Hamiltonian and tested against selected objective func- moelastic properties as the material is heated. In making this
tions (lattice constant, cohesive energy, elastic constants, and comparison, it is necessary to choose which thermodynamic
force & energy rmse). The target values for the objective state variables will be held fixed and which will be allowed

3/14
 a) Γ point
q = [0 0 0]
E (eV/atom)

Press. (kbar)
0.1 100

0.0 0

0.9 1.0
V / V0
b) ΓH branch c) ΓP branch
q = 0.01[-1 1 1] q = 0.07[1 1 1]
E (eV/atom)

E (eV/atom)
Press. (kbar)

Press. (kbar)
0.1 100 0.1 100

0.0 0 0.0 0

q = 0.1[-1 1 1] q = 0.14[1 1 1]
E (eV/atom)

E (eV/atom)
Press. (kbar)

Press. (kbar)
0.1 100 0.1 100

0.0 0 0.0 0

q = 0.2[-1 1 1] q = 0.21[1 1 1]
E (eV/atom)

E (eV/atom)
Press. (kbar)

Press. (kbar)
0.1 100 0.1 100

0.0 0 0.0 0

0.9 1.0 0.9 1.0

V / V0 V / V0

Figure 3. Plots of the equation of state data from first-principles calculations (VASP computations) and our magneto-elastic
ML-IAP (LAMMPS computations) for seven different spin-spirals: a) Γ point b) vectors along the ΓH high-symmetry line
, and c) vectors along the ΓP high-symmetry line. Visualizations of the corresponding spin-spiral supercells and associated
q-vectors are shown to the right of and above each plot, respectively.

to vary with temperature. Spin-lattice dynamics algorithms refer to calculations performed in this pressure-controlled CE
have been developed for simulations in a canonical ensemble as "pressure-controlled conditions" (PCC). In both conditions,
(CE) which preserves the number of particles, the volume, and the temperature of the spin and lattice subsystems is set using
the temperature in the system42 . Our first set of simulation two separate Langevin thermostats (one acting on the spins,
conditions, referred to as "fixed-volume conditions" (FVC), the other on the lattice)42 . Finally, this enables us to define a
hold the volume fixed while running dynamics in the CE en- third set of conditions: in addition to controlling the pressure,
semble at specified values of the lattice and spin temperatures. the spin thermostat can be set to match a given magnetiza-
A disadvantage of this choice is that the pressure steadily in- tion value (i.e., the experimental magnetization) rather than
creases as heat is added to the material, in contradiction to the a temperature. We refer to this as "pressure-controlled and
experimental observations, which are conducted at constant magnetization-controlled conditions" (PCMCC).
pressure. To this date, an isobaric spin-lattice algorithm has
In practice, FVC, PCC and PCMCC only differ in their equi-
not been developed (preserving the system’s pressure rather
libration conditions (control of pressure and / or magnetiza-
than its volume). However, our methodology as implemented
tion), as each of the corresponding simulations are performed
in LAMMPS enables us to compute the magnetic contribution
in a canonical ensemble. We illustrate the predictive capability
to the pressure. By alternating thermalization (coupled spin-
of our magneto-elastic ML-IAP in α-iron for these equilibra-
lattice dynamics in a CE) and pressure equilibration (frozen
tion conditions in Figure 4.a-f (FVC : , PCC : , PCMCC
spin configuration in an isobaric ensemble) steps, it is possible
: ). The agreement of the following magneto-elastic prop-
to control the pressure of our spin-lattice system. Hence, we
erties with experimental results is assessed: magnetization

4/14
 a) b) c)
1.2 150 12.5 1250

1.0 125 10.0 1000

Spin Temp. Ts (K)

)
1

P(Tl ) (GPa)
0.8 100

.K
7.5
M(Tl , P)

750

1
Cp (J.mol
0.6 75 5.0
500
0.4 50 2.5
0.2 25 250
0.0
0.0 0 0
d) e) f)
200 70 130

60 120
(C 11-C 12)/2 (GPa)

180
Bulk (GPa)

50 110

C 44 (GPa)
160
40 100

140 30 90

20 80
120
10 70
0 500 1000 0 500 1000 0 500 1000
Lattice Temp. Tl (K) Lattice Temp. Tl (K) Lattice Temp. Tl (K)

Figure 4. Plots a-f show magnetoelastic data obtained with our magneto-elastic ML-IAP. The green ( ), blue ( ), and red ( )
markers indicate the choice of equilibration conditions: "fixed-volume conditions" (FVC), "pressure-controlled conditions"
(PCC) and "pressure-controlled and magnetization-controlled conditions" (PCMCC), respectively. In all plots, experimental
data (extracted from five different references20, 21, 49–51 ) is denoted by the filled triangles (N), and the dotted black lines ( )
represent the experimental Curie temperature. The plots in a-b) show magnetization and specific heat comparisons between
different ensembles and experiments. The light blue region in (b) indicates the low temperature regime T .200 K where
quantum effects reduce the experimental heat capacity below the classical Dulong-Petit limiting value of 3R52 . The data in plot
c) illustrates that for FVC ( ) the increase in temperature can raise the pressure up to 10 GPa. For both isobaric ensembles
(PCC and PCMCC) the pressure is barostated to 0 GPa. For PCMCC ( ) below the critical point, spin temperature is set to
reproduce the experimental magnetization. Plots d-f show (d) bulk modulus, (e) (c11 − c12 )/2 shear constant, and (f) c44 shear
constant for the three aforementioned sets of conditions.

(Figure 4.a), heat-capacity Cp (Figure 4.b), the bulk modulus value52 . The magnetic contribution is computed by extracting
(Figure 4.d), and two shear constants, (c11 − c12 )/2 and c44 the purely magnetic energy of the system and evaluating its
(Figure 4.e-f). variance, as detailed in Eriksson et al.53 . At low temperature,
deviation between simulations and experiment (highlighted by
We first work under the FVC ( ), keeping a constant vol- the semi-transparent blue region in Figure 4.b) occurs due to
ume and equal spin and lattice temperatures (Figure 4.c). At quantum effects which reduce the experimental heat capacity
constant volume, our model predicts a Curie temperature of below the classical Dulong-Petit limiting value of 3R. The
approximately 716K (Figure 4.a). Specific heat calculations FVC heat-capacity is determined at constant volume, although
shown in Figure 4.b were done by taking into account both we use the symbol Cp on the axis label because the enhanced
lattice and magnetic contributions. The SNAP contribution simulations described below are indeed conducted at constant
to the heat-capacity was first isolated and determined to be pressure conditions. Figure 4.c shows the substantial pressure
26.4 Jmol−1 K−1 , in good agreement with the Dulong-Petit

5/14
evolution with temperature increase (up to 12 GPa, almost of about 25-30 GPa in the bulk modulus is observed as we
corresponding to the α → ε transition), which have a strong move across the critical point. This jump was found to be
impact on the underlying elastic properties. Interestingly, at strongly impacted by the underlying mechanical potential.
the Curie temperature (here 716K), the increasing pressure The prediction accuracy could possibly be improved by in-
exhibits an inflection point, confirming the importance of spin cluding additional, finite-temperature objective functions in
fluctuations on the thermoelastic properties. The temperature the fitting procedure. The PCMCC prediction of the shear
dependence of three elastic constants is shown in Figure 4.d-f. constant c44 closely matches the PCC data. This tends to indi-
For the bulk modulus, FVC does not agree well with exper- cate that this shear constant c44 is not impacted significantly
imental data, especially at higher temperatures. The FVC by the spin dynamics. For both pressure controlled conditions
results tend to overestimate the stiffness, which most likely (PCC and PCMCC) the maximum deviation from experiments
arises from the build-up of thermal stresses in the material. occurs near 700K and is approximately 14%.
Under these conditions a nearly temperature-invariant c44 re-
sponse is predicted, which is in strong contrast to trends in Discussion
experiment. Despite these shortcomings, the FVC calcula-
tions actually match the experimental data for shear constant We presented a data-driven framework for automated gener-
(c11 − c12 )/2 relatively well throughout the entire tempera- ation of magneto-elastic ML-IAPs which enable large-scale
ture range. In general, the fixed volume assumption made spin-lattice dynamics simulations for any magnetic material
under FVC fails to account for thermal expansion, leading to in LAMMPS. This framework was demonstrated by gener-
incorrect elastic predictions. ating a robust magneto-elastic ML-IAP for α-iron. First we
investigated the magneto-static accuracy (energy and pres-
We correct this shortcoming of the model by working under
sure) with respect to equivalent first-principles calculations. It
PCC ( ) which allows for thermal expansion. The cell vol-
was demonstrated that the generated magneto-elastic ML-IAP
umes are relaxed at each finite temperature, until the pressure
(which represents the corresponding 5-N dimensional PES)
in the system drops to 0 GPa (Figure 4.c). As shown in Fig-
is in close agreement with first-principles magneto-elastic
ure 4.a, the thermal expansion incorrectly moves the onset of
calculations. Subsequently, we investigated the magneto-
Curie transition to approximately 536K. As the average inter-
dynamic accuracy by comparing predicted finite tempera-
atomic distance increases, the strength of the exchange inter-
ture magneto-elastic properties (magnetization, heat-capacity,
action is lowered, thus decreasing the transition temperature.
bulk modulus, and shear constants) across the ferromagnetic-
In comparison, PCC fares better in reproducing the experi-
paramagnetic phase transition from spin-lattice dynamics sim-
mental bulk modulus up to the Curie transition (no hardening
ulations against data from experiments. In the course of this,
observed). PCC also does better in terms of the shear constant
we analyzed the choice of simulation conditions (control of
c44 , as it is able to reproduce the thermal softening seen in
pressure and magnetization) and highlighted the importance
experiments. However, for shear constant (c11 − c12 )/2, PCC
of thermal and magnetic pressure contributions. This is an im-
underestimates the extent of the thermal softening. Overall,
portant advance over traditional spin-lattice dynamics, where
PCC does better than FVC in terms of elastic properties, but
a fixed lattice negates any contribution from thermal expan-
deviates more in terms of magnetic predictions compared to
sion or spin pressure due to disorder. We demonstrated that
experiment. By shifting the Curie transition towards lower
spin-lattice dynamics simulations of controlled pressure and
temperatures, it reduces the range of validity of our elastic
constrained magnetization yields close qualitative and quan-
calculations.
titative agreement with the measured magneto-elastic prop-
In order to improve the magnetic predictions of α-iron, we erties. Most remarkably, our framework enables quantita-
finally consider the PCMCC scheme ( ). In addition to allow- tive predictions of the critical behavior across a second-order
ing for thermal expansion similarly to the PCC, we also set phase transition within a classical spin-lattice dynamics sim-
the spin thermostat temperature in order to reproduce the ex- ulation, such as the divergent behavior of the heat capacity
perimental magnetization. As shown in Figure 4.c, below the around the Curie temperature (Figure 1). Such a positive
Curie transition, the spin temperature increases more slowly agreement between computational and experimental data was
than the lattice temperature, while above the Curie transition, only achieved by properly partitioning the potential energy
it increases at the same rate as the lattice temperature. Fig- surface into magnetic and mechanical degrees of freedom.
ure 4.a shows that the obtained magnetization under PCMCC Additionally, our finite-temperature predictions highlight the
closely matches that of experiment. Most prominently, the importance of proper spin pressure accounting.
resulting Cp agrees well with experiments (Figure 4.b). Up We conclude the discussion of our results by pointing out
to approximately 600 K, PCMCC agrees very well with the limitations of the present method and future prospects. First,
experimental values for (c11 − c12 )/2 (Figure 4.e) but at 800- note that the agreement to the experimental Curie transition
1000K a slight hardening is observed, which contradicts ex- (Tc ≈ 716K in a fixed volume calculation) could have been ad-
perimental data. For the bulk modulus, PCMCC correctly justed by parameterizing the spin potential on a smaller range
predicts the nearly linear trend up to the Curie temperature. of the high-symmetry lines (see Figure 5), or by adding an ob-
We note that in all three sets of conditions, a rapid increase jective function aimed at matching the experimental value in

6/14
the spin-potential fitting procedure. However, this additional where q is the spin-spiral vector, R 0 j is the position of atom
constraint would have worsened the agreement of our model j relative to a central atom 0, s j is the spin on atom j, and
with the DFT energy and pressure results (as displayed on θ is a constant angle between the spins and the spin-spiral
Figure 3) and would contradict the overall objective of this vector (often referred to as "cone angle")67 . x̂x, ŷy, and ẑz are the
work. unit vectors along [100], [010], and [001], respectively. Our
The main limitation of our work lies in the simplicity of the calculations are restricted to θ = π/2, corresponding to flat
spin Hamiltonian model used. Extended spin Hamiltonians, spin-spirals in the (001) plane.
such as spin-cluster expansions, might be a promising route to First-principles calculations of the per-atom energy and the
improving the accuracy of the magnetic component of the PES pressure corresponding to spin-spiral states are performed us-
by both accounting for longitudinal spin-norm fluctuations ing DFT by leveraging the frozen-magnon approach68, 69 and
and many-body spin interactions54, 55 . Enhanced magnetic the generalized Bloch theorem70 as implemented in VASP71 .
thermostats have also been proposed in order to better match We consider a primitive cell of one atom. A 10 × 10 × 10 k-
the experimental magnetic transition versus temperature56, 57 , point grid, an energy cutoff of 320 eV, and 224 bands proved
and could be an efficient way of improving our spin-lattice sufficient to reach the level of accuracy expected in our model
simulations (for example, by replacing the magnetization- (as can be seen in Figure 5).
controlled conditions defined in the "Results" section). Classical calculations are performed by using Eq. (2) to
A straightforward extension of this work could combine generate supercells accommodating the spin-spirals corre-
recently developed extended spin Hamiltonians with first- sponding to the q -vectors used in the DFT calculations. Based
principles studies, and apply our formalism to extend our on a given supercell and a spin Hamiltonian, the per-atom
α-iron magneto-elastic ML-IAP to account for defect con- energy and pressure are computed using the SPIN package of
figurations58, 59 , Cr clustering60, 61 , and magneto-structural LAMMPS26, 43 .
phase-transitions11, 62 .
In summary, we have presented a new computational frame- Spin Hamiltonian
work for near quantum-accuracy simulations of magneto- A spin Hamiltonian is used to model the energy, mechanical
elastic materials properties. By leveraging the flexibility of forces, and pressure contributions of magnetic configurations.
ML-IAPs, our data-driven workflow enables to model the in- Rosengaard and Johansson72 and Szilva et al.73 showed that
terplay between magnetic and phononic dynamics for a large adding a biquadratic term to the classical Heisenberg Hamil-
class of magnetic materials. Furthermore, our straightforward tonian improves the accuracy of magnetic excitations in 3-d
connection to the LAMMPS package makes it possible to per- transition ferromagnets. We adopted their Hamiltonian form:
form large-scale quantitative magneto-elastic predictions over N
controlled pressure and temperature spaces, hitherto study Hmag = − ∑ J (ri j ) [ssi · s j − 1]
unexplored magneto-dynamics properties of materials. i6= j
N h i
− ∑ K (ri j ) (ssi · s j )2 − 1 , (3)
Methods i6= j
Density functional theory calculations where s i and s j are classical atomic spins of unit length lo-
Parameterizing both the ML-IAP and the magnetic Heisenberg cated on atoms i and j, J (ri j ) and K (ri j ) (in eV) are magnetic
Hamiltonian relies on data computed using spin-dependent exchange functions, and ri j is the interatomic distance be-
DFT calculations. They were performed using VASP63, 64 . In tween magnetic atoms i and j. The two terms in Eq. 3 are
all calculations the PBE65 exchange-correlation functional offset by subtracting the spin ground state (corresponding to
was employed. We used PAW pseudopotentials66 with 8 va- a purely ferromagnetic situation), as detailed in Ma et al.40 .
lence electrons and a core radius of rc = 2.3 aB . The plane Although this offset of the exchange energy does not affect the
wave cutoff was set to 320 eV and the convergence in each self- precession dynamics of the spins, it allows to offset the cor-
consistency cycle was set to 10−8 . The Fermi-Dirac smearing responding mechanical forces. Without this additional term,
scheme with a width of 0.026 eV was used. The first Brillouin the magnetic contribution to the forces and the pressure are
zone was sampled on a 10 × 10 × 10 grid of k-points. The not zero at the energy ground state. For the exchange interac-
number of bands used was 224. tion terms J (ri j ) and K (ri j ), the interatomic dependence is
taken into account through the following function based on
Spin-spiral calculations an approximation of the Bethe-Slater curve74, 75 :
Spin-spirals define a subset of non-collinear magnetic states.
In this work, we leverage spin-spirals as a convenient tool to  r 2   r 2     
r 2
f (r) = 4α 1−γ exp − Θ (Rc − r) ,
perform one-to-one comparisons between first-principles and δ δ δ
classical magneto-elastic calculations. They can be defined as (4)
follows:
where α denotes the interaction energy, δ the interaction de-
s j = sin θ cos(qq · R0 j )x̂x + sin θ sin(qq · R0 j )ŷy + cos θ ẑz , (2) cay length, γ a dimensionless curvature parameter, r a distance,

7/14
and Θ (Rc − r) a Heaviside step function for the radial cutoff LAMMPS DFT 2%
Rc . This assumes that the interaction decays rapidly with the
interatomic distance, consistent with former calculations73, 76 . DFT Exp. Loong
We set Rc = 5Å to include five neighbor shells, as Pajda et LAMMPS 2% Exp. Lynn
al.76 showed that the exchange interaction decays slower
along the [111] direction in α-iron.

∆E (meV)
Using Eq. (3) and leveraging the generalized spin-lattice 200
Poisson bracket as defined by Yang et al.77 , the magnetic
precession vectors (ωω i ), mechanical forces F i ), and their
(F
corresponding virial components (W r N ) are derived:


0

1 Ni

Mag. (µB )
= J (ri j ) s j + K (ri j ) (ssi · s j ) s j , (5)
h̄ ∑
ωi
j 2.0
Ni
dJ (ri j )
Fi = ∑ [ssi · s j − 1] ei j 1.5
j dri j

∆P (GPa)
dK (ri j ) h i
0
+ (ssi · s j )2 − 1 e i j , (6)
dri j
N
W rN


= ∑ ri · F i , (7) −20
i=1 1.0 0.5 Γ 0.5 1.0
−1
where rN denotes a 3N size vector of all atomic positions H ←− kqk ( Å ) −→ P
and r i the position vector of atom i. We note that the virial
components enable computing the spin contribution to the Figure 5. Comparison of spin-spiral results along sections
pressure. of the ΓH and ΓP high-symmetry lines. The upper plot dis-
The spin Hamiltonian is used to reproduce spin-spiral en- plays the per-atom energy, the middle one the atomic moment
ergy and pressure reference results obtained from DFT. They fluctuations (in Bohr magneton per atom), and on the bottom
are sampled along two high-symmetry lines, ΓH and ΓP, and the evolution of the pressure. The energy and pressure fluc-
for two different lattice constant values (corresponding to the tuations are plotted with respect to the magnetic ground state
equilibrium bulk value and to a lattice compression of 2%). at the Γ point. The green and red dots represent experimental
This allows us to encapsulate in the model the influence of measurements obtained by Loong et al.78 and Lynn79 . In
lattice compression on the spin stiffness and the Curie temper- all three plots, the dashed lines correspond to the DFT re-
ature, which was experimentally and theoretically predicted sults, and the continuous lines to our classical model results,
to be small80–82 . Figure 5 displays the excellent agreement whereas the line color (black or blue) corresponds to the lat-
obtained between our first-principles spin-spiral energies and tice compression (0 or 2%, respectively). In the middle plot,
experimental measurements. the green dashed horizontal line represent the experimental
Our current spin Hamiltonian does not account for longi- equilibrium value (2.2 µB per atom), which is the constant
tudinal spin fluctuations (LSF), i.e. the norm of atomic spins value chosen in our model. In all three plots, the red vertical
remains constant in our calculations. As can be seen in Fig- dashed lines are delimiting the q-vectors on which our spin
ure 5, this is not the case for our DFT results, as the LSF Hamiltonian was parametrized.
can become important when departing from the Γ point. We
thus decided to parameterize our model only on spin-spirals
corresponding to q -vectors for which the spin norm deviates and thus a better agreement for the Curie temperature. How-
from the ferromagnetic value (≈ 2.2 µB /atom at the Γ point) ever, this would worsen the pressure agreement.
by less than 5%. The red dashed lines in Figure 5 delimit this Spin-orbit coupling effects were included by accounting
q -vector range. for an iron-type cubic anisotropy83 :
Finally, we used the single objective genetic algorithm N 
within the DAKOTA software package47 to optimize the six Hcubic = − ∑ K1 (ssi · x̂x)2 (ssi · ŷy)2 + (ssi · ŷy)2 (ssi · ẑz)2 + ...
coefficients of J (ri j ) and K (ri j ) in order to obtain the best i=1
 (8)
possible agreement between our reference DFT spin-spiral (c)
(ssi · x̂x)2 (ssi · ẑz)2 + K2 (ssi · x̂x)2 (ssi · ŷy)2 (ssi · ẑz)2 ,
energy and pressure results and our spin model. Figure 5
displays the obtained result. As can be seen in Figure 4, for
(c)
a fixed-volume calculation, our spin Hamiltonian predicts a with K1 = 0.001 eV and K2 = 0.0005 eV the intensity coef-
Curie temperature of 716K. Note that a better match of the ficients corresponding to α-iron. The cubic anisotropy was
DFT spin-spiral energies would yield a larger spin-stiffness, only included to run calculations, but ignored in the fitting

8/14
procedure as its intensity is below the range of accuracy of rations or forces they are applied to, therefore allowing for
our ML-IAP. larger group weights be (cautiously) interpreted more valu-
In all our classical spin-lattice dynamics calculations, our able at meeting the set of targeted objective functions. This
system size remained small compared to the typical magnetic optimized Fe-SNAP interatomic potential is contained as Sup-
domain-wall width in iron83 . Thus, long-range dipole-dipole plemental Material along with LAMMPS input scripts used
interactions could safely be neglected. in the following section.

Spin-lattice dynamics
SNAP potential Calculations are performed following the spin-lattice dy-
For this work, an interatomic potential for iron was devel- namics approach as implemented in the SPIN package of
oped that is specifically parameterized for use in coupled LAMMPS26, 43 , and set by the spin-lattice Hamiltonian be-
spin and molecular dynamics simulations. Training data for low:
a Spectral Neighborhood Analysis Potential (SNAP)48, 84, 85
was collected to constrain the fit to the pressure and tem-
N N
perature phase space of < 20GPa and < 2000K. The set of |pp|2
Hsl (rr , p , s ) = Hmag (rr , s ) + ∑ + ∑ VSNAP (ri j ) (9)
non-colinear, spin-polarized VASP calculations includes BCC-
i=1 2mi i, j=1
, HCP- and liquid-iron, Table 1 displays the quantity of each
training type and target properties that are captured therein. where Hmag is the spin Hamiltonian defined by the com-
Optimization of a SNAP potential necessitates that the gen- bination of Eq. (3) and Eq. (8). The term VSNAP (ri j ) is our
erated training database be broken into these groups (rows SNAP ML-IAP. The second term on the right in Eq. (9), rep-
in Table 1) such that the weighted linear regression can (de- resents the kinetic energy, where the particle momentum is
)emphasize different parts in search of a global minima in given as p and the mass of particle i is mi . Based on this spin-
objective function errors. Each training group is assigned a lattice Hamiltonian and leveraging the generalized spin-lattice
unique weight for its’ associated energies and atomic forces Poisson bracket as defined by Yang et al.77 , the equations of
for each candidate potential, optimization of these weights motion can be defined as:
is controlled by DAKOTA. In order to avoid double count-
ing, and properly simulate the magnetic properties of iron in
drr i p
classical MD, we have adapted the SNAP fitting protocol48 to = i (10)
isolate the non-magnetic energy and forces from the generated dt mi
training data. To do so, the fitted biquadratic spin Hamiltonian
is evaluated for every atom in the training set, and its’ contri- N 
bution to the total energy and per-atom forces is subtracted. d pi dVSNAP (ri j ) dJ(ri j )
= ∑ − + (ssi · s j ) + ...
This is akin to previous uses of an ion core repulsion86 or dt j,i6= j dri j dri j
(11)
electrostatic interaction term87 as a reference potential while dK(ri j )

γL
2
fitting SNAP models. (ssi · s j ) e i j − p i + f (t)
dri j mi
Optimization of the SNAP potential was achieved using
a single objective genetic algorithm within the DAKOTA
software package47 . Radial cutoff distance, training group

dssi 1
= ω i + η (t)) × s i + ...
(ω
weights and number of bispectrum descriptors were varied dt (1 + λ 2 )
to minimize a set of objective functions, as percent error to  (12)
available DFT or experimental89 data, that encapsulate the λ s i × (ω
ω i × si)
desired mechanical properties of Fe. These objective func-
tions specific to BCC-Fe are listed in Table 2, and the RMSE Particle positions are advanced according to Eq. (10). The
energy and force regression errors are included in optimiza- derivative of the momentum, given in Eq. (11), is dependent
tion as well. In all objectives, our linear SNAP model with 31 not only on the mechanical potential but the magnetic ex-
bispectrum descriptors achieves accuracy in all mechanical change functions as well. Here γL is the Langevin damping
properties within a few percent of experiment/DFT. Addi- constant for the lattice and f is a fluctuating force following
tionally, lattice constants and cohesive energies of FCC and Gaussian statistics given below43 .
HCP phases were fit, but given far less priority with respect to
the BCC mechanical properties resulting in ∼ 6 − 7% errors
with respect to DFT. Importantly, each of the objective func- h f (t)i = 0 (13)
tions were evaluated including the magnetic spin contributions 0 0
h fα (t) fβ (t )i = 2kB Tl γL δαβ δ (t − t ) (14)
to avoid unforeseen changes in property predictions. A full
breakdown of the optimal training group weights and mean The fluctuating force f is coupled to γL via the fluctuation
absolute energy/force errors are given in Table 1. Group dissipation theorem as shown in Eq. (14). Here kB is the
weights listed have been adjusted by the number of configu- Boltzmann constant, Tl is the lattice temperature, and α and

9/14
 # of Config. # of Forces Target Property Energy Fit Weight Forces Fit Weight Energy MAE (eV) Forces MAE (eV·Å−1 )
Eq. of State 403 65286 Volumetric Deform 4.2 · 103 2.0 · 105 1.6 · 10−2 2.4 · 10−1
DFT-MD, 300K 40 15360 Bulk phonons 2.9 · 105 1.1 · 105 5.2 · 10−4 2.4 · 10−1
Liquid w/ Spins 10 3000 Magnetic Disorder 5.5 · 101 1.9 · 104 2.0 · 10−1 5.9 · 10−1
Liquid w/o Spins 52 15300 Structural Disorder 3.3 · 103 2.0 · 104 2.2 · 10−1 8.0 · 10−1
Point Defects 10 3096 Defect Energetics 1.4 · 102 3.5 · 104 2.8 · 10−2 1.1 · 10−1
Martensitic Transform 168 1008 α →ε 4.0 · 102 2.3 · 103 9.2 · 10−2 2.3 · 10−1

Table 1. Training set for linear SNAP model adapted form Ref. [88 ] to include explicit spin degrees of freedom. Regression
of SNAP coefficients takes into account both configuration energies and forces from DFT, optimization of group weights is
applied to either term independently. Weighted linear regression is carried out via reported optimal fit weights, values have
already been scaled by the number of training points each group contributes.

in the z-direction. To measure the magnetic properties for

SNAP Exp/DFT Units Error %
the canonical ensemble we initially thermalize the system
c11 243.25 239.55 GPa 1.54% under NVT dynamics at the target spin/lattice temperatures
for 40 picoseconds and then sample the target properties for
c12 135.65 138.1 GPa 1.77%
10 picoseconds. For pressure-controlled simulations (see PCC
c44 118.73 120.75 GPa 1.67% and MCPCC in the "Results" section), after the initial 40
picosecond of temperature equilibration we freeze the spin
Bulk modulus 171.52 169.55 GPa 1.16%
configuration and run isobaric-isothermal NPT dynamics in
0.5(c11 -c12 ) 53.8 51.9 GPa 3.66% order to allow the system to thermally expand (still account-
ing for the effect of the "magnetic" pressure, generated by the
Poisson ratio 0.358 0.36 - 1.10%
spin Hamiltonian). The pressure damping parameter is set to
bcc energy -8.25 -8.26 eV 0.02% 10 picoseconds. The pressure equilibration run is terminated
once the system pressure drops below 0.05 GPa. After this,
bcc lat. const. 2.838 2.83 Å 0.30%
the spin configuration is unfrozen and another equilibration
run is carried out under NVT dynamics for 20 picoseconds.
Table 2. Objective functions of the DAKOTA optimization Unfreezing the spin configurations causes a small jump in the
with ground truth values taken from the present DFT calcula- pressure, typically within the range of +/- 2 GPa. To reduce
tions(at zero Kelvin) or experiments89 . Percent error is used this pressure fluctuation, a series of uniform isotropic box
as the objective function to avoid artificial importance scaling deformations are performed under the NVE ensemble. During
based on units of the target property. this procedure the box is deformed in 0.02% increments every
2 picoseconds until the magnitude of the pressure is reduced
to approximately 0 GPa.
β are coordinates. Shown in Eq. (12) is the stochastic Landau- For the magnetization-controlled conditions (MCPCC in
Lifshitz-Gilbert equation which describes the precessional the "Results" section), the spin temperature is adjusted to
motion of spins under the influence of thermal noise. In match the experimental magnetization values. Spin tempera-
Eq. (12), λ is the transverse damping constant and ω i is a spin ture adjustments are made based on the magnetization curve
force analogue as shown in Eq. (5). The variable η (t) is a obtained in the pressure-controlled conditions (PCC in the
random vector whose components are drawn from a Gaussian "Results" section). The corresponding spin-lattice tempera-
probability distribution given below: ture relationship is shown in Eqs. (17-20). Here the fitting
coefficients are given as a1 = 471.6, a2 = 6362, a3 = 2774,
a4 = 1119, a5 = 13.6, a6 = 1043.3, and a7 = 0.1, respectively.
hη
η (t)i = 0 (15) The functions Ts,pre and Ts,post prescribe how the spin temper-
0 0 ature varies before and after the critical point. At the critical
hηα (t)ηβ (t )i = DS δαβ δ (t − t ) (16)
point we use a switching function fsw to smoothly transition
where the amplitude of the noise DS can be related to the from Ts,pre to Ts,post .
temperature of the external spin bath Ts according to DS =
2πλ kB Ts /h̄42 .
SD-MD calculations are carried out using a 20x20x20 BCC Ts,post (Tl ) = Tl − a1 (17)
 
cell. The BCC lattice is oriented along each of the coordi- 2
Ts,pre (Tl ) = a2 e −((Tl −a3 )/a4 ) − a5 (18)
nate directions. The MD timestep in all cases is set to 0.1

femtoseconds. For both spin and lattice thermostats we use a fsw (Tl ) = 0.5 + 0.5tanh (Tl − a6 )/a7 (19)
damping constant of 0.1. Initially all spins start out aligned Ts (Tl ) = fsw Ts,post + (1 − fsw )Ts,pre (20)

10/14
 After the magnetic measurements we compute elastic con- 11. Surh, M. P., Benedict, L. X. & Sadigh, B. Magnetostruc-
stants by performing both uniaxial and shear deformations tural transition kinetics in shocked iron. Phys. review
along each of the coordinate directions and planes. The mag- letters 117, 085701 (2016).
nitude of these deformations in all cases is 2% of the box 12. Moses, E. I., Boyd, R. N., Remington, B. A., Keane, C. J.
length. Following each deformation the box is relaxed for 3 & Al-Ayat, R. The national ignition facility: Ushering in
picoseconds. After this relaxation the stresses are sampled for a new age for high energy density science. Phys. Plasmas
2 picoseconds. 16, 041006, DOI: 10.1063/1.3116505 (2009).
13. Tschentscher, T. et al. Photon beam transport and scien-
Data Availability tific instruments at the european xfel. Appl. Sci. 7, DOI:
The data that support the findings of this study are available 10.3390/app7060592 (2017).
from the corresponding author upon reasonable request. 14. Tan, X., Chan, S., Han, K. & Xu, H. Combined effects
of magnetic interaction and domain wall pinning on the
Code Availability coercivity in a bulk nd 60 fe 30 al 10 ferromagnet. Sci.
reports 4, 6805 (2014).
The code which was used to train the SNAP potential
is available from: https://github.com/FitSNAP/ 15. Gràcia-Condal, A. et al. Multicaloric effects in metamag-
FitSNAP. netic heusler ni-mn-in under uniaxial stress and magnetic
field. Appl. Phys. Rev. 7, 041406 (2020).
16. Alfè, D. & Gillan, M. J. First-principles calculation of
References
transport coefficients. Phys. Rev. Lett. 81, 5161–5164,
1. Tatsumoto, E. & Okamoto, T. Temperature dependence DOI: 10.1103/PhysRevLett.81.5161 (1998).
of the magnetostriction constants in iron and silicon iron.
17. Militzer, B., Hubbard, W. B., Vorberger, J., Tamblyn,
J. Phys. Soc. Jpn. 14, 1588–1594 (1959).
I. & Bonev, S. A. A massive core in jupiter predicted
2. Bahl, C. R. H. & Nielsen, K. K. The effect of demagneti- from first-principles simulations. The Astrophys. J. 688,
zation on the magnetocaloric properties of gadolinium. J. L45–L48, DOI: 10.1086/594364 (2008).
Appl. Phys. 105, 013916 (2009). 18. Schöttler, M. & Redmer, R. Ab initio calculation of the
3. Tavares, S., Fruchart, D., Miraglia, S. & Laborie, D. Mag- miscibility diagram for hydrogen-helium mixtures. Phys.
netic properties of an aisi 420 martensitic stainless steel. Rev. Lett 120, 115703 (2018).
J. alloys compounds 312, 307–314 (2000). 19. Chandler, D. Introduction to modern statistical mechanics
4. Huang, S., Holmström, E., Eriksson, O. & Vitos, L. Map- (1987).
ping the magnetic transition temperatures for medium- 20. Wallace, D. C., Sidles, P. & Danielson, G. Specific heat
and high-entropy alloys. Intermetallics 95, 80–84 (2018). of high purity iron by a pulse heating method. J. applied
5. Rao, Z. et al. Unveiling the mechanism of abnormal physics 31, 168–176 (1960).
magnetic behavior of fenicomncu high-entropy alloys 21. Touloukian, Y. & Buyco, E. Thermophysical properties
through a joint experimental-theoretical study. Phys. Rev. of matter, vol. 4, specific heat. IFI/Plenum, New York
Mater. 4, 014402 (2020). (1970).
6. Jaime, M. et al. Piezomagnetism and magnetoelastic 22. Horstemeyer, M. F. The near Future: ICME for the
memory in uranium dioxide. Nat. communications 8, 1–7 Creation of New Materials and Structures. In Integrated
(2017). Computational Materials Engineering (ICME) for Metals,
7. Nussle, T., Thibaudeau, P. & Nicolis, S. Dynamic mag- chap. 10, 410–423, DOI: 10.1002/9781118342664.ch10
netostriction for antiferromagnets. Phys. Rev. B 100, (John Wiley & Sons, Ltd, 2012).
214428 (2019). 23. van der Giessen, E. et al. Roadmap on multiscale materi-
8. Lejman, M. et al. Magnetoelastic and magnetoelectric als modeling. Model. Simul. Mater. Sci. Eng. 28, 043001
couplings across the antiferromagnetic transition in mul- (2020).
tiferroic bifeo 3. Phys. Rev. B 99, 104103 (2019). 24. Alder, B. J. & Wainwright, T. E. Studies in Molecular
9. Patrick, C. E., Marchant, G. A. & Staunton, J. B. Spin Dynamics. I. General Method. The J. Chem. Phys. 31,
orientation and magnetostriction of tb 1- x dy x fe 2 from 459–466, DOI: 10.1063/1.1730376 (1959).
first principles. Phys. Rev. Appl. 14, 014091 (2020). 25. Rapaport, D. C. The art of molecular dynamics simulation
10. Graham, R., Morosin, B., Venturini, E. & Carr, M. Ma- (Cambridge university press, 2004).
terials Modification and Synthesis Under High Pressure 26. Plimpton, S. Fast parallel algorithms for short-range
Shock Compression. Annu. Rev. Mater. Sci. 16, 315–341, molecular dynamics. J. computational physics 117, 1–19
DOI: 10.1146/annurev.ms.16.080186.001531 (1986). (1995).

11/14
27. Voter, A. F., Montalenti, F. & Germann, T. C. Extending 40. Ma, P.-W., Woo, C. & Dudarev, S. Large-scale simulation
the time scale in atomistic simulation of materials. Annu. of the spin-lattice dynamics in ferromagnetic iron. Phys.
Rev. Mater. Res. 32, 321–346 (2002). review B 78, 024434 (2008).
28. Zepeda-Ruiz, L. A., Stukowski, A., Oppelstrup, T. & 41. Ma, P.-W., Dudarev, S. & Woo, C. Spilady: A parallel
Bulatov, V. V. Probing the limits of metal plasticity with cpu and gpu code for spin–lattice magnetic molecular
molecular dynamics simulations. Nature 550, 492–495 dynamics simulations. Comput. Phys. Commun. 207,
(2017). 350–361 (2016).
29. Huan, T. D. et al. A universal strategy for the creation of 42. Ma, P.-W. & Dudarev, S. Atomistic spin-lattice dynamics.
machine learning-based atomistic force fields. npj Com- Handb. Mater. Model. Methods: Theory Model. 1017–
put. Mater. 3, DOI: 10.1038/s41524-017-0042-y (2017). 1035 (2020).
30. Smith, J. S., Isayev, O. & Roitberg, A. E. Ani-1: an 43. Tranchida, J., Plimpton, S., Thibaudeau, P. & Thompson,
extensible neural network potential with dft accuracy at A. P. Massively parallel symplectic algorithm for cou-
force field computational cost. Chem. Sci. 8, 3192–3203, pled magnetic spin dynamics and molecular dynamics. J.
DOI: 10.1039/C6SC05720A (2017). Comput. Phys. 372, 406–425 (2018).
31. Zhang, L., Han, J., Wang, H., Car, R. & E, W. Deep Po- 44. Dos Santos, G. et al. Size-and temperature-dependent
tential Molecular Dynamics: A Scalable Model with the magnetization of iron nanoclusters. Phys. Rev. B 102,
Accuracy of Quantum Mechanics. Phys. Rev. Lett. 120, 184426 (2020).
143001, DOI: 10.1103/PhysRevLett.120.143001 (2018).
45. Zhou, Y., Tranchida, J., Ge, Y., Murthy, J. & Fisher,
32. Bartók, A. P., Payne, M. C., Kondor, R. & Csányi,
T. S. Atomistic simulation of phonon and magnon ther-
G. Gaussian Approximation Potentials: The Accuracy
mal transport across the ferromagnetic-paramagnetic tran-
of Quantum Mechanics, without the Electrons. Phys.
sition. Phys. Rev. B 101, 224303 (2020).
Rev. Lett. 104, 136403, DOI: 10.1103/PhysRevLett.104.
136403 (2010). 46. Ma, P.-W., Dudarev, S. & Wróbel, J. S. Dynamic simula-
33. Jaramillo-Botero, A., Naserifar, S. & Goddard, W. A. tion of structural phase transitions in magnetic iron. Phys.
General Multiobjective Force Field Optimization Frame- Rev. B 96, 094418 (2017).
work, with Application to Reactive Force Fields for Sili- 47. Eldred, M. S. et al. Dakota, a multilevel parallel object-
con Carbide. J. Chem. Theory Comput. 10, 1426–1439, oriented framework for design optimization, parameter es-
DOI: 10.1021/ct5001044 (2014). timation, uncertainty quantification, and sensitivity analy-
34. Lubbers, N., Smith, J. S. & Barros, K. Hierarchi- sis. Tech. Rep., Citeseer (2006).
cal modeling of molecular energies using a deep neu- 48. Thompson, A. P., Swiler, L. P., Trott, C. R., Foiles, S. M.
ral network. The J. Chem. Phys. 148, 241715, DOI: & Tucker, G. J. Spectral neighbor analysis method for
10.1063/1.5011181 (2018). automated generation of quantum-accurate interatomic
35. Thompson, A. P., Swiler, L. P., Trott, C. R., Foiles, S. M. potentials. J. Comput. Phys. 285, 316–330 (2015).
& Tucker, G. J. Spectral neighbor analysis method for 49. Crangle, J. & Goodman, G. The magnetization of pure
automated generation of quantum-accurate interatomic iron and nickel. Proc. Royal Soc. London. A. Math. Phys.
potentials. J. Comput. Phys. 285, 316–330, DOI: 10. Sci. 321, 477–491 (1971).
1016/j.jcp.2014.12.018 (2015).
50. Seki, I. & Nagata, K. Lattice constant of iron and austen-
36. Kohn, W. & Sham, L. J. Self-Consistent Equations In- ite including its supersaturation phase of carbon. ISIJ
cluding Exchange and Correlation Effects. Phys. Rev. international 45, 1789–1794 (2005).
140, A1133–A1138, DOI: 10.1103/PhysRev.140.A1133
(1965). 51. Ridley, N. & Stuart, H. Lattice parameter anomalies at
the curie point of pure iron. J. Phys. D: Appl. Phys. 1,
37. Li, X.-G., Chen, C., Zheng, H., Zuo, Y. & Ong, S. P. Com- 1291 (1968).
plex strengthening mechanisms in the nbmotaw multi-
principal element alloy. npj Comput. Mater. 6, 1–10 52. Ashcroft, N. W., Mermin, N. D. & Wei, D. Solid State
(2020). Physics (Cengage Learning Asia Pte Limited, 2016).
38. Cusentino, M., Wood, M. & Thompson, A. Suppres- 53. Eriksson, O., Bergman, A., Bergqvist, L. & Hellsvik, J.
sion of helium bubble nucleation in beryllium exposed Atomistic spin dynamics: Foundations and applications
tungsten surfaces. Nucl. Fusion 60, 126018 (2020). (Oxford university press, 2017).
39. Dragoni, D., Daff, T. D., Csányi, G. & Marzari, N. 54. Drautz, R. & Fähnle, M. Spin-cluster expansion:
Achieving dft accuracy with a machine-learning inter- Parametrization of the general adiabatic magnetic energy
atomic potential: Thermomechanics and defects in bcc surface with ab initio accuracy. Phys. Rev. B 69, 104404
ferromagnetic iron. Phys. Rev. Mater. 2, 013808 (2018). (2004).

12/14
55. Drautz, R. Atomic cluster expansion of scalar, vectorial, 71. Marsman, M. & Hafner, J. Broken symmetries in the
and tensorial properties including magnetism and charge crystalline and magnetic structures of γ-iron. Phys. Rev.
transfer. Phys. Rev. B 102, 024104 (2020). B 66, 224409 (2002).
56. Woo, C., Wen, H., Semenov, A., Dudarev, S. & Ma, P.-W. 72. Rosengaard, N. & Johansson, B. Finite-temperature study
Quantum heat bath for spin-lattice dynamics. Phys. Rev. of itinerant ferromagnetism in fe, co, and ni. Phys. Rev.
B 91, 104306 (2015). B 55, 14975 (1997).
57. Bergqvist, L. & Bergman, A. Realistic finite temperature 73. Szilva, A. et al. Interatomic exchange interactions for
simulations of magnetic systems using quantum statistics. finite-temperature magnetism and nonequilibrium spin
Phys. Rev. Mater. 2, 013802 (2018). dynamics. Phys. review letters 111, 127204 (2013).
58. Marinica, M.-C., Willaime, F. & Crocombette, J.-P. 74. Kaneyoshi, T. Introduction to amorphous magnets (World
Irradiation-induced formation of nanocrystallites with Scientific Publishing Company, 1992).
c 15 laves phase structure in bcc iron. Phys. review letters
108, 025501 (2012). 75. Yosida, K., Mattis, D. C. & Yosida, K. THEORY OF
MAGNETISM.: Edition en anglais, vol. 122 (Springer
59. Chapman, J. B., Ma, P.-W. & Dudarev, S. L. Effect
Science & Business Media, 1996).
of non-heisenberg magnetic interactions on defects in
ferromagnetic iron. Phys. Rev. B 102, 224106 (2020). 76. Pajda, M., Kudrnovskỳ, J., Turek, I., Drchal, V. & Bruno,
60. Chapman, J. B., Ma, P.-W. & Dudarev, S. L. Dynamics P. Ab initio calculations of exchange interactions, spin-
of magnetism in fe-cr alloys with cr clustering. Phys. Rev. wave stiffness constants, and curie temperatures of fe, co,
B 99, 184413 (2019). and ni. Phys. Rev. B 64, 174402 (2001).
61. Klaver, T., Drautz, R. & Finnis, M. Magnetism and 77. Yang, K.-H. & Hirschfelder, J. O. Generalizations of
thermodynamics of defect-free fe-cr alloys. Phys. Rev. B classical poisson brackets to include spin. Phys. Rev. A
74, 094435 (2006). 22, 1814 (1980).
62. Kalantar, D. et al. Direct observation of the α- ε tran- 78. Loong, C.-K., Carpenter, J., Lynn, J., Robinson, R. &
sition in shock-compressed iron via nanosecond x-ray Mook, H. Neutron scattering study of the magnetic exci-
diffraction. Phys. review letters 95, 075502 (2005). tations in ferromagnetic iron at high energy transfers. J.
63. Kresse, G. & Furthmüller, J. Efficiency of ab-initio total applied physics 55, 1895–1897 (1984).
energy calculations for metals and semiconductors using a 79. Lynn, J. Temperature dependence of the magnetic excita-
plane-wave basis set. Comput. Mater. Sci. 6, 15 – 50, DOI: tions in iron. Phys. Rev. B 11, 2624 (1975).
https://doi.org/10.1016/0927-0256(96)00008-0 (1996).
80. Leger, J., Loriers-Susse, C. & Vodar, B. Pressure effect
64. Kresse, G. & Joubert, D. From ultrasoft pseudopotentials on the curie temperatures of transition metals and alloys.
to the projector augmented-wave method. Phys. Rev. B 59, Phys. Rev. B 6, 4250 (1972).
1758–1775, DOI: 10.1103/PhysRevB.59.1758 (1999).
81. Morán, S., Ederer, C. & Fähnle, M. Ab initio electron
65. Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized theory for magnetism in fe: Pressure dependence of spin-
gradient approximation made simple. Phys. review letters wave energies, exchange parameters, and curie tempera-
77, 3865 (1996). ture. Phys. Rev. B 67, 012407 (2003).
66. Blöchl, P. E. Projector augmented-wave method. Phys.
82. Körmann, F., Dick, A., Hickel, T. & Neugebauer, J. Pres-
review B 50, 17953 (1994).
sure dependence of the curie temperature in bcc iron
67. Zimmermann, B. et al. Comparison of first-principles studied by ab initio simulations. Phys. Rev. B 79, 184406
methods to extract magnetic parameters in ultrathin films: (2009).
Co/pt (111). Phys. Rev. B 99, 214426 (2019).
83. Skomski, R. et al. Simple models of magnetism (Oxford
68. Halilov, S., Perlov, A., Oppeneer, P. & Eschrig, H. University Press on Demand, 2008).
Magnon spectrum and related finite-temperature mag-
netic properties: A first-principle approach. EPL (Euro- 84. Wood, M. A. & Thompson, A. P. Extending the accuracy
physics Lett. 39, 91 (1997). of the snap interatomic potential form. The J. Chem. Phys.
148, 241721 (2018).
69. Kurz, P., Förster, F., Nordström, L., Bihlmayer, G. &
Blügel, S. Ab initio treatment of noncollinear magnets 85. Zuo, Y. et al. Performance and cost assessment of ma-
with the full-potential linearized augmented plane wave chine learning interatomic potentials. The J. Phys. Chem.
method. Phys. Rev. B 69, 024415 (2004). A 124, 731–745 (2020).
70. Sandratskii, L. Noncollinear magnetism in itinerant- 86. Wood, M. A., Cusentino, M. A., Wirth, B. D. & Thomp-
electron systems: theory and applications. Adv. Phys. son, A. P. Data-driven material models for atomistic
47, 91–160 (1998). simulation. Phys. Rev. B 99, 184305 (2019).

13/14
87. Deng, Z., Chen, C., Li, X.-G. & Ong, S. P. An elec-
trostatic spectral neighbor analysis potential for lithium
nitride. npj Comput. Mater. 5, 1–8 (2019).
88. Goryaeva, A. M., Maillet, J.-B. & Marinica, M.-C. To-
wards better efficiency of interatomic linear machine
learning potentials. Comput. Mater. Sci. 166, 200–209
(2019).
89. Adams, J. J., Agosta, D., Leisure, R. & Ledbetter, H.
Elastic constants of monocrystal iron from 3 to 500 k. J.
applied physics 100, 113530 (2006).

Acknowledgements
Sandia National Laboratories is a multimission laboratory
managed and operated by National Technology & Engineer-
ing Solutions of Sandia, LLC, a wholly owned subsidiary
of Honeywell International Inc., for the U.S. Department of
Energy’s National Nuclear Security Administration under con-
tract DE-NA0003525. This paper describes objective techni-
cal results and analysis. Any subjective views or opinions that
might be expressed in the paper do not necessarily represent
the views of the U.S. Department of Energy or the United
States Government. AC acknowledges funding from the Cen-
ter for Advanced Systems Understanding (CASUS) which is
financed by the German Federal Ministry of Education and
Research (BMBF) and by the Saxon State Ministry for Sci-
ence, Art, and Tourism (SMWK) with tax funds on the basis
of the budget approved by the Saxon State Parliament.

Author contributions statement

AC, MAW, MPD and JT performed the DFT calculations.
JBM, MCM, JT and MAW generated the Database of con-
figurations. JT implemented the extended spin Hamiltonian
and the magnetic pressure computation in LAMMPS, and
parametrized it on first-principles calculations. MAW and
SN trained the SNAP potential. JT and SN performed the
magneto-static calculations. SN, APT, MAW and JT per-
formed the magneto-dynamics calculations. All authors partic-
ipated in conceiving the research and writing the manuscript.

Competing interests
The authors declare no competing interests.

14/14