betaFeedback

WorksPhysics and Astronomy

Accurate Machine Learning Predictions of Coercivity in High-Performance Permanent Magnets

Abstract

Increased demand for high-performance permanent magnets in the electric vehicle and wind turbine industries has prompted the search for cost-effective alternatives.Discovering new magnetic materials with the desired intrinsic and extrinsic permanent magnet properties presents a significant challenge to researchers because of issues with the global supply of rare-earth elements, material stability, and a low maximum magnetic energy product BHmax_{max}.While first-principle density functional theory (DFT) predicts materials' magnetic moments, magneto-crystalline anisotropy constants, and exchange interactions, it cannot compute coercivity (HcH_c).Although it is possible to calculate HcH_c theoretically with micromagnetic simulations, the predicted value is larger than the experiment by almost an order of magnitude, due to the Brown paradox.To circumvent these, we employ machine learning (ML) methods on an extensive database obtained from experiments, DFT calculations, and micromagnetic modeling.The use of a large dataset enables realistic HcH_c predictions for materials such as Ce-doped Nd2_2Fe14_{14}B, comparing favorably against micromagnetically simulated coercivities.Remarkably, our ML model accurately identifies uniaxial magneto-crystalline anisotropy as the primary contributor to HcH_c. With DFT calculations, we predict the Nd-site dependent magnetic anisotropy behavior in Nd2_2Fe14_{14}B, confirming that Nd 4g4g-sites mainly contribute to uniaxial magneto-crystalline anisotropy, and also calculate Curie temperature (TC_{C}).Both calculated results are in good agreement with experiment.The coupled experimental dataset and ML modeling with DFT input predict HcH_c with far greater accuracy and speed than was previously possible using micromagnetic modeling.Further, we reverse-engineer the inter-grain exchange coupling with micromagnetic simulations by employing the ML predictions.

Introduction

With the rapid advance of computational capabilities, there is considerable research interest in machine learning (ML) methods for predicting material properties using extensive databases1,2. Of specific interest is the remarkable speed of these techniques, which outperform traditional first-principle methods like density functional theory (DFT) by an order of magnitude. ML methods can deal with complex structures, and are desirable for discovering high-performance permanent magnet materials much needed for the electric vehicle and wind turbine industries. Although recent advances in first-principle methods such as DFT have enabled successful prediction of intrinsic properties, e.g., magnetic moments, magneto-crystalline anisotropy, and exchange interactions, the prediction of coercivity (HcH_{\mathrm{c}}) is a daunting task. Theoretically, HcH_{\mathrm{c}} can be computed by solving the phenomenological Landau–Lifshitz–Gilbert equation (LLGE) with micromagnetic simulations, but these simulations overestimate the experimental HcH_{\mathrm{c}} by an order of magnitude due to the Brown paradox3–5. Previous papers employing ML to predict the extrinsic properties6–9 have been limited exclusively to micromagnetic simulated materials.

ML requires datasets that include information about material properties such as crystal structure, micromagnetic grain size and boundaries, saturation magnetization (MsM_{\mathrm{s}}), the uniaxial magnetocrystalline anisotropy constant (KuK_{\mathrm{u}}), the exchange stiffness constant (AexA_{\mathrm{ex}}), and the Curie temperature (TCT_{\mathrm{C}}). To build predictive models based on our dataset, we utilize classical ML and artificial neural network (ANN) algorithms10–12. These models establish patterns and relationships between the independent and dependent (HcH_{\mathrm{c}}) material properties. They are trained on subsets of the known data and tested on the complementary subsets to assess model accuracy and reliability.

An extensive survey of the literature resulted in a dataset of 300 experimentally known materials (see SupplementaryTable I), to our knowledge the largest current experimental ML magnetic dataset. Our second dataset consists of 8770 micromagnetically computed permanent magnet materials. Various predictive techniques, including ML, statistical inference, and micromagnetic modeling (mumax3^{3} program) are applied to both datasets to predict and compare HcH_{\mathrm{c}}14,15. In experimental materials, we find standard non-linear models such as the decision tree (DT), extreme gradient boosting (XGB), and random forest (RF) produce excellent results with R2∼0.87R^2 \sim0.87 (where R2R^2 is a standard statistical measure of accuracy in regression), but tuning the XGB regressor improves the R2R^2 measure to 0.89. Most importantly, ML clearly demonstrates that HcH_c is related directly to KuK_u, weakly to AexA_{\mathrm{ex}}, and inversely to MsM_s.

We predict the HcH_c of cerium (Ce)-doped Nd2_2Fe14_{14}B 2 : 14 : 1 materials to demonstrate the complete pipeline enabled by the new ML toolkit coupled with ab initio calculations. First, for a pure neo-magnet, the site contribution to magnetocrystalline anisotropy is analyzed with DFT calculations showing that 4g-sites mainly contribute to the uniaxial magneto-crystalline anisotropy. Second, our computed TCT_C for a pure compound using Green’s function in the atomic sphere approximation (ASA) is in good agreement with experiment. Finally, we employ DFT-computed parameters in ML for predicting the HcH_c of Ce-doped compositions. The ML-predicted HcH_c matches with experiment, demonstrating that the ab initio computed input parameters and the ML methodology are sufficient to predict experimental HcH_c, even without access to experimental conditions and advanced internal structural properties. Conversely, the ML prediction for selected candidate materials is used to engineer their grain boundary size (GBS) and inter-grain coupling.

Micromagnetism

Micromagnetics is the study of the behavior of magnetic materials typically in the nanometer range. During its early formulation17,18, the field emphasized qualitative aspects of magnetism: the role of domain structures, domain walls, and magnetic vortices in ferromagnetic materials. The transition to computer simulation in micromagnetics was a significant advance, providing detailed examination of the forces at play inside a magnetic material19.

Naively, the estimate 2KuMs\frac{2K_u}{M_s} gives an upper bound on HcH_c3,20. However, this estimate disregards impurities and multi-grain structures in materials, leading to a gross overestimate of HcH_c from theory alone: the Brown paradox. Micromagnetic simulations which include demagnetization (shape anisotropy) are crucial for a good understanding of magnetic materials, but the interplay among these complex effects and the Brown paradox still hinder accurate predictions of HcH_c.

Theory

Magnetodynamics is described by a nonlinear partial differential equation for the spatio-temporal magnetization vector M(r,t)\mathbf{M}(\mathbf{r},t). The time evolution of M(r,t)\mathbf{M}(\mathbf{r},t) is given by a phenomenological Landau–Lifshitz–Gilbert (LLG) equation17,18

∂M(r,t)∂t=γ1+α2M(r,t)×Heff(r,t)−αγ1+α2M(r,t)×[M(r,t)×Heff(r,t)](1)\frac{\partial\mathbf{M}(\mathbf{r},t)}{\partial t}=\frac{\gamma}{1+\alpha^2}\mathbf{M}(\mathbf{r},t)\times\mathbf{H}_{\mathrm{eff}}(\mathbf{r},t)-\frac{\alpha\gamma}{1+\alpha^2}\mathbf{M}(\mathbf{r},t)\times\left[\mathbf{M}(\mathbf{r},t)\times\mathbf{H}_{\mathrm{eff}}(\mathbf{r},t)\right] \tag*{(1)}

Here, M(r,t)\mathbf{M}(\mathbf{r},t) is the unit vector describing the magnetization of the material with MsM_s as the saturation magnetic moment per unit volume, while Heff(r,t)\mathbf{H}_{\mathrm{eff}}(\mathbf{r},t), γ\gamma, and α\alpha respectively are the effective static magnetic field, the gyromagnetic ratio, and the damping parameter (quantifying the rate at which the magnetization relaxes back to equilibrium). In (1), the first term is the precession of the magnetic moment around the external magnetic field. The second term is the damping, which relaxes the magnetic moment to the equilibrium. For time-independent scenarios, such as the computation of a hysteresis loop, the α\alpha term is set to 0. Then, Heff=H+Hms+Hex+Ha\mathbf{H}_{\mathrm{eff}}=\mathbf{H}+\mathbf{H}_{\mathrm{ms}}+\mathbf{H}_{\mathrm{ex}}+\mathbf{H}_{\mathrm{a}}. The terms are as follows. H\mathbf{H} is the externally applied field, which is taken as a parameter. Hms\mathbf{H}_{\mathrm{ms}} is a long-range magnetic field

Hms(r)=14π∫∇∇′1∣r−r′∣⋅M(r′) dr′(2)\mathbf{H}_{\mathrm{ms}}(\mathbf{r})=\frac{1}{4\pi}\int\nabla\nabla'\frac{1}{\lvert\mathbf{r}-\mathbf{r}'\rvert}\cdot\mathbf{M}(\mathbf{r}')\,\mathrm{d}\mathbf{r}' \tag*{(2)}

corresponding to self-interaction of the induced magnetic field with the magnetization across the material21. The exchange field is the Laplacian of the magnetization, which is obtained from the classical Heisenberg model

Hex=2AexMs∇M(r,t).(3)\mathbf{H}_{\mathrm{ex}}=\frac{2A_{\mathrm{ex}}}{M_s}\nabla\mathbf{M}(\mathbf{r},t). \tag*{(3)}

Although this expression was originally deduced for localized spins, it is still valid for itinerant systems to the first order approximation. AexA_{\mathrm{ex}} is a measure of the strength of magnetic exchange interaction. HaH_a is the uniaxial anisotropy term, written as

Ha=2Ku1Ms(u⋅M(r,t))u+2Ku2Ms(u⋅M(r,t))3u(4)\mathbf{H}_{\mathrm{a}}=\frac{2K_{u1}}{M_s}\left(\mathbf{u}\cdot\mathbf{M}(\mathbf{r},t)\right)\mathbf{u}+\frac{2K_{u2}}{M_s}\left(\mathbf{u}\cdot\mathbf{M}(\mathbf{r},t)\right)^3\mathbf{u} \tag*{(4)}

where Ku1K_{u1} and Ku2K_{u2} are anisotropy constants, and u\mathbf{u} indicates the direction of the anisotropy vector, making it easier for the magnetization to align with the direction of u\mathbf{u} in the case of ferromagnetic materials. Most papers only envisage a single constant: Ku1=KuK_{u1}=K_u and Ku2=0K_{u2}=0. Other terms, such as the Dzyaloshinskii-Moriya interaction (DMI), were neglected due to the lack of experimental data for current materials and the relatively weak effect in modern magnetic materials22.

In order to estimate fundamental magnetic parameters in cases of partial experimental information, the following relations were utilized. AexA_{\mathrm{ex}} is inversely proportional to the lattice constant (aa) as given by

Aex=JS2an,(5)A_{\mathrm{ex}}=\frac{JS^2}{a}n, \tag*{(5)}

where SS is the spin quantum number, and nn is the number of magnetic ions per unit cell. Equivalently, AexA_{\mathrm{ex}} can be expressed approximately in terms of TCT_C as

Aex∼3TC2za.(6)A_{\mathrm{ex}}\sim\frac{3T_C}{2za}. \tag*{(6)}

Here zz is the number of the nearest neighbors of a magnetic ion. Finally, the equation

BHmax⁡=μ0Ms24(7)BH_{\max}=\frac{\mu_0M_s^2}{4} \tag*{(7)}

is used to estimate MsM_s from the vacuum permeability μ0\mu_0 and the maximum energy product BHmax⁡BH_{\max}, which is given by the maximum product of the magnetic flux density and the magnetic field strength at any point on the hysteresis loop23.

The micromagnetic equations are solved in the continuum approximation M(r,t)=Msm(r,t)\mathbf{M}(\mathbf{r},t)=M_s\mathbf{m}(\mathbf{r},t)24–26. Heff\mathbf{H}_{\mathrm{eff}} is deduced from the magnetic free energy functional F[m]F[\mathbf{m}] as

Heff=−1μ0MsδF[m]δm\mathbf{H}_{\mathrm{eff}}=-\frac{1}{\mu_0M_s}\frac{\delta F[\mathbf{m}]}{\delta\mathbf{m}}
F[m]=∫V[Aex(∇m)2−μ0M⋅Hex−Ku(m⋅u)2−μ0M⋅H−μ02M⋅Hdem+fDMI(m)+⋯ ] d3r,(8)F[\mathbf{m}]=\int_V\left[A_{\mathrm{ex}}(\nabla\mathbf{m})^2-\mu_0\mathbf{M}\cdot\mathbf{H}_{\mathrm{ex}}-K_u(\mathbf{m}\cdot\mathbf{u})^2-\mu_0\mathbf{M}\cdot\mathbf{H}-\frac{\mu_0}{2}\mathbf{M}\cdot\mathbf{H}_{\mathrm{dem}}+f_{\mathrm{DMI}}(\mathbf{m})+\cdots\right]\,d^3\mathbf{r}, \tag*{(8)}

where fDMIf_{\mathrm{DMI}} is DMI. Additional terms may be added as needed to count for additional interactions. The F[m]F[\mathbf{m}] is minimized with respect to m\mathbf{m} using the steepest descent algorithm as implemented in the micomagnetic program26. Then the hysteresis loop is obtained by evaluating m\mathbf{m} in each equilibrium magnetic state for different values of the external applied magnetic field.

Experimental materials

An experimental database of 300 magnetic materials was assembled[27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57] [20, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81]. A material was chosen if KuK_u, MsM_s, and HcH_c were explicitly included in the experimental results. For the R2_2Fe14_{14}B family (where R is a rare-earth element) and some other materials, only the experimental anisotropy field (HaH_a) is available, so the estimate Hc∼Ha/4H_c\sim H_a/4 was used82.

The exchange strength AexA_{\mathrm{ex}} was determined in one of three different ways. The first was a direct reporting of the stiffness exchange, which was rarely available. For some materials, the value was interpolated from the values for similar alloys and compounds. The third method derived AexA_{\mathrm{ex}} from the Curie-stiffness relation in Eq. (6), using the experimental Curie temperature TCT_C and known lattice constant83. This latter method was used for R2_2Fe14_{14}B, including Ce-doped Nd2_2Fe14_{14}B71, binary alloys69, SmCo5_5, 1 : 5 compositions66, La/Pr/Co-doped hexaferrites61, and other element doped hexaferrites[56, 57] [60, 84].

Curie-stiffness relation in Eq. (6), using the experimental Curie temperature TCT_C and known lattice constant83. This latter method was used for R2_2Fe14_{14}B, including Ce-doped Nd2_2Fe14_{14}B71, binary alloys69, SmCo5_5, 1 : 5 compositions66, La/Pr/Co-doped hexaferrites61, and other element doped hexaferrites[56, 57] [60, 84].

There was significant variation in the reported HcH_c values, depending on the experimental conditions and fabrication methods. For example, the atmospheric composition during the annealing of iron (Fe) magnets affected HcH_c as oxygen modified the material composition during the hardening process. Oxygenated crystalline defects, such as nucleation and pinning of the domain walls, play critical roles in the HcH_c mechanism. The nucleation field85 (the magnetic field at which the atomic spin ceases to align along the magnetic easy axis) lowers the measured HcH_c, while pinning does the opposite.

Cooling rates have significant effects on the material grain size[27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57] [20, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81]. Several dozen materials (most notably Nd and Fe based ferromagnets) do have grain size recorded in the database. Grain composition, as visualized in Figure 1, is known to be heavily predictive of a material’s HcH_c86. However, correlating the grain measurements from different sources made it apparent that a single number cannot completely represent the grain structure, and that the particular measurement techniques employed may induce additional inaccuracies. In any material with multiple measurements for a property, a representative median was selected from a specific experiment (defaulting to papers with more comprehensive materials property estimates). This choice was made before the application of statistical techniques, to ensure that no bias contaminated the dataset.

A magnetic cuboid showing different micromagnetic grains

Figure 1. A magnetic cuboid of size 128×128×128 nm3128\times128\times128\ \mathrm{nm}^3 showing different micromagnetic grains. The colors within each grain refer to different magnetization directions. The snapshot of spin texture is captured during the micromagnetic simulation.

For the better training of ML networks, an additional supplementary database was computed, using micromagnetic modeling with HmsH_{\mathrm{ms}}, HexH_{\mathrm{ex}}, HH, HdemH_{\mathrm{dem}}, and HaH_a in the HeffH_{\mathrm{eff}}. Using the uniaxial approximation, AexA_{\mathrm{ex}}, KuK_u, and MsM_s were supplied as the inputs for stiffness exchange, magnetocrystalline anisotropy, and saturation magnetization. The input parameters were uniformly sampled from the cuboid determined by the ranges of AexA_{\mathrm{ex}}, KuK_u, and MsM_s in the 300 experimental materials. There are then two measures of HcH_c: the experimentally measured coercivity Hc(exp)H_c(\mathrm{exp}), possessed only by the experimental materials, and the computationally determined HcH_c, calculated for both the hypothetical and the experimental materials.

For micromagnetic modeling, simple structures were chosen for the magnetic samples. Due to the limited grain structure data, for the majority of the database a single structure composed of multiple grains would have been chosen for the materials. This would have induced scaling on the HcH_c, which instead may be modeled directly with ML. As no grains or boundaries were involved in the computation, a magnetic cube of 32×32×3232\times32\times32 at nm3\mathrm{nm}^3 scale was found to be sufficient to avoid loss of precision in the uniform uniaxial anisotropy alignment. For Ce-substituted compositions, we used simulation cells of size 128×128×128 nm3128 \times128 \times128\ \mathrm{nm}^{3} for grain boundary engineering as shown in Figure 1.

Micromagnetic results

We compute the HcH_c for experimentally known systems, and list the comparison of the calculated and experimental values for selected materials in Table I. Theoretically, the upper limit for HcH_c is the anisotropy field87

Ha=2KuMs,H_a = \frac{2K_u}{M_s},

however, the true experimental value is an order of magnitude smaller due to the uncertainty in the coercivity mechanism in permanent magnets, commonly known as the Brown paradox3. Experimentally measured HcH_c values fit well with the empirical Kronmüller equation88 Hc=cHa−Neff4πMsH_c = cH_a - N_{\mathrm{eff}}4\pi M_s89,90, where cHacH_a is the field to nucleate a reverse domain, and Neff4πMsN_{\mathrm{eff}}4\pi M_s is the demagnetization field, with cc and NeffN_{\mathrm{eff}} being renormalization factors. For example, Nd2_2Fe14_{14}B (sintered) has c∼0.25(0.37)c \sim0.25(0.37) and Neff∼0.26(1)N_{\mathrm{eff}} \sim0.26(1).89,90

The micromagnetically simulated HcH_c values differ from the experimental values by a factor of up to ∼5\sim5. Interestingly, these values are very similar to the experimental HaH_a. Moreover, for some materials, the use of the estimated values for MsM_s, AexA_{\mathrm{ex}}, and KuK_u using the empirical relations as discussed in Eqs. (5)–(7) will result in additional error. In general, HcH_c depends non-linearly on grain size and domain wall width or particle size in magnetic materials91–95. In modern manufacturing, grain sizes are larger relative to domain sizes. This is especially common in neo-magnets, leading to a decrease of HcH_c with increasing grain size96, where higher grain surface area hosts more defects. Grain boundary size also affects the demagnetization factor97, further reducing HcH_c.

The experimental features for 211 of the 300 materials in the dataset are known, while the remaining 90 materials require the use of the aforementioned theoretical models to determine AexA_{\mathrm{ex}} from experimental results. These materials show larger discrepancies in the micromagnetically predicted HcH_c, which is worth investigating both experimentally and theoretically. We show a comparison of these results (for selected key magnetic materials) in Table I. As all micromagnetically predicted HcH_c values are overestimated, the only meaningful comparison is obtained with a scaled coercivity cHccH_c, which for c=0.25c = 0.25 fits well with the experimental values obtained for the 210 materials. The cHccH_c estimate is generally a good match for 2 : 14 : 1 compositions, except for Co-based Nd2_2Co14_{14}B and Gd2_2Co14_{14}B. However, the appropriate scaling factor may vary for different compositions. Overall, the HcH_c variation with independent features is similar in both theory and experiment, although there are some exceptions such as La2_2Fe14_{14}B. For 1 : 5 compositions, a similar trend is evident.

Machine Learning

Classical machine learning algorithms

Machine learning (ML) encompasses a variety of advanced statistical techniques. It creates a correspondence between a space of independent variables, XX, and a space of dependent variables, YY, by taking a ground truth function fo:Xo→Yf_o : X_o \to Y representing a sequence of observations on a limited subset Xo⊂XX_o \subset X of the data to construct a more general function f:X→Yf : X \to Y which extends the function fof_o. This ground truth function may also be represented as a list of observations {xi↦yi}i⊂Xo×Y\{x_i \mapsto y_i\}_i \subset X_o \times Y. The method by which ff is constructed from fof_o is referred to as ML, and it is written symbolically as f=ML(fo)=ML({xi↦yi}i)f = ML(f_o) = ML(\{x_i \mapsto y_i\}_i), where ML is optionally subscripted to indicate the algorithm or hyperparameters in use.

The goal of ML is to represent the structure underlying fof_o as ff. This is usually quantified by splitting the observations into a training set fo′f'_o and testing set fo∗f^*_o so that fo′∩fo∗=∅f'_o \cap f^*_o = \varnothing. Then the function is constructed relative to a norm, termed a loss function, so that ∥fo′−ML(fo′)∥\lVert f'_o - ML(f'_o)\rVert is minimized in some appropriately defined subspace of X→YX \to Y to avoid overfitting. The quality of the fit with f=ML(fo′)f = ML(f'_o) is measured by evaluating ∥fo∗−f∥\lVert f^*_o - f\rVert using the same class of norms. In this paper, all ML algorithms use the standard Euclidean norm.

We use the scikit-learn library for the Python programming language for predictive data analysis98. The dataset (both experimental and theoretical) is split into a 70 : 30 train-to-test data set ratio for the model training. Performance does not change drastically by modifying the split ratio to 80 : 20. The classical linear ML models used are:

  1. Linear regression;

  1. Lasso regularization;

  1. Ridge regularization.

The non-linear classical ML models used are:

  1. Decision tree (DT) regression;

  1. Random forest (RF) regression;

  1. Gradient boost (GB) regression;

  1. XGBoost (XGB) regression.

In ML training with cross-validation, the training dataset is further divided into two parts - the training and the validation datasets. Most ML models have arbitrary hyperparameters, which are not modified directly during training. The validation dataset allows a meta-training of the ML models by random or grid search on the training data and evaluation on the validation set without contaminating the testing set by predicting an ML hyperparameter choice for test set evaluation. The most relevant cross-validation method is k-fold validation. The dataset is randomly split into k disjoint subsets, then one set is chosen as the validation dataset, while the others are chosen as training datasets.

MaterialMsexpM_s^{\mathrm{exp}} (MA/m)AexexpA_{\mathrm{ex}}^{\mathrm{exp}} (pJ/m)KuexpK_u^{\mathrm{exp}} (MJ/m3^3)HcexpH_c^{\mathrm{exp}} (T)Hcmumax3H_c^{\mathrm{mumax3}} (T)cHccH_c (T)HcdiffH_c^{\mathrm{diff}} (T)
La2_2Fe14_{14}B1.097.411.40.502.110.530.03
Ce2_2Fe14_{14}B0.936.001.700.543.140.790.25
Pr2_2Fe14_{14}B1.247.944.661.546.461.620.08
Nd2_2Fe14_{14}B1.278.234.651.246.251.560.32
Gd2_2Fe14_{14}B0.719.330.850.232.400.600.37
Tb2_2Fe14_{14}B0.568.786.132.9321.625.412.48
Dy2_2Fe14_{14}B0.569.424.241.7014.793.701.99
Ho2_2Fe14_{14}B0.648.142.420.757.261.821.07
Lu2_2Fe14_{14}B0.937.661.210.292.300.580.28
Y2_2Fe14_{14}B1.128.011.460.192.160.540.35
Th2_2Fe14_{14}B1.126.771.460.172.100.530.36
La2_2Co14_{14}B0.791.501.190.343.060.770.42
Pr2_2Co14_{14}B1.041.515.202.509.432.36-0.14
Nd2_2Co14_{14}B1.081.512.423.694.221.05-2.64
Gd2_2Co14_{14}B0.231.521.030.309.482.372.07
Y2_2Co14_{14}B0.851.521.190.342.840.710.37
SmCo5_50.861.201.727.5038.809.702.20
YCo5_50.787.115.503.9013.333.33-0.57
LaCo5_50.716.476.305.2516.894.22-1.03
CeCo5_50.605.136.405.7020.665.170.57
PrCo5_50.946.968.105.3216.084.02-1.30
NdCo5_50.937.090.240.150.470.12-0.03

Table 1. Comparison of calculated and experimental HcH_c for different rare-earth permanent magnets. For realistic comparisons, the mumax3^3 values are scaled by a factor of c=0.25c = 0.25 in column cHccH_c. The HcdiffH_c^{\mathrm{diff}} column presents the difference between cHccH_c and HcexpH_c^{\mathrm{exp}}.

Artificial neural networks

An artificial neural network (ANN) is suitable for high complexity problems such as HcH_c prediction. Figure 2 shows a schematic diagram (a portion used in the calculations) for an ANN consisting of an input layer (6 neurons), two hidden layers (containing 4 and 3 neurons respectively), and the output layer with one neuron. A non-linear activation function leaky-relu [f(z)=max⁡(0,z)+0.01∗min⁡(0,z)f(z) = \max(0,z) + 0.01 * \min(0,z)] is used between the input and hidden layers to counter typical convergence problems found in small ANNs. We used the state of the art gradient descent benchmark optimizer Adam99, as implemented in tensorflow100 in the Keras API101.

Schematic diagram showing an artificial neural network

Figure 2. Schematic diagram showing an artificial neural network (only a portion used in the calculations) with 6 neurons in the input layer, 4 and 3 neurons in the second and the third deep (hidden) layers, and a neuron for regression in the output layer. Two different ANN’s were used in actual calculations, as discussed in the main text.

Performance comparison

For the regression model, given an ML fit, we use two common statistical measures of error applied to the HcH_c predictions: (i) Mean Square Error (MSE):

MSE=1N∑iN(yi−ypredi)2(9)\mathrm{MSE} = \frac{1}{N}\sum_i^N (y_i-y_{\mathrm{pred}_i})^2 \tag*{(9)}

which is also used for training all ML models, and (ii) R-squared (R2R^2):

R2=1−MSEV(y)(10)R^2 = 1 - \frac{\mathrm{MSE}}{V(y)} \tag*{(10)}

where V(y)V(y) is the variance written as

V(y)=⟨(yi−⟨y⟩)2⟩(11)V(y) = \left\langle\left(y_i-\langle y\rangle\right)^2 \right\rangle \tag*{(11)}

with ⟨y⟩\langle y\rangle as the mean of the predicted values. Equivalently, we may write R2=1−RSS/TSSR^2=1-RSS/TSS, where RSSRSS is the Residual Sum of Squares RSS=∑iN(yi−ypredi)2RSS=\sum_{i}^{N}(y_i-y_{\mathrm{pred}i})^2, and TSSTSS is the Total Sum of Squares TSS=∑iN(yi−⟨y⟩)2TSS=\sum_{i}^{N}(y_i-\langle y\rangle)^2. The loss function is usually written as a sum of the MSE and the first or second-order norms of the target function parameters. This provides convergence of the model, and penalizes overfitting for parameterized machine learning.

Results and discussion

Machine learning on experimental data

Here, we explore the experimental data and the relation between target and independent variables. Figure 3 visualizes HcH_c as a function of KuK_u, MsM_s, and AexA_{\mathrm{ex}}. The actual values of the variables differ by several orders of magnitude. They are scaled using the sci-kit minmaxscalar in which a dimensionless scaled feature variable xscaledx_{\mathrm{scaled}} in the range [0,1] is obtained using the following relation:

The distribution of $H_c$ as a function of $M_s$, $A_{\mathrm{ex}}$, and $K_u$

Figure 3. The distribution of HcH_c as a function of MsM_s, AexA_{\mathrm{ex}}, and KuK_u in experimental magnetic materials. The physical quantities are rendered as normalized and dimensionless. The color map shows the magnitude of HcH_c.

xscaled=x−xmin⁡xmax⁡−xmin⁡,(12)x_{\mathrm{scaled}}=\frac{x-x_{\min}}{x_{\max}-x_{\min}}, \tag*{(12)}

where xmin⁡x_{\min} and xmax⁡x_{\max} are minimum and maximum values of the feature variable xx. Figure 3 shows an increase of HcH_c with KuK_u, a decrease with MsM_s, and very little variation with AexA_{\mathrm{ex}}.

ModelR2R^2MSE
Linear regression0.620.64
Lasso0.381.07
Ridge0.630.64
Lasso-CV0.610.68
Ridge-CV0.630.66
Elasticnet0.620.66
DT regressor0.890.19
DT pruned0.840.27
RF regressor0.870.23
gradient boosting (GB) regressor0.800.34
Tuned GB regressor0.870.22
XGB regressor0.870.23
Tuned XGB regressor0.890.18
ANN (Adam)0.640.62
light gradient boosted machine (LGBM)0.700.60
Fine tuned ANN0.850.25

Table 2. Comparison of R2R^2 and MSE metrics in different ML models on test data sets for experimental materials. The non-linear regression models show better values of R2R^2 and MSE than the linear models. The error is evaluated over the difference of the logarithm of the coercivity in T.

In gradient-based algorithms, such as ANN, additive relationships are easier to model than multiplicative ones, so the independent and dependent variables are logarithmically scaled. Accordingly, the measures of model performance are given in terms of the logarithm of the HcH_c. As a result of this change, the absolute measure of R2R^2 is increased by ∼0.01\sim0.01 for all gradient-based models.

Figure 4 shows a comparison of ML and experimental HcH_c for the test dataset. The XGB-tuned model shows better performance than ANN-Adam. Table II shows the performance of various ML models. The linear models show poor performance, with R2≪1R^2 \ll1. This demonstrates the complex non-linear relation between the dependent and independent variables. The non-linear models DT, XGB, and RF perform better, with R2R^2 values of 0.89, 0.87, and 0.87, respectively. These values still witness a decrease from the R2R^2 scores of more than 0.9 on the training dataset.

ML predicted and experimental coercivities of magnetic materials obtained with tuned XGB and ANN

Figure 4. ML predicted and experimental coercivities of magnetic materials obtained with tuned XGB (top) and ANN (bottom). Selected key materials are labeled in the figure. Non-linear regressors all yield similar results.

The ANN model has R2R^2 scores that are similar to other classical models (see Table II). The ANN architecture using (dense layers=64, activation function=leaky_relu, dense layers=32, activation function=leaky_relu, dense layers=16, activation function=leaky_relu, and output layer=1, and learning rate=0.007, epochs=1000, batch size=24) showed similar R2R^2 scores to RF. This weak performance of the ANN architecture is attributed to the smallness of the dataset size.

Feature importance. The model performance can be further validated from the dependence of HcH_c on the independent features. We compute the importance of RF features by utilizing the model-agnostic interpretive features of scikit-learn on the RF and XGB regression models10,2. The feature importance is computed by the

Gini index

Gini index=1−∑ipi2(13)\text{Gini index}=1-\sum_i p_i^2 \tag*{(13)}

where pip_i is the probability of class ii in the data. Figure 5 depicts the relative importance of the independent parameters for HcH_c, as computed by the XGB model. ML accurately identifies KuK_u as the leading contributing feature to HcH_c. More surprisingly, the effect of MsM_s is about four times smaller than that of KuK_u. The uneven distribution of magnetic materials (shown in Figure 3), and a constant noise level, may explain the unequal weighting of KuK_u and MsM_s. Finally, we note that AexA_{\mathrm{ex}} has a comparatively minimal effect on HcH_c.

Feature importances in the XGB model

Figure 5. Histogram of important features contributing to HcH_c in the XGB model. The KuK_u clearly leads, followed distantly by MsM_s and AexA_{\mathrm{ex}}.

Machine learning prediction from micromagnetic data

In order to explore the relation between HcH_c and the independent variables, we used 8770 micromagnetic simulation-generated data points as a training set. The correlation between variables is given by Pearson’s correlation coefficient, which is defined as

r=∑(X1−⟨X1⟩)(X2−⟨X2⟩)∑(X1−⟨X1⟩)2∑(X2−⟨X2⟩)2(14)r=\frac{\sum(X_1-\langle X_1\rangle)(X_2-\langle X_2\rangle)}{\sqrt{\sum(X_1-\langle X_1\rangle)^2\sum(X_2-\langle X_2\rangle)^2}} \tag*{(14)}

where XiX_i and ⟨Xi⟩\langle X_i\rangle denote the variable ii and its average value. A positive (or negative) value of rr indicates a positive (or negative) correlation between the two variables: the higher the rr-value, the higher the correlation. Figure 6 shows a correlation heatmap for the variables. The distribution for the 8770-data set is right-skewed, as shown in Figure 7, with a peak around 0.2–0.4 T. It is similar to the distribution for the experimental materials (not shown here): most of the materials have values around the peak, and a few materials, such as 1 : 5 compositions, have larger values of 2–6 T. For better ML training, logarithmic transformation is appropriate for skewed data, which is employed in our model fits.

Heat map showing the correlation between $H_c$, $A_{\rm ex}$, $K_u$, and $H_c$

Figure 6. Heat map showing the correlation between the HcH_c, AexA_{\rm ex}, KuK_u, and HcH_c. The color palette corresponds to the correlations between the variables. HcH_c is strongly correlated with KuK_u, negatively correlated with MsM_s, and weakly correlated with AexA_{\rm ex}. MsM_s shows a slightly negative correlation with KuK_u and AexA_{\rm ex}. KuK_u shows a slightly positive correlation with AexA_{\rm ex}.

Distribution of $H_c$ in the training set of 8770 data points

Figure 7. Distribution of HcH_c in the training set of 8770 data points. The yy-axis is scaled by 10310^{3}. The data points are right-skewed.

Significantly, the strongest (absolute) correlation is observed between HcH_c and KuK_u at 0.66, followed by MsM_s with a correlation of −0.62, and then by AexA_{\mathrm{ex}} at 0.2. This pattern suggests a high dependence of HcH_c on KuK_u. The negative correlation between HcH_c and MsM_s aligns with the theoretical relationship Hc∝Ms−1H_c \propto M_s^{-1}. Similarly, the correlation coefficient (rr) between HcH_c and KuK_u or AexA_{\mathrm{ex}} validates the proportionalities Hc∝KuH_c \propto K_u and Hc∝AexH_c \propto A_{\mathrm{ex}}. These findings are consistent with the trends observed in actual materials, as illustrated in Figure 3.

Next, we discuss the model performance in micromagnetic simulation data, as given in Table III. Both linear and non-linear models perform well, as the R2R^2 score is above 0.9. This is expected in a linear regression model where multivariate analysis demonstrates very little multicollinearity in randomly generated independent variables. The multicollinearity can be measured with the variance inflation factor (VIF) VIFα=(1−Rα2)−1VIF_\alpha=(1-R_\alpha^2)^{-1}, where

ModelR2R^2MSEMSE
Linear regression0.950.22
Lasso0.542.13
Ridge0.950.22
Lasso-cv0.950.25
Ridge-cv0.950.22
Elasticnet0.940.26
DT regressor0.970.13
DT pruned0.970.15
RF regressor0.980.08
GB regressor0.980.09
Tuned GB regressor0.980.08
XGB regressor0.980.98
Tuned XGB regressor0.980.08
ANN0.980.09

Table III. TABLE III. mumax⁡3\operatorname{mumax}^{3} data on hypothetical materials: Comparison of R2R^2 and MSE of HcH_c for different ML models. All the errors are evaluated over the ln⁡\ln of the HcH_c in T. The fine-tuned model hyperparameters are given in Appendix A.

Rα2R_{\alpha}^{2} is the R2R^{2} [Eq. (10)] value of HcH_c considered as a function of variable α\alpha. The computed VIF factors are 1.53 for MsM_s, 4.36 for AexA_{\rm ex}, and 1.53 for KuK_u.

More advanced decision tree regressors improve the performance. XGB, as the most accurate model, was used as a benchmark, with the hyperparameters given in Appendix A for HcH_c prediction from experimental data. The deviance in the MSE of the training and testing data sets is shown in Figure 14. ANNs exhibit performance similar to other non-linear regressors. The ANN is as follows: (dense layers=64, activation function=leaky_relu, dense layers=32, activation function=leaky_relu, dense layers=16, activation function=leaky_relu, and output layer=1, and learning rate=0.007, epochs=1000, batch size=56). To obtain a deeper insight into ML predictability, the experimental, scaled-micromagnetic, and ML-

Deviance as a function of boosting iterations

Figure 14. Deviance as a function of boosting iterations for the micromagnetic simulated dataset with XGB model. Deviance is MSE between the actual and predicted data set computed in each iteration. Both the training and test data set converge well after 600 boosting iterations.

predicted HcH_c values are compared with only the known important rare-earth-based materials in Table IV. In general, ML-predicted values are smaller than the micromagnetically computed values. For 2 : 14 : 1 rare-earth magnets, ML improves the results. For 1 : 5 rare-earth magnets, it slightly underestimates the experiment. The full comparison of experimental and ML predicted data is given in Figure 8

Scatter plot showing the comparison of experimental and ML-mumax3 predicted $H_c$ for experimental materials

Figure 8. Scatter plot showing the comparison of experimental and ML-mumax3^{3} (ML-micromagntics data) predicted HcH_c for experimental materials.

MaterialHc(exp)H_c(\mathrm{exp})Hc(XGB)H_c(\mathrm{XGB})Hc(ANN)H_c(\mathrm{ANN})cHccH_c
La2_2Fe14_{14}B0.520.520.480.53
Ce2_2Fe14_{14}B0.540.840.730.79
Pr2_2Fe14_{14}B1.541.591.541.62
Nd2_2Fe14_{14}B1.241.501.451.56
Gd2_2Fe14_{14}B0.230.560.520.60
Tb2_2Fe14_{14}B2.934.005.655.41
Dy2_2Fe14_{14}B1.702.763.803.70
Ho2_2Fe14_{14}B0.751.501.761.82
Lu2_2Fe14_{14}B0.290.590.520.58
Y2_2Fe14_{14}B0.190.570.480.54
Th2_2Fe14_{14}B0.170.560.480.53
La2_2Co14_{14}B0.340.650.680.77
Pr2_2Co14_{14}B2.502.282.242.36
Nd2_2Co14_{14}B3.691.000.901.06
Gd2_2Co14_{14}B0.300.672.792.37
Y2_2Co14_{14}B0.340.620.610.71
NdLaCeFe14_{14}B0.651.331.251.39
LaCeYFe14_{14}B0.410.991.261.39
NdPrFe14_{14}B0.801.331.271.39
Sm2_2Co17_{17}1.251.601.531.62
Sm2_2Fe17_{17}N3_32.302.463.132.95
SmCo5_57.506.766.669.70
YCo5_53.902.632.683.33
LaCo5_55.254.184.144.22
CeCo5_55.704.105.105.17
PrCo5_55.324.053.814.02
NdCo5_50.150.120.100.12
RMSE0.901.101.10
MAE0.680.780.80

Table IV. TABLE IV. HcH_c (in T) comparisons among experiment, ML prediction, and scaled mumax⁡3\operatorname{mumax}^{3} for different permanent magnetic materials. The ML (tuned XGB and ANN) predictions are obtained by training the hypothetical materials using the random independent variables MsM_s, AexA_{\mathrm{ex}}, and KuK_u, and the mumax⁡3\operatorname{mumax}^{3} computed HcH_c. RMSE and MAE quantify the discrepancy between the experimental ML and the mumax⁡3\operatorname{mumax}^{3} scaled cHccH_c.

Additionally, XGB was trained with different fractions of the dataset, from 200 to 12000, by computing the mean absolute error (MAE) error metric:

MAE=∑iN∣yi−ypredi∣N(15)\mathrm{MAE} = \sum_{i}^{N} \frac{\lvert y_i-y_{\mathrm{pred}i}\rvert}{N} \tag*{(15)}

Figure 9 shows the convergence of the MAE (obtained with the XGB model) plotted against the training data set size, indicating that the model performs well for datasets with more than 6000 points.

MAE convergence for $H_c$ with respect to training data set size in XGB model

Figure 9. MAE convergence for HcH_c with respect to training data set size in XGB model. The xx-axis is in the scale of 10410^{4}.

Next, we discuss the loss function RMSE to examine the model performance. We computed the root mean square error (RMSE) as

RMSE⁡=1N∑i=1N(yi−y^i)2,(16)\operatorname{RMSE} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(y_i-\hat{y}_i)^2}, \tag*{(16)}

where yy and y^\hat{y} refer to the experimental and ML/mumax⁡3\operatorname{mumax}^{3} predicted values. Remarkably, the RMSE and MAE between experimental and ML (tuned XGB) predicted values are smaller than the RMSE between experimental and the scaled-mumax⁡3\operatorname{mumax}^{3}, indicating that ML helps to correct the mumax⁡3\operatorname{mumax}^{3} predictions. It also shows how the HcH_c is correlated with other features. Moreover, ML is an order of magnitude computationally faster than mumax⁡3\operatorname{mumax}^{3}, and we can use the trained model for predicting HcH_c in new materials without any mumax⁡3\operatorname{mumax}^{3} calculations.

ML application with DFT

Our computational HcH_c predictions involve two steps: DFT for Ce-doped Nd2_2Fe14_{14}B, and then ML trained on micromagnetically generated databases with input parameters from the DFT computations.

DFT methods and crystal structure

We used the Vienna simulation package (VASP)103,104 with the projector augmented wave (PAW) formalism in the generalized gradient approximation (GGA) of Perdew-Burke-Ernzerhof (PBE) semi-local exchange-correlation functionals105,106 including onsite electron-electron correlation Dudarev-Hubbard107 Ueff=U−JU_{\mathrm{eff}} = U - J for the 4f4f states of Nd, and spin-orbit interaction. The kinetic energy cut-off is 520 eV for the plane wave expansion, with a 4×4×24 \times4 \times2 kk-mesh for the Brillouin zone integration. The experimental structure was used in calculations to avoid overestimating bond lengths lengths and lattice constants in the PBE.

The crystal structure of Nd2_2Fe14_{14}B is tetragonal with 68 atoms (4 formula units), having space group P42/mnmP4_2/mnm (136) and lattice constants a=b=8.80a=b=8.80 and c=12.20c=12.20 Å, taking α=β=γ=90∘\alpha=\beta=\gamma=90^\circ at a temperature T=285T=285 K108, with experimental parameters for full Ce-doped Nd2_2Fe14_{14}B109. The crystal structure of the 2:14:1 neo-magnet is shown in Figure 10, which consists of rare-earth Nd in the 2 inequivalent sites 4f4f and 4g4g, transition element Fe in the 6 inequivalent sites at 16k16k, 16k16k, 8j8j, 8j8j, 4e4e, and 4c4c, and finally B at the 4f4f sites. Fully self-consistent spin-orbit calculations, i.e., PBE+UU+spin-orbit coupling (SOC), were performed to obtain the magnetic anisotropy energy along with the spin and orbital magnetic moments.

Crystal structure of Nd₂Fe₁₄B

Figure 10. Crystal structure of Nd2_2Fe14_{14}B occupied by Nd at two non-equivalent sites 4ff and 4gg, Fe at six inequivalent sites 16kk, 16kk, 8jj, 8jj, 4cc, and 4ee (labeled by Wyckoff positions), and B at the 4ff site.

Magnetic properties

Here we discuss the PBE+UU+SOC results with Nd2_2Fe14_{14}B, using ML for HcH_{\mathrm{c}} prediction. Hubbard UU-correction is necessary for highly correlated 3dd and 4ff elements110–113 to accurately predict KuK_{\mathrm{u}}. We performed test calculations for various values of UU from 4–7 eV for Nd 4ff orbitals. Table V shows the spin and orbital magnetic moments of Nd, Fe, and B at different crystallographic sites for different values of UeffU_{\mathrm{eff}}. The non-magnetic B atom carries negligible magnetic moments. Nd exhibits an Nd3+^{3+} state with spin magnetic moment (μs\mu_{\mathrm{s}}) of ∼−3.3\sim-3.3 and strongly quenched orbital moment (μl\mu_{\mathrm{l}}) of ∼1.5 μB\sim1.5\,\mu_{\mathrm{B}} due to the crystalline electric field. The spin and orbital moments have opposite signs, consistent with Hund’s rule for a less than half-filled 4ff shell. The net magnetic moment is robust, and does not vary significantly with UU.

Table VI lists KuK_{\mathrm{u}}, MsM_{\mathrm{s}}, TCT_{\mathrm{C}}, and their comparison with experimental values. KuK_{\mathrm{u}} is positively correlated with UU. The experimental KuK_{\mathrm{u}} is 4.5 MJ/m3^{3}114 at 300 K, which corresponds to the computed value for U∼4U \sim4 eV. TCT_{\mathrm{C}} is computed with the static Green’s function (GF) as implemented in ASA in the local density approximation (LDA)115 for the exchange-correlation functional in the linearized muffin-tin orbital (LMTO) program116,117. The pair exchange interaction JRR′J_{RR^{\prime}} between magnetic RR and R′R^{\prime} ions is computed using the Lichtenstein formula118, from which TCT_{\mathrm{C}} is estimated. The calculated Weiss mean-field theory119 value of 718.1 K is larger than the experimental value, which is expected in the mean-field approximation (MFA). According to a spin-wave theory by Tyablikov120`` the random-phase approximation (RPA) corrects the value, bringing it down to 539.1 K. The experimental value lies within the MFA and RPA limits.

Next, we discuss the site-resolved spin-orbit anisotropy energy which is computed as Eanis=E100−E001E_{\mathrm{anis}} = E_{100} - E_{001}111, where E100E_{100} and E001E_{001} are PBE+UU+SOC-computed atomic site energies for spin-quantization along the [100] and [001] directions. Generally, 4ff-elements contribute to KuK_{\mathrm{u}}, and 3dd-elements contribute to the magnetic moments in rare-earth-based magnets. The spin moments of the rare-earth ion are anti-parallel to the spin moment of the 3dd ion, which is also the case for neo-magnets. The site contribution of the Nd-element to crystalline anisotropy energy is given in Table VII. In our PBE+UU+SOC calculations, we did not impose symmetry, which means that all the atoms are inequivalent under the P1P_{1} crystal symmetry. All eight Nd atoms split into eight different sites. The individual on-site energies differ significantly for the atoms of the same crystallographic site. Interestingly, at least one Nd (Nd1_1) at 4ff has a negative contribution to crystalline anisotropy energy, as given in Table VII. Theoretically, it can be inferred that Nd has a tendency to be planar at the 4ff site, differing from the 4gg site, which is strictly uniaxial along the crystalline cc direction. These results are consistent with experiment121.

Table VIII shows the computed values of the magnetic moments and magnetic anisotropy in the Ce-substituted neo-magnet at Nd-4ff sites. The Ce-atom carries a small spin magnetic moment ∼1 μB\sim1\,\mu_{\mathrm{B}} compared to the Nd ∼3 μB\sim3\,\mu_{\mathrm{B}} moment. Moreover, the orbital moment does not entirely cancel the spin moment in Nd, unlike in Ce, where the net spin + orbital moment vanishes. Therefore, there is an increase in the net magnetic moment with Ce. On the other hand, magnetic anisotropy is reduced because the Ce-site contribution is much smaller than that of Nd. The calculated value of KuK_{\mathrm{u}} slightly underestimates the experiment (see Table I), which is reasonable, given the choice of the UU-values used for the Ce and Nd 4ff states. We used Ueff=2U_{\mathrm{eff}} = 2 eV for Ce and 4 eV for Nd in the calculations.

ML coercivity prediction

Central to this paper is the hypothesis that macroscopic measures of HcH_{\mathrm{c}} are not well predicted by theory, while the microscopic magnetic parameters MsM_{\mathrm{s}}, KuK_{\mathrm{u}}, and AexA_{\mathrm{ex}} are. However, knowledge of the microscopic parameters is sufficient to produce realistic estimates of HcH_{\mathrm{c}} by accumulating a database and modeling it with ML. Furthermore, once an ML prediction is made, information about the material grain structure may be reverse-engineered with micromagnetic modeling. Although material fabrication variations have a significant impact on HcH_{\mathrm{c}}, this section demonstrates the surprising effectiveness of the methodology as applied to Nd2_2Fe14_{14}B, with and

atom→\rightarrowNd(4f4f)Nd(4g4g)Fe(16k16k)Fe(8j8j)Fe(4e4e)Fe(4c4c)B(4g4g)total
U↓U\downarrowμs\mu_s μl\mu_lμs\mu_s μl\mu_lμs\mu_s μl\mu_lμs\mu_s μl\mu_lμs\mu_s μl\mu_lμs\mu_s μl\mu_lμs\mu_s μl\mu_lμst\mu_{st} μlt\mu_{lt}
4 eV-3.26 1.52-3.28 1.532.27, 2.36 0.043, 0.0492.29, 2.71 0.0442.03 0.0452.48 0.054-0.17 0.00105.24 14.75
5 eV-3.26 1.52-3.28 1.532.27, 2.36 0.042, 0.0482.29, 2.70 0.0442.03 0.0452.47 0.052-0.17 0.00105.25 14.75
7 eV-3.27 1.52-3.28 1.532.27, 2.36 0.042, 0.0482.29, 2.71 0.0422.03 0.0442.48 0.049-0.17 0.00105.25 14.68

Table V. Spin and orbital magnetic moments (μs\mu_s and μl\mu_l) of individual atoms in different Wyckoff positions, and total magnetic spin and orbital magnetic moments (μst\mu_{st} and μlt\mu_{lt}) per unit cell (in μB\mu_B), in Nd2_2Fe14_{14}B computed using the PBE+UU+SOC method with various values of UU for the Nd (4f4f) electronic states. There are two 16k16k and 8j8j sites, therefore, the pairs are given. The μl\mu_l of the pair at 8j8j-sites do not differ, and only one value is given to represent both atoms.

UUKuK_uMsM_sMs(Expt20)M_s(\mathrm{Expt}^{20})TCT_CTC(Expt20)T_C(\mathrm{Expt}^{20})
4 eV6.1629.9837.7a^{a}, 32.5b^{b}718.1c^{c}, 539.1d^{d}585
5 eV8.5630.00⋯\cdots⋯\cdots⋯\cdots
7 eV11.5830.01⋯\cdots⋯\cdots⋯\cdots

Table VI. Calculated KuK_u in MJ/m3^3, MsM_s in μB\mu_B/f.u. using PBE+UU+SOC, TCT_C in Kelvin (K) using Green’s function-ASA (LDA), and comparison with available experiment.

PropertiesNd2_2Fe14_{14}BCe2_2Fe14_{14}B
μs\mu_s105.24120.73
μl\mu_l14.754.55
μs+μl\mu_s+\mu_l119.99125.28
KuK_u6.161.21
MsM_s1.311.24

Table VIII. Total spin, orbital, and spin + orbital magnetic moments μs\mu_s, μl\mu_l, and μs+μl\mu_s+\mu_l in μB\mu_B/cell, KuK_u in MJ/m3^3, and MsM_s in MA/m of Ce2_2Fe14_{14}B and Nd2_2Fe14_{14}B.

AtomWyckoff-positionEanisoE_{\mathrm{aniso}}
Nd1_14f4f1.3780
Nd2_24f4f-1.1914
Nd3_34f4f0.5336
Nd4_44f4f6.2204
Nd5_54g4g0.4865
Nd6_64g4g3.8064
Nd7_74g4g6.2094
Nd8_84g4g3.0622

Table VII. Site resolved spin-orbit anisotropy energy (EanisE_{\mathrm{anis}}) in meV of the Nd contribution to magneto-crystalline anisotropy energy. In PBE+UU+SOC calculations, the crystalline symmetry lowers to the P1P1 point group implying that the atoms (including the Nd) split into eight sites. The negative value for Nd at the 4f4f site indicates it produces a planar contribution to the crystalline anisotropy energy, consistent with experiment121.

without Ce doping.

The ab initio-calculated values were fed to the tuned XGB ML model to produce the output predictions. Additionally, micromagnetic simulation was used to determine a hysteresis loop. The results are depicted in Table IX. As can be seen, the reduction trend of HcH_c

MaterialML-ExpML-mumax3cHccH_\mathrm{c}
Nd2_2Fe14_{14}B3.922.392.35
CeNdFe14_{14}B0.881.271.29
Ce2_2Fe14_{14}B0.390.400.36
SmCo5_58.875.8818.12
YCo5_51.451.511.60
CeZrFe11_{11}6.254.935.99
CeZrFe11_{11}N1.512.152.59
CeTiFe11_{11}1.502.092.70
CeTiFe11_{11}N1.511.922.45

Table IX. Comparison of ML predicted and scaled-micromagnetic (cHccH_\mathrm{c}) HcH_\mathrm{c} in T for materials with calculated DFT parameters. ML-exp/ML-mumax3 represent to predictions made by training experimental/scaled-micromagnetic data.

Hysteresis curve for Nd$_{2-x}$Ce$_x$Fe$_{14}$B

Figure 11. Hysteresis curve for Nd2−x_{2-x}Cex_xFe14_{14}B (x=0x = 0, 1.0, 2.0) calculated from micromagnetic simulation. Ce leads to the reduction in the coercivity. M/MsM/M_s is the reduced magnetization.

with Ce is similar to that found experimentally122. Further, we computed the ML [trained with experiment (ML-Exp) and scaled micromagnetic data (ML-mumax3^3) separately] and scaled micromagnetic simulated HcH_c for SmCo5_5, YCo5_5, CeZrFe11_{11}, CeZrFe11_{11}N, CeTiFe11_{11}, and CeTiFe11_{11}N using DFT intrinsic parameters as given in Table IX. The ML-exp predicts inconsistent HcH_c due to the small data size training, as evident from model performance metrics in Tab II. The ML-mumax3^3 trained on big data sets produces more sensible HcH_c than scaledmicromagnetics and ML-exp. For SmCo5_5, micromagnetic simulation yielded a factor of three larger HcH_\mathrm{c} than the ML-mumax3 model and experiments. For YCo5_5, the ML-mumax3 model produced slightly lower HcH_\mathrm{c} than the experiment, attributed to the DFT underestimate of K1K_1. Overall the ML-mumax3 model consistently tends to correct the micromagnetically predicted HcH_\mathrm{c} towards experiments for known materials (see Table I).

If the fabricated material is grown as a crystal with well-defined grains, it is possible to reverse-engineer the structure. The dependence of HcH_\mathrm{c} on micromagnetic parameters viz, grain sizes and inter-grain coupling is visualized in Figure 12 in 128 × 128 × 128 nm3^3 magnetic cell.

Coercivity as a function of grain coupling and grain size

Figure 12. Coercivities for different inter-grain couplings for Nd2_2Fe14_{14}B, measured as a function of the ratio of inter-grain exchange stiffness to intra-grain exchange stiffness and different grain sizes. As the material is complex, a cuboid of 128 × 128 × 128 nm3^3 was generated for each grain size, resulting in some random undulations according to grain initialization. The plot displays a decreasing coercivity with increasing grain size as observed in experiment122.

Here the grain coupling refers to a reduction factor to the stiffness exchange coupling AexA_\mathrm{ex}. We note that we did not find much difference in the results with the use of 64 × 64 × 64 and 128 × 128 × 128 nm3^3 simulation cells. Although non-linear, we find a similar trend of HcH_\mathrm{c} with inter-grain exchange coupling and grain sizes. For a fixed value of grain size, qualitatively, HcH_\mathrm{c} decreases with inter-grain exchange coupling, while it is the opposite with grain size. That inter-grain exchange coupling may not be uniform in actual material due to the void or imperfection between the grains, which affects the perfect spin alignment resulting in MsM_\mathrm{s}. A larger inter-grain exchange coupling further reduces K1K_1 and HcH_\mathrm{c}. With an increase in grain size, HcH_\mathrm{c} experiences a decrease, similar to that reported in the Dy-substituted neo-magnet123.

Various combinations of grain size and inter-grain exchange coupling reproduce the actual coercivity. For instance, with grain size ∼30\sim30 nm, and a 0.58 ratio for inter-grain to intra-grain AexA_\mathrm{ex} reproduces the realistic HcH_\mathrm{c} of 0.40 T, as demonstrated in Figure 13. Strictly speaking, a larger inter-grain coupling tends to reduce the HcH_\mathrm{c}, regardless of grain size. This suggest that we can explore the other magnetic parameters AexA_\mathrm{ex} and grain size by using ML-predicted results employing micromagnetic simulations, which demonstrates the fundamental usefulness of ML, not just as a black box, but also as a pre-processing input to a micromagnetic model to reverse-engineer the micromagnetic structure.

Hysteresis curve for Ce$_2$Fe$_{14}$B

Figure 13. Hysteresis curve for Ce2_2Fe14_{14}B calculated naively with micromagnetic simulation with grain size set to 30 nm and intergrain stiffness to intragrain stiffness set at 0.58 to recreate the ML HcH_\mathrm{c} prediction.

Conclusion

Experimental dataset input and ML modeling coupled with DFT predicts HcH_\mathrm{c} with far greater accuracy and speed than was previously possible using micromagnetic modeling. This technique provides a robust computational foundation for predicting novel permanent magnet materials and optimizing their properties. Using micromagnetic simulation, we first studied magnetization as a function of the applied magnetic field for real and hypothetical magnetic materials, mainly focusing on 1 : 5 and 2 : 14 : 1 rare-earth-based permanent magnets such as Nd2_2Fe14_{14}B. Calculations of HcH_c based on the hysteresis loop overestimate by a factor of ∼5\sim5; the Brown Paradox. The paradox is side-stepped by the judicious use of ML, obtaining a more accurate prediction of the target variable by learning its dependence on the input features. We find that HcH_c is directly proportional to KuK_u, inversely proportional to MsM_s, and weakly correlated with AexA_{\mathrm{ex}}.

We apply the ML modeling to Nd2_2Fe14_{14}B by first computing its independent variables with DFT calculations. These calculations show that the Nd 4g-sites mainly contribute to the uniaxial magneto-crystalline anisotropy. They also yield a value for the TCT_C, which agrees with the experiment. The DFT predictions suggest the possibility of tuning rare-earth magnetic properties by substituting non-critical elements at specific sites. For instance, Ce-doping at 4f sites shows only a slight reduction in HcH_c, consistent with the ML prediction. Finally, we engineer the grain boundary size and inter-exchange coupling with the aid of ML-predicted HcH_c, which indicates that the reduction in inter-grain exchange coupling reduces HcH_c. On the other hand, reducing the grain size increases HcH_c qualitatively.

Acknowledgments

This work is supported by the Critical Materials Innovation Hub, an Energy Innovation Hub funded by the US Department of Energy, Office of Energy Efficiency and Renewable Energy, Advanced Materials and Manufacturing Technologies Office. The Ames National Laboratory is operated for the U.S. Department of Energy by Iowa State University of Science and Technology under Contract No. DE-AC02-07CH11358. GNN acknowledges support from the U.S. Department of Energy, Office of Science, Office of Workforce Development for Teachers and Scientists, Office of Science Graduate Student Research (SCGSR) program. The SCGSR program is administered by the Oak Ridge Institute for Science and Education for the DOE under contract number DE-SC0014664.

ml model and hyperparameter tuning

In this appendix we provide information about the hyperparameters used in the fine tuning of the ML models. With the experimental data, we used the hyperparameters given in Table X for tuned XGB model training.

n_estimators=500
min_weight_fraction_leaf=0
max_depth=5
learning_rate=0.006

Table X. Hyperparameters used for XGB model training of experimental magnetic materials.

In mumax3^3 data training we use the hyperparameters given in Table XI.

ModelHyperparameters
Tuned RFn_estimators=1200, max_depth=5, random_state=1, min_samples_leaf=3, bootstrap=True, max_features=’auto’
Tuned XGBn_estimators=1200, max_depth=5, min_weight_fraction_leaf=0, learning_rate=0.007
Tuned XGBn_estimators=1200, max_depth=5, min_child_weight=0, learning_rate=0.007, random_state=42

Table XI. Fine-tuned hyperparameters for different ML models used in the training of micromagnetic simulated data.

The hyperparameter used in the fine-tuned GBR model training is shown in Figure 14. The deviance (MSE) converges well in both the test and training datasets after 600 boosting iterations, confirming a very good model performance.

DFT parameters for 1:5 and 1:12 intermetallics

tbl:XII provides the DFT computed parameters for 1 : 5 and 1 : 12 intermetallic compounds.

MaterialMsM_s (A/m)AexA_{\mathrm{ex}} (pJ/m)KuK_u (MJ/m3^3)
SmCo5_5${}^{124}` | 1069768.213 | 7 | 40.27
YCo5_5${}^{125}` | 770449.2281 | 7 | 2.65
CeZrFe11_{11}597220.260277.36
CeZrFe11_{11}N632804.637673.42
CeTiFe11_{11}587261.715473.29
CeTiFe11_{11}N625792.124773.20

Table 12. DFT computed intrinsic magnetic parameters for 1 : 12 compositions taken from our previous work113 except for 1 : 5 type materials as referred in Table. AexA_{\mathrm{ex}} is kept fixed for all materials.

References

  1. [27](1968).
  2. [28]R. A. Ristau, K. Barmak, L. H. Lewis, K. R. Coffey, and J. K. Howard, On the relationship of high coercivity and L1 ordered phase in CoPt and FePt thin films, J. Appl. Phys. 86, 4527 (1999).DOI
  3. [29]B. C. Sales, B. Saparov, M. A. McGuire, D. J. Singh, and D. S. Parker, Ferromagnetism of Fe₃Sn and alloys, Sci. Rep. 4, 7024 (2014).DOI
  4. [30]H. Jaballah, W. Bouzidi, R. Fersi, N. Mliki, and L. Bessais, Structural, magnetic and magnetocaloric properties of (Pr,Sm)₂Fe₁₇ compound at room temperature, J. Phys. Chem. Solids 161, 110438 (2022).DOI
  5. [31]Y. Iwama, M. Takeuchi, and M. Iwata, New determination of magnetic anisotropy constants of alnico magnet alloys, J. Phys. Colloques 32, C1 (1971).
  6. [32]Y. Liu, J. Song, G. Liu, J. Chen, C. Wang, H. Wang, J. Wang, and X. Zhang, High strength and low coercivity of cobalt with three-dimensional nanoscale stacking faults, Nano Lett. 21, 6480 (2021), pMID: 34324350.DOI
  7. [33]H. Sato, T. Yoshioka, H. Tsuchiyura, Y. Mizuno, K. Koike, K. Takahashi, and H. Kato, Rare-earth moment reduction and local magnetic anisotropy in Pr₂Fe₁₄B and Tm₂Fe₁₄B, J. Magn. Magn. Mater. 545, 168684 (2022).DOI
  8. [34]N. Yu, M. Zhu, L. Song, Y. Fang, K. Song, Q. Wang, and W. Li, Coercivity temperature dependence of Sm₂Co₁₇-type sintered magnets with different cell and cell boundary microchemistry, J. Magn. Magn. Mater. 452, 272 (2018).DOI
  9. [35]T. Kikuchi, T. Nakamura, T. Yamasaki, M. Nakanishi, T. Fujii, J. Takada, and Y. Ikeda, Magnetic properties of La–Co substituted M-type strontium hexaferrites prepared by polymerizable complex method, J. Magn. Magn. Mater. 322, 2381 (2010).DOI
  10. [36]K. M. U. Rehman, M. Riaz, X. Liu, M. W. Khan, Y. Yang, K. M. Batoo, S. F. Adil, and M. Khan, Magnetic properties of Ce doped m-type strontium hexaferrites synthesized by ceramic route, J. Magn. Magn. Mater. 474, 83 (2019).DOI
  11. [37]P. Wang, X. Wang, L. Qiao, J. Zhang, G. Wang, B. Duan, T. Wang, and F. Li, High-frequency magnetic properties and microwave absorption performance of oxidized Pr₂Co₁₇ flakes/epoxy composite in x-band, J. Magn. Magn. Mater. 468, 193 (2018).DOI
  12. [38]J. Zhang, H. Gao, Y. Yan, X. Bai, F. Su, W. Wang, and X. Du, Morphology and magnetic properties of CeCo₅ submicron flakes prepared by surfactant-assisted high-energy ball milling, J. Magn. Magn. Mater. 324, 3272 (2012).DOI
  13. [39]A. Gabay, X. Hu, and G. Hadjipanayis, Preparation of YCo₅, PrCo₅ and SmCo₅ anisotropic high-coercivity powders via mechanochemistry, J. Magn. Magn. Mater. 368, 75 (2014).DOI
  14. [40]Y. Hirayama, A. Panda, T. Ohkubo, and K. Hono, High coercivity Sm₂Fe₁₇N₃ submicron size powder prepared by polymerized-complex and reduction–diffusion process, Scr. Mater. 120, 27 (2016).DOI
  15. [41]J. Fliegans, Coercivity of NdFeB-based sintered permanent magnets : experimental and numerical approaches, Theses, Université Grenoble Alpes (2019).DOI
  16. [42]Q. Cheng, J. Lin, and M. Su, On synthesis and magnetic properties of Nd(Fe, Mo, Ti)₁₂Zₓ (Z=N, H), J. Alloys Compd. 280, 310 (1998).
  17. [43]T. Oyama, ed., The Chemistry of Transition Metal Carbides and Nitrides, 1 (Springer Dordrecht) pp. XXVI, 536.
  18. [44]T. Wu, S. Cao, M. Kou, Y. Xie, G. Ding, S. Guo, B. Zheng, R. Chen, M. Zhong, and A. Yan, Magnetic performance and microstructure of ndfeb sintered magnet by diffusing Tb₁₀Pr₉₀−ₓ(Cu,Al,Ga)ₓ alloys, J. Alloys Compd. 934, 167888 (2023).DOI
  19. [45]J. Tsui, K. Strnat, and J. Schweizer, High magnetic coercivity of neodymium- and didymium-cobalt alloys sintered with Pr and Sm additives, APL 21, 446 (2003).
  20. [46]C. Zhou and F. E. Pinkerton, Magnetic properties of Ce–Nd–Fe–Mo alloys and their nitrides, J. Magn. Magn. Mater. 369, 127 (2014).DOI
  21. [47]G. Gkouzia, D. Günzing, R. Xie, T. Webels, A. Kovács, A. T. N’Diaye, M. Major, J. P. Palakkal, R. E. Dunin-Borkowski, H. Wende, H. Zhang, K. Ollefs, and L. Alff, Element-specific study of magnetic anisotropy and hardening in SmCo₅−ₓCuₓ thin films, Inorg. Chem. 62, 16354 (2023), pMID: 37739403.DOI
  22. [48]K. Suresh, R. Gopalan, A. Singh, G. Bhikshamaiah, V. Chandrasekaran, and K. Hono, Coercivity of Sm(Co₀.₉Cu₀.₁)₄.₈ melt-spun ribbons, J. Alloys Compd. 436, 358 (2007).DOI
  23. [49]J. Téllez-Blanco, R. Grössinger, and R. Sato Turtelli, Structure and magnetic properties of SmCo₅−ₓCuₓ alloys, J. Alloys Compd. 281, 1 (1998).DOI
  24. [50]M. D. Kuz’min, K. P. Skokov, I. Radulov, C. A. Schwöbel, S. Foro, W. Donner, M. Werwiński, J. Ruzs, E. Delczeg-Czirjak, and O. Gutfleisch, Magnetic anisotropy of La₂Co₇, J. Appl. Phys. 118, 053905 (2015).DOI
  25. [51]M. Q. Huang, S. G. Sankar, W. E. Wallace, M. E. McHenry, Q. Chen, and B. M. Ma, Structure and magnetic properties of RCo₇−ₓZrₓ (R=Y, Gd, Nd, or Ho, x=0-0.8), J. Appl. Phys. 87, 5305 (2000).DOI
  26. [52]M. T. Onyszczak, T. N. Lamichhane, S. L. Bud’ko, P. C. Canfield, and A. Palasyuk, Structural and magnetic properties of the CeCo₅−CeZn₅ solid solution and potential improvements upon iron substitution, J. Magn. Magn. Mater. 482, 192 (2019).
  27. [53]Y. Wong, H. Chang, Y. Lee, W. Chang, C. Chiu, and C. Mo, Coercivity enhancement of thicker sintered ndfeb magnets by grain boundary diffusion with low-melting Tb₇₅−ₓCeₓCu₂₅ (x=0–45) alloys, J. Magn. Magn. Mater. 515, 167287 (2020).DOI
  28. [54]W. Silva, N. Ferreira, J. Soares, R. da Silva, and M. Macêdo, Investigation of structural and magnetic properties of nanocrystalline mn-doped SrFe₁₂O₁₉ prepared by proteic sol–gel process, J. Magn. Magn. Mater. 395, 263 (2016).DOI
  29. [55]S. Katlakunta, S. S. Meena, S. Srinath, M. Boududina, R. Sandhya, and K. Praveena, Improved magnetic properties of Cr³⁺ doped SrFe₁₂O₁₉ synthesized via microwave hydrothermal route, Mater. Res. Bull. 63, 58 (2015).DOI
  30. [56]S. Chawla, S. Meena, P. Kaur, R. Mudsainiyan, and S. Yusuf, Effect of site preferences on structural and magnetic switching properties of Co-Zr doped strontium hexaferrite SrCoₓZrₓFe₍₁₂−₂ₓ₎O₁₉, J. Magn. Magn. Mater. 378, 841 (2015).
  31. [57]R. Alange, P. P. Khirade, S. D. Birajdar, A. V. Humbe, and K. Jadhav, Structural, magnetic and dielectrical

Paper details

vlnc:W_pr9nyud3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents