2026 Volume 39 Issue 4
Article Contents

Shuaibin Li, Musen Li, Yongle Li. Phase Transition of Fe-Co Alloy under Magnetic Field with Machine Learning Force Field[J]. Chinese Journal of Chemical Physics, 2026, 39(4): 529-536. doi: 10.1063/1674-0068/cjcp2505066
Citation: Shuaibin Li, Musen Li, Yongle Li. Phase Transition of Fe-Co Alloy under Magnetic Field with Machine Learning Force Field[J]. Chinese Journal of Chemical Physics, 2026, 39(4): 529-536. doi: 10.1063/1674-0068/cjcp2505066

Phase Transition of Fe-Co Alloy under Magnetic Field with Machine Learning Force Field

通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Figures(4)  /  Tables(1)

Article Metrics

Article views(190) PDF downloads(4) Cited by(0)

Access History

Phase Transition of Fe-Co Alloy under Magnetic Field with Machine Learning Force Field

Abstract: The phase transition behavior of alloys can be modulated by magnetic fields, which affects the phase transition process by arranging magnetic moments, thereby changing the thermodynamic and kinetic behaviors of the phase transition. In this work, based on our previous works, we have successfully simulated the process of the phase transition of Fe-Co alloys combining molecular dynamics and spin-lattice dynamics, under a machine-learning force field, with magnetic interaction parameters obtained from the first-principles density functional theory, with and without external magnetic field. From the simulated results, we have demonstrated the mechanism of both its solid-solid and solid-liquid phase transitions, under either 0 T or 10 T external magnetic field.

    • Magnetic alloys have long been employed in high-performance magnetic devices and aerospace components due to their advantageous properties, such as high saturation magnetization, mechanical strength, and thermal stability [1−4]. Among these alloys, the magnetic ordering exhibits complex interactions with structural phase transitions in binary Fe-Co alloy systems with near-equiatomic compositions, particularly those involving transformations between body-centered cubic (BCC) and face-centered cubic (FCC) phases [5, 6]. An external magnetic field adds further complexity to these phase transition dynamics. Magnetic field influences phase stability and transformation pathways by modulating spin configurations and altering free energy landscapes. As a result, investigating the impact of magnetic fields on phase transitions at the atomic scale is essential for optimizing the performance of magnetic alloys in extreme environments.

      Molecular dynamics (MD) simulations have become an effective tool for the study of phase transition processes at the atomic level [7−9]. Traditional MD relies on force fields with empirical functional form to describe interatomic interactions; however, these functions frequently underestimate the complex magnetic and spin interactions intrinsic to Fe-Co alloys, particularly under the external magnetic fields. In recent years, machine-learning force fields, such as those based on the atomic cluster expansion (ACE) framework [10, 11], have emerged as promising alternatives. These models offer enhanced accuracy and transferability, making them well-suited for studying complex phenomena in magnetic materials. Advancements in the integration of spin-lattice dynamics [12] with conventional molecular dynamics have led to novel avenues for investigating the impact of spin and magnetic fields on phase transition processes [13].

      The present study is an extension of previous research conducted within our group, which focused on the solidification nucleation of pure iron under an external magnetic field [14] or the solidification behavior of alloys without any external field [15]. A central challenge in our work is to accurately describe and simulate the force field of alloys under an external magnetic field. In this study, we have effectively established a machine-learning force field within ACE framework, containing magnetic-interaction parameters obtained by fitting first-principles density functional theory (DFT) calculations. Using this machine-learning magnetic force field, we have simulated the phase transition process of Fe-Co alloys under an external magnetic field. And in this work, we have investigated the influence of the magnetic field on the phase transformation of Fe-Co alloys via the MD simulations with consideration of magnetic interactions, in which the spin information is parameterized for each atom. And the Fe-Co alloy under simulation is equiatomic, and the simulated phase transition contains both melting and the BCC-FCC solid phase transition.

    II.   METHODS
    • The magnetic behavior of the simulated system, both the atomic-atomic and response to external magnetic field, is simulated by spin-spin interaction and the spin-external magnetic field interactions [16]. In this study, the model of Zeeman effect is employed to simulate the response of the atomic spin to the external magnetic field, and the effective interaction term of the Heisenberg model, with lattice points assigned to the position of each atom, is used to simulate the magnetic interactions between the atoms [13].

      In Eq.(1), the first term in the right-hand side is the contribution of the Hamiltonian containing atomic spins, named spin Hamiltonian, and its functional form is shown in Eq.(2). The second term in Eq.(1) represents the sum of the kinetic energies of the particles, and the last term is the sum of the interatomic potentials, which is the machine-learning force field in this work. The last two terms of Eq.(1) are the same as those in traditional MD simulations using a given force field, while $ \mathcal{H}\mathrm{_{mag}} $ needs to assign the spins for the magnetic atoms and fit the magnetic interaction parameter J(rij) from first-principles calculations. The ${\mathcal{H}}_{\rm{mag}}$ in Eq.(2) is the Hamiltonian of N-interacting spins on a fixed lattice, where the first term is frequently referred to as the Zeeman term, corresponds to the interaction energy for spins under an external uniform magnetic field ${\boldsymbol{H}}_{\rm{ext}}$, and the second term represents the exchange interaction between spins. The parameterized equation for the exchange interaction is given by Eq.(3), which is parameterized by three coefficients, $ \alpha $, $ \delta $, and the dimensionless constant $ \gamma $. This functional form allows for the introduction of the cutoff radius ${R}_{\rm{c}}$, and for the calculation of the exchange interaction for a given spin i, it is only necessary to consider another spin j when ${r}_{ij} \lt {R}_{\rm{c}}$. The Heaviside step function, denoted by ${\Theta }\left({R}_{\rm{c}}-{r}_{ij}\right)$, is employed to specify this condition.

      The fitting of the magnetic exchange interaction parameter J(rij) contains two steps. First, the electronic structure calculations with spin interactions are performed using the spanish initiative for electronic simulations with thousands of atoms (SIESTA) [17], which employs norm-conserving pseudopotentials [18] and the Perdew-Burke-Ernzerhof (PBE) [19] generalized gradient approximation (GGA) functional. Secondly, the TB2J package [20] is utilized to calculate the magnetic interaction $ J $, including exchange and Dzyaloshinskii-Moriya (DM) interaction. TB2J employs the magnetic force theorem in the Green’s function formalism to extract exchange interaction parameters between spins, based on electronic structure results obtained from SIESTA. Subsequently, the nonlinear least squares method [21] is implemented to fit the parametric equations of the exchange interaction $ J $ to obtain the magnetic interaction parameters. Finally, MD simulations of the system are performed using the large scale atomic molecular massively parallel simulator (LAMMPS) [22], taking into account the spin effects and an applied magnetic field of 10 T. The electronic structure calculation systems using SIESTA are BCC-structured Fe, Co, and Fe-Co alloy in 2-atom single cells. The cutoff energy was chosen to be 200 Ry, the Monkhorst–Pack k-point grid was configured as 9 × 9 × 9, and the initial spin value was set to 3 μB. In this study, the double-zeta plus polarization (DZP) [23] basis sets from the SIESTA Pseudopotentials and Basis Set Database were employed.

      For fitting the machine learning force field, the system’s energy, virial, and force information were calculated using first-principles DFT, fulfilled by the Vienna ab initio simulation package (VASP) [24] with projection-based augmented wave pseudopotentials [18], and PBE [19] GGA functional. In all calculations, a plane-wave cutoff energy of 400 eV was used, with a force convergence criterion of 0.01 eV/Å, an energy convergence threshold of 10−6 eV, and Brillouin zone sampling performed using a 13 × 13 × 13 Monkhorst–Pack k-point grid. Subsequently, the PACE (performant implementation of the atomic cluster expansion) force field was obtained by fitting the data through a machine learning method based on the ACE framework, as implemented in the pacemaker code (https://github.com/ICAMS/python-ace). The fitting procedure adopted a regularized linear regression approach, in which the total energies, atomic forces, and virial stresses obtained from first-principles calculations were simultaneously fitted to a linear combination of ACE basis functions.

      The initial configuration of the MD-simulated system was derived by constructing BCC Fe-Co alloy supercells with periodic boundary conditions. The atomic ratio of Fe to Co was 1:1, and the MD simulations were performed using 10 × 10 × 10 (containing 2000 atoms) and 20 × 20 × 20 (containing 16000 atoms) supercells, respectively, with a lattice constant of 2.84 Å. Physical properties of this supercell, such as density, were simulated. The simulations were run under the isothermal-isobaric (NPT) ensemble: the system temperature was regulated through a Nosé-Hoover chain thermostat [25, 26], while pressure (1 atm) was regulated via an Andersen barostat [27]. Magnetic spins were incorporated using a Langevin thermostat [28], as described by the magnetic spins associated with the atoms. An external magnetic field of 10 T was applied to the simulated system in the z direction, and the initial spins of Co were set to 1.7 μB and Fe to 2.2 μB. The temperature rise process of Fe-Co alloy supercells from 1 K is simulated with a heating rate of 1 K/ps. The velocity-Verlet algorithm [29] was employed to solve the propagating classical equations of motion and to sample the real-time position and velocity of the particles. Equilibration simulations were executed at 1 K for 100 ps for each iteration before heating the system. The time step during the run was 0.1 fs. Additionally, for the 16000-atom system, 100 ps equilibration steps were conducted after every 500 K temperature during heating, before proceeding with further heating. The Visual Molecular Dynamics (VMD) [30] package and the Open Visualization Tool (OVITO) [31] package were utilized for structure monitoring and analysis of the various output data generated in MD simulations. The incorporation of the SPIN package [13] within the LAMMPS framework was crucial for the identification of magnetic parameters during magnetic field MD simulations.

    III.   RESULTS AND DISCUSSION
    • The electronic structures of Fe, Co, and Fe-Co primitive cells containing two-atom BCC structures were obtained with optimized atomic structures by first-principles calculations, respectively. After calculating the magnetic interactions, the results were parameterized and fitted using the trust region reflective algorithm [32]. The fitted results are displayed in FIG. 1 and collected in Table I. FIG. 1 shows that the maximum value of J for Fe-Co is larger than that of pure Fe or Co, and this causes larger magnetic parameters fitted from DFT than geometric averages of pure Fe-Fe and Co-Co. The cutoff radius ${R}_{\rm{c}}$ is incorporated into the parametric Eq.(3) through the Heaviside step function to reduce computational complexity. To ensure the accuracy, we have chosen the Rc by considering Ref.[13] and analyzing the decrease of magnetic interactions with distance within the alloy lattice, as shown in FIG. 1. In this work, ${R}_{\rm{c}}$ is set to 4 Å for Fe-Fe, 4.5 Å for Co-Co, and 5 Å for Fe-Co after test. And long-range magnetic exchange parameters are from optimized single-atom configuration. Table I shows the parameters of the magnetic interactions obtained from fitting the first-principles calculated results, denoted as fitted parameters, and for comparison, the parameters obtained by geometric mean parameters ($ {D}_{12}=\sqrt{{D}_{11}\times {D}_{22}} $) of literature values are also included, in which the parameters of Fe-Fe and Co-Co are taken from Ref.[13] and denoted as geometric mean parameters.

      Table I reveals a substantial discrepancy between the fitted parameters and geometric mean parameters. Firstly, the fitted coupling strength (α) of Fe-Co pair is 0.03904 eV, which is considerably higher than geometric mean value, 0.02380 eV. This indicates that the strength of the magnetic coupling between Fe-Co is underestimated once using geometric mean, and can only be properly considered via first-principles calculations. The reason would be the geometric mean of that parameter is homogeneous, but in essence the interaction between Fe and Co within condensed phase alloy is anisotropic. The second parameter, the interaction range (δ), is 2.820 Å through fitting, which is larger than the geometric mean, 1.707 Å. This suggests the Fe-Co magnetic interaction exists over a wider range, and is also hard to capture from simple geometric mean. The third parameter, nonlinear modulation factor γ, defined as the ratio of magnetic to electronic relaxation times, shows the fitted value of 0.470, substantially larger than the geometric mean, 0.0015. This discrepancy reflects that the Fe-Co magnetic interaction is long range and is characterized by a more complex coupling of orbital contributions to the electronic structure, which cannot be ascertained through simple empirical interpolation in the homogeneous systems. Actually, the parameters exhibit unique magnetic response properties in the heterogeneous alloy systems, so only the first-principles calculations can give the accurate estimation.

    • Given the magnetic interaction parameters are obtained, we performed the first round of MD simulation using a relatively smaller 10 × 10 × 10 supercell under an external magnetic field of 10 T. A phase transition curve, as shown in FIG. 2, was obtained from MD simulations of heating process. FIG. 2 (a) and (b) depict the system density as a function of temperature with a phase transition determined by a sudden change. Furthermore, the variation of the average potential energy per atom with temperature is also demonstrated in FIG. 2 (c) and (d). Using fitted parameters, MD simulations, shown in FIG. 2(a, c), revealed a transition point at 915±15 K. In contrast, simulations with geometric mean parameters, shown in FIG. 2(b, d), demonstrated a transition point at 940 K (±10 K). Structural analysis revealed a BCC to FCC phase transition (see Supplementary materials, SM). Both density curves show the density of the alloy changes gradually before the phase transition in the BCC phase, followed by a sudden jump according to higher packing efficiency for atoms in the FCC phase. And after the phase transition, density decreases with temperature again, with relatively larger fluctuation. But until the system is heated to 2500 K, there is no solid-liquid phase transition.

      This trend is also observed in the per-atomic potential energy plot. During the phase transition, both parameter sets give similar behavior of the density and per-atomic potential energy, with the difference in numbers. This deviation becomes larger in FCC phase, reflecting that when magnetic interaction is strong or an external magnetic field maintains system order, similar atomic structures can be yielded with both parameter sets. But when the thermal motion of the system becomes large, geometric mean parameters would underestimate magnetic interaction, leading to smaller change of values with larger fluctuations for both density and per-atomic potential. The absence of a critical point from a solid to a liquid state as the temperature increased was evidenced by the application of both fitted parameter and geometric average parameters in the MD simulations.

      Since the above MD simulation gives only a single critical point, intuitively it would stem from the insufficient spatial size of the simulation. To verify this deduction, heating is adopted on a large system, that is, a 20 × 20 × 20 supercell. The density and per-atomic potential energy changing with temperature are plotted in FIG. 3 (a) and (b), respectively. Since the fitted parameters are proved more reliable, we do not simulate the large system with geometric mean parameters. As illustrated in FIG. 3, all three physical properties exhibit two pronounced abrupt changes with increasing temperature, reflecting a two-stage phase transition. In FIG. 3(a), the system density initially decreases gradually, reflecting the solid-state characteristics of the first phase. At the first phase transition point, at about 1622±18 K, the density undergoes a sudden jump again, as observed in the simulations of small 10 × 10 × 10 system, indicating the solid-solid phase transition. Upon further heating, a second drop in density at 2395±10 K is observed. At this second critical point, the system undergoes a solid-liquid phase transition, which can be seen from the even smaller density compared with BCC solid phase. More confidently, a breakdown of long-range order is observed from the radial distribution function (RDF), as discussed below.

      FIG. 3(b) illustrates the variation of per-atom potential energy with temperature. In the three distinct temperature regions, unlike the non-linear feature of the density plot, the per-atomic potential energy increases approximately linearly. But at the critical points, pronounced jumps are observed, corresponding to structural rearrangement and melting, respectively. These energy transitions align closely with the density trends, shown in FIG. 3(a), providing further evidence that the system undergoes two distinct first-order phase transitions during the heating process. Through analyzing the heating plot of the large system, we verify the accuracy of our machine learning force field with the fitted parameters.

      In order to understand the phase transition mechanism of Fe-Co alloys under a magnetic field, we give the RDF, $ g\left(r\right) $ [33−35], for both small and large Fe-Co alloy systems at different temperatures during the heating process. FIG. 4(a–d) represent the MD RDFs for the small system, where FIG. 4(a, b) correspond to the MD RDFs with fitted parameters and FIG. 4(c, d) is the MD RDFs with geometric mean parameters. Furthermore, FIG. 4(e, f) show the MD RDFs with fitted parameters for large system.

      FIG. 4(a) shows the RDF curve at 800 K, a temperature far below the solid-solid phase transition, exhibiting periodic sharp peaks that indicate the long-range order in the BCC phase. In contrast, FIG. 4(b) shows the RDF curve at 950 K after the phase transition, where the peaks become broader and significantly attenuated. And the first peaks for the three different pairs of atoms, Co-Co, Co-Fe, and Fe-Fe are closer to each other, reflecting a more symmetric local environment after the phase transition. Except for the Fe-Co peak, both Co-Co and Fe-Fe peaks shift to the origin point after the phase transition, reflecting closer packing. The second peak of the three curves is significantly changed from that at 800 K, with a new shoulder peak at about 3 Å. Both the changes are typical features of the FCC structure. And although the system changes its symmetry, periodic peaks can still be recognized after the first phase transition during heating, revealing again that the MD phase transition for small system is solid-solid. This transition is attributed to entropic contributions, which make the FCC structure thermodynamically more favorable at high temperatures in Fe-Co alloys [36, 37]. FIG. 4(c, d) present the RDFs obtained using geometric mean parameters at the same temperatures. It can be observed that the results exhibit consistent trends, different only in numbers. And it is interesting that the 950 K RDF from geometric mean parameters cannot exhibit shoulder peak after the first peak, meaning the 950 K system has even higher symmetry. In conclusion, we can realize again the geometric mean parameters underestimate the anisotropy within the alloy.

      FIG. 4 (e) and (f) correspond to the RDF of the large system at temperature below (2300 K) and above (2600 K) the solid-liquid critical point, respectively. Since the system becomes larger, curves in both plots become smoother. FIG. 4(e) shows the RDF curve before the phase transition, during which a clear periodic structure can still be observed, indicating that the system remains in the solid state. And this FCC RDF is smoother and with less fluctuation than that from small system. Actually, the FCC RDF from small system looks more like that of liquid phase, reflected at the long range, the peaks become nearly smeared off. Furthermore, the RDF curve of this large system is consistent with the shape of the solid-state RDF of the small 10×10×10 system in the precession. The RDF after the solid-liquid phase transition is illustrated in FIG. 4(f). At this temperature, the RDF curve exhibits liquid-state structural characteristics, that is, long-range disorder. Notably, the RDF demonstrates that the Co-Fe interaction maintains a certain degree of structural correlation in the high-temperature liquid state, evidenced by the RDF of Co-Fe still exhibiting a certain peak at high temperature with a certain degree of order. This finding indicates that the magnetic field plays a structural stabilizing role in the alloy with respect to the Co-Fe interaction.

      From RDF analysis, we have successfully confirmed that the critical points obtained from MD simulations under external magnetic field are for solid-solid transition at lower temperature, and solid-liquid transition at higher temperature. And we can also confirm the structures found at each temperature region.

    IV.   CONCLUSION
    • In this work, we have successfully obtained the magnetic interaction parameters of binary Fe-Co alloy system by fitting first-principles DFT results. We then used the fitted parameters in combination with a machine learning force field to perform MD simulations of the Fe-Co alloy system. The phase transition of the Fe-Co alloy system under a 10 T external magnetic field was successfully simulated, exhibiting two phase transitions during heating: one is BCC-to-FCC solid-solid transition, and the other is solid-liquid transition. Our findings are consistent with classical Fe-Co phase transition theory. We have also found that only the 20×20×20 supercell is enough to accurately capture the heating process. By comparing the fitted parameters with the geometric mean parameters from pure Fe and Co, we have confirmed that the geometric mean parameters are useful for the simulation of equilibrium state, yielding slightly different structures, but are not reliable for simulation of phase transition processes. Finally, our results provide an efficient means for simulating magnetic materials or materials under magnetic fields.

      Supplementary materials: The atomic structures for both the initial structures for both 10×10×10 and 20×20×20 supercell, and structures before and after the phase transition of the 10×10×10 supercell.

    Figure (4)  Table (1) Reference (37)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return