Abstract
Databases for twodimensional materials host numerous ferromagnetic materials without the vital information of Curie temperature since its calculation involves a manually intensive complex process. In this work, we develop a fully automated, hardwareaccelerated, dynamictranslation based computer code, which performs first principlesbased computations followed by Heisenberg modelbased Monte Carlo simulations to estimate the Curie temperature from the crystal structure. We employ this code to conduct a highthroughput scan of 786 materials from a database to discover 26 materials with a Curie point beyond 400 K. For rapid data mining, we further use these results to develop an endtoend machine learning model with generalized chemical features through an exhaustive search of the model space as well as the hyperparameters. We discover a few more high Curie point materials from different sources using this datadriven model. Such material informatics, which agrees well with recent experiments, is expected to foster practical applications of twodimensional magnetism.
Similar content being viewed by others
Introduction
The recent experimental demonstration of ferromagnetism in twodimensional (2D) materials: CrI_{3}^{1} and Cr_{2}Ge_{2}Te_{6}^{2} at low temperatures, has opened a new horizon of nanotechnology research since these materials inherit the potential to revolutionize engineering fields like spintronics^{3}, valleytronics^{4}, sensing and memory technologies^{5}. In their classical work, Mermin and Wanger^{6} showed that under an isotropic Heisenberg model, longrange magnetic order must be absent in 2D. However, the more recent discovery of even roomtemperature ferromagnetism in monolayer VSe_{2}^{7} and MnSe_{2}^{8} has been possible as the strong magnetocrystalline anisotropy of these 2D materials lifts the Mermin–Wanger restriction. So far, a plethora of 2D ferromagnetic (FM) materials^{9,10,11,12,13,14} have been computationally predicted, including a few generalpurpose 2D materials databases^{15,16,17} containing hundreds to thousands of entries. However, none of these databases contain the most crucial parameter for 2DFM materials relevant for practical applications: the transition temperature or Curie point (T_{C}). This is due to the fact that the computational determination of T_{C} is a highly complex process, which involves a manual heuristicsbased search for the groundstate and lowenergy spin configurations. Identification of different magnetic exchanges (direct, super or double) within the neighbouring atoms and mapping them appropriately in a Monte Carlo based spinflipping simulator has also been a manually intensive exercise. The choice of the model Hamiltonian (Ising instead of Heisenberg) used to simulate the spinflipping with temperature, also raises a question on the reliability of the Curie temperatures of 2D materials predicted so far in the literature^{10,14}.
Recently, an algorithm^{18} has been proposed which can search and predict the collinear, experimentally verified ground and lowenergy spin states for bulk materials, almost optimally and exhaustively. Building on this, we develop a code, which performs first principlesbased computations followed by Heisenberg modelbased Monte Carlo simulations to predict the Curie point accurately from any magnetic 2D material crystal structure. Software engineering on this code makes it capable to execute such rigorous calculations in a highthroughput manner, even on a workstationgrade computer with GPU (graphical processing unit) acceleration. We use this code to determine the Curie points of materials from a suitable database^{16}. To our surprise, almost 47% of the 786 materials classified as FM, turned out to be antiferromagnetic (AFM) upon close inspection by our code. The T_{C} and other magnetic properties could be successfully determined for 157 materials, among which 26 materials reveal beyond 400 K Curie point. Close agreement with experimentally measured T_{C} for a few materials validates our highthroughput methodology. In pursuit of faster discovery of highT_{C} materials, we further develop a machinelearning (ML) pipeline using these 157 data points. Using this ML model, we identify a few high T_{C} 2DFM materials from the literature and other databases.
The informatics, which optimally balances the rigorousness and efficiency, gives us unprecedented opportunity to compare the magnetic properties of a very large number of materials with diverse structures, which may lead to many new insights on 2D magnetism. For example, we understand why the inclusion of the higherorder neighbours is important for T_{C} calculation for certain materials and why it is not for the others. We observe several violations of the Goodenough–Kanamori^{19,20} rules for superexchange, the origin of which is open for further exploration. We also demonstrate that a machinelearning model can capture the complex process of temperaturedependent spinflipping with exceptional accuracy. Our work thus significantly upgrades the computational materials toolbox to foster practical applications of twodimensional magnetism.
Results
Highthroughput computational framework
We first explain the workflow of our automated code as illustrated in Fig. 1. The unit cell of the material is first fed to a recently developed module^{18} of the opensource python library pymatgen^{21}, which generates different FM and AFM spin configurations of the material based on symmetry analysis. Eliminating the heuristicsbased approach, pymatgen not only helps to automate the process of the T_{C} calculation, but also makes it more rigorous and thus reliable [see Methods]. The codegenerated spin configurations for experimentally synthesized materials CrI_{3}^{1} and Cr_{2}Ge_{2}Te_{6}^{2} and newly predicted material Cr_{3}Te_{4}^{10} are shown in Supplementary Figs. 1–3 as examples. These structures are then relaxed using collinear density functional theory with Hubbard correction (DFT + U) and their energies are calculated. At this stage, if the ground state is found to be AFM, the material is discarded. Here we also calculate the magnetic moment (µ) of each atom of the structure. Since the magnetocrystalline anisotropic energy (MAE) is essential for the existence of longrange magnetic order in 2D materials, it is calculated using noncollinear DFT including the effects of spin–orbit coupling (SOC) in the next step. These calculations also reveal the easy magnetization axis (EMA) of the 2DFM material, which can be important for specific applications. All these calculations provide us enough information to fit the DFT energy values to the following Heisenberg Hamiltonian:
Here, J_{1}, J_{2}, J_{3} and J_{4} are the nearestneighbour (N1), 2nd nearestneighbour (N2), 3rd nearestneighbour (N3) and 4th nearestneighbour (N4) exchange coupling constants and S_{i}, S_{j}, S_{l}, S_{m} and S_{n} are the spins at sites i, j, l, m, and n, respectively. k_{x}, k_{y,} and k_{z} are the magnetic anisotropy constants in the x, y and z directions. S is computed as µ/(2µ_{B}), where µ is the local magnetic moment of the magnetic ions.
For 2D materials, the classical Monte Carlo (MC) based solution of the Heisenberg Model is known to accurately predict the transition temperature^{22}. First, the 2DFM unit cell is multiplied to make a supercell large enough to eliminate size effects, and using a GPU accelerated search, all the neighbours of all the sites are mapped into an 1D array structure. This is then used to perform a Heisenberg modelbased classical MC simulation using a semicompiled dynamictranslation based module, from which the T_{C} of the material can be obtained. For a few materials, no inplane anisotropy is observed which are classified as XY magnets. Mermin–Wagner theorem^{6} prohibits spontaneous symmetry breaking in these kinds of systems where the spin degree of freedom is ≤2. Instead, XY magnets exhibit a Berezinskii−Kosterlitz−Thouless (BKT) transition to a quasilongrange ordered lowtemperature phase. For these materials, T_{C} is calculated from the following equation obtained from Monte Carlo simulations of the XY model^{23}:
where k_{B} is the Boltzmann constant and E_{FM} and E_{AFM} are energies of the FM and the most stable AFM solutions normalized by the number of atoms.
Though the Ising model can provide good results for materials with extremely high anisotropy, significant overestimation of T_{C} may happen for materials with moderate to low anisotropy^{10,11,14}. The anisotropic Heisenberg model takes care of this problem and effectively balances the contributions between exchange and anisotropy. However, it requires further MAE calculations and the MC simulation becomes computationally much more expensive. Although this is a highthroughput study, leveraging software engineering [see Methods], we decide to use the rigorous Heisenberg model without compromising the accuracy.
Database search
We identify the database C2DB^{16} as the ideal database to conduct our study based on the following reasons. (1) The authors have performed a preliminary classification between FM and AFM materials and a large number (786) of materials are classified as FM. (2) These materials have been explored by a “systematic combinatorial approach” where almost all known layered exfoliable materials are covered^{15}, and by substituting the atoms, the authors have predicted a lot of new materials. This kind of variation is ideal to train machinelearning models, which is one of our primary goals. Also, recent synthesis of janus^{24} and other species substituted^{25} 2D materials with no bulk analogues have made practical application of this kind of “synthetic” 2D materials possible. (3)The authors have calculated several properties of interest including thermodynamic stability and energy above the convex hull which helps us to estimate the chance of potential synthesis of these materials. The electronic structures have also been computed, which tells us about the presence of important properties like halfmetallicity.
Interestingly, after a close examination by our code, 368 of the 786 materials classified as FM in the database turns out to be actually AFM. The pymatgen magnetism module explores the symmetry allowed spinconfiguration space almost exhaustively, and in the process also explores large AFM supercells, which probably the authors of C2DB could not afford to do in their generalpurpose study. A few discrepancies can also arise from the difference in DFT settings between the studies. Analyses of a considerable amount of materials have failed and thus have been discarded due to various computational limitations [see Methods]. Sheet 1 of Supplementary Spreadsheet 1 lists all the examined materials with FM/AFM classification as found by our method. In the end, the T_{C} of 157 2DFM materials could be successfully computed, among which 12 materials are found to be XY magnets. These materials belong to more than 20 different prototype structures, which are shown in Fig. 2.
Discovery of highT _{c} materials
Almost all experimentally synthesized 2D materials are there in C2DB, except Cr_{2}Ge_{2}Te_{6}, which too we have included manually in this set. We have not included those 2DFM materials whose ferromagnetism cannot be accurately modelled by the Heisenberg or the XY model, such as known itinerant material Fe_{3}GeTe_{2}^{26,27}. Sheet 1 of Supplementary Spreadsheet 2 lists all the calculated properties, as well as the computed Curie temperatures of these 2DFM materials. To include interactions of N neighbours in the Hamiltonian, N + 1 spin configurations are required, and the number of configurations generated by pymatgen is also listed in Supplementary Spreadsheet 2. Given enough configurations, we have included up to the 4th nearestneighbour interaction in this work. The T_{C}s are calculated using two methodologies: (1) commonly used nearestneighbour approach: including only the N1 interaction and fitting the energies of the FM and the most stable AFM configurations, (denoted as ‘TC’ in the spreadsheet) and (2) multineighbour approach: interactions including up to N4 (listed under the ‘TC_exact’ column). Apparently, the TC_exact should provide a better estimate of the real T_{C} of the material. However, to our surprise, we observe that TC, which is computationally much economic, is close to TC_exact in most of the cases, except a few as described below. (1) Materials with more than one distinct metal layer (prototypes CH, GaSe, CdI_{2}MXene, Ti_{2}CO_{2,} and Ti_{2}CH_{2}O_{2}). For prototypes CH and GaSe, there is no layer containing anions between the metal layers, which gives rise to strong interlayer direct exchange in addition to the intralayer superexchange. In prototypes CdI_{2}, Ti_{2}CO_{2} and Ti_{2}CH_{2}O_{2}, which are all MXenes, the distinct metal layers are connected by anionic layers, where both the interlayer as well as intralayer exchange interactions play a pivotal role in deciding the T_{C}. Clearly, considering only the N1 interactions in these materials is not accurate enough, as reflected in the huge differences between TC and TC_exact. In passing, we note that, for a few MXenes, pymatgen could only generate two configurations, thus, the TC value has been repeated as TC_exact. (2) Materials with square or rectangular lattice (prototype FeOCl, FeSe, GeS_{2} and NiSe). Here, the N1s or N2s are the atoms situated in the diagonally opposite corners of the square or rectangle, where superexchange is expected to be feeble at best and only strong direct exchange could persist. The difference in distance between N1s and N2s are also very small in these materials, which again makes the inclusion of the higher neighbour interactions necessary [see Supplementary Figs. 4 and 5]. In a few materials (Janus) the effect of higherorder neighbours is not possible to take into account due to moderate distortion in the lattice and T_{C} value has been repeated as TC_exact.
The calculated T_{C} (TC_exact) by our automated code for the experimentally synthesized materials CrI_{3}^{1}, Cr_{2}Ge_{2}Te_{6}^{2}, and MnSe_{2}^{8} matches the experimental reports very closely without any manual tinkering of parameters, validating the generalization and accuracy of our method. For Tphase VSe_{2} our predicted T_{C} is only 114.33 K, whereas room temperature ferromagnetism has been reported^{7}. However, it must be noted that the authors have reported strong substrate dependence of the magnetism and T_{C} in this study which explains this apparent discrepancy. Also, our code confirms the magnetism to be inplane in this material which matches the experimental report. With Supplementary Table 1, Supplementary Note 1 and Supplementary Figs 1, 2, and 6, we explain in detail how our code works for 2 FM (CrI_{3}^{1}, CrGeTe_{3}^{2}) and 2 AFM (FePS_{3}^{28}, NiPS_{3}^{29}) experimentally synthesized 2D materials.
We finally discover a total of 26 materials with T_{C} > 400 K and 32 materials with T_{C} ≥ 300 K, making these materials suitable for practical device applications. Interestingly, many of these materials are known to show a “low” amount of magnetism in bulk forms, such as materials containing Rh, Ru, Mo, W, Sc, Ti, and Zr, which were ignored in previous heuristicsbased searches^{10}. However, our study suggests that the materials containing the abovementioned metals can indeed show a “decent” (0.59–3.96 µ_{B}/atom) magnetic moment in 2D crystal form along with highT_{C}, possibly because of the enhanced electron localization. Also, for some of these materials the anisotropy is not great, but the difference in energy between the FM and AFM states, which ultimately translates to exchange parameters, helps to lift the T_{C} beyond the roomtemperature.
Since the magnetic properties of these materials can greatly depend on the value of U (Hubbard Correction), we also calculate the T_{C} of the 26 promising highT_{C} materials with much more accurate materialspecific U valuess^{30} [see Methods] and present the results in Sheet 3 of Supplementary Spreadsheet 2. Apart from a single material (MoIN_Pmmn), we observe that the T_{C} of the rest of the materials remains either close to 400 K or becomes much higher than that. We also noticed in some cases the T_{C} value has been significantly enhanced with the application of these tailormade U values. Thus, we expect that some of the materials whose T_{C} values fall in the 250–400 K range might exhibit much higher T_{C} if it is calculated using materialspecific U.
Machinelearning model
Due to an increasing amount of available data, machinelearning has recently found many applications in the field of solidstate materials science^{31}. Very recently, training on about 2500 experimentally reported Curie temperature of bulk materials, accurate ML models to predict T_{C} of bulk materials have been developed^{32}. 2D magnetism is fundamentally different from the magnetism of the bulk materials as most of the time anisotropy does not play such a significant role there. Therefore, In this emerging field, one doesn’t have the luxury of a sufficient amount of data points to train on. Based on the 157 data points obtained from our databasesearch we develop a machinelearning model to predict the T_{C} from the crystal structures. To decide the best model and features, we use the autoML library automatminer (https://github.com/hackingmaterials/automatminer). This tool takes structures as input and decorates the dataset with easily computable and chemically and physically meaningful features^{33}. Then the dataset is cleaned and reduced and is sent to the autoML library TPOT^{34}, which stochastically searches the model and the hyperparameter space using a genetic algorithm and finds the best model for the given dataset. After this extensive search [see Methods], we find an excellently fitted pipeline with average crossvalidation (CV) score 94.57 K^{2}, which is reported as the mean square error (MSE) on the training set. For the FM/AFM classification problem, we also try to find a suitable pipeline using the same method, with our examined 525 FM + AFM data points. The fitted pipeline reports an average CV score of a lowly 72.89% (accuracy) on the training set, which is understandable considering the complexity of the problem and the size and skewness of the data.
To test the generalization and predictive power of these ML pipelines, we construct a test set from reported 2DFM materials. Also, quite a few selected materials are included in the test set from a separate database^{17} with completely new structures and complex compositions. After inspecting these materials using our code, we find a few materials to be AFM which have been claimed as FM. This discrepancy can originate from the use of different DFT settings. This, along with fitting with a large number of spin configurations also causes a difference in fitted J values as observed for few materials (see Supplementary Fig. 3). Details of 22 materials identified as FM are provided in Sheet 2 of Supplementary Spreadsheet 2. The new prototypes encountered in this set are illustrated in Fig. 3.
Figure 4a illustrates the accuracy of the ML pipeline prediction against the DFTMC calculated T_{C} for all the train and test data. The ML predicted T_{C} is also listed in the ‘TC_exact predicted’ column of Supplementary Spreadsheet 2. The distribution of absolute errors in train and test data have been plotted in Fig. 4b. Although the pipeline has fitted to train data with high accuracy, generalization to test data does not seem so well. The MSE value also turns out to be 30335.46 K^{2}. Although the small train data size could be partially responsible for this, the main reason is possibly the introduction of unseen crystal structures. For instance, although we have not trained with even a single material containing La, the prediction for LaBr_{2} is exceptionally close probably because we had a lot of crystals with similar structures in our train data. The same argument can be applied to Mn_{2}H_{2}NO_{2} and Cr_{3}Te_{4}. The ML classifier pipeline has also been tried on the test data containing a total of 123 FM + AFM samples, which yields a decent 73.17% accuracy. Sheet 2 of Supplementary Spreadsheet 1 tabulates all the materials tried and their ultimate fate as well as the classification prediction.
During this exercise, we identify CrO_{2} _P4/mmm and ZnNi_{2}O_{5} as highT_{C} materials, while the claim of Cr_{3}Te_{4} possessing highT_{C} has also been verified, albeit the Curie temperature turns out to be much lower than the reported Ising model predicted value^{10}, but higher than the experimentally reported value for bulk Cr_{3}Te_{4}.
Violation of GoodenoughKanamori rule
The classical GoodenoughKanamori semiempirical rules^{19,20} for magnetic materials essentially states that when the magneticionanionmagneticion bond angle is close to 180°, a strong AFM structure due to superexchange should prevail, whereas if this angle is close to 90°, the material should show ferromagnetism. However, in this work, we notice quite a few violations of this rule. We have tabulated these bond angles for all the train and test materials in Supplementary Spreadsheet 2 and the apparent violation cases are marked in red. The most significant violation of the GoodenoughKanamori criterion can be seen in the strong highT_{C} FM material CrO_{2} _P4/mmm, where the cation–anion–cation bond angle is precisely 180°. Although a full investigation is out of the scope of this work, it appears that these postulates were developed for bulk materials and are failing here because of highly covalent bonds of 2DFM materials. Especially CrO_{2} _P4/mmm exhibits a highly planar structure, and thus should manifest highly covalent bonds. However, this material seems to be dynamically unstable and possesses a much higher total energy compared to the experimentally available bulk phase which might make it unlikely to be experimentally synthesized^{9}.
Hightemperature structural stability
According to the thermodynamic stability classification in C2DB, we choose three highly stable highT_{C} materials, namely CrIN_Pmmm, RhCl_{2}_C2/m, and Mn_{2}H_{2}CO_{2}_P3m1 along with newfound ZnNi_{2}O_{5}_Pmmn for further structural stability evaluation at high temperatures. An abinitio molecular dynamics (AIMD) run at 400 K for a total of 6.5 ps is conducted for this. Although CrIN, RhCl_{2}, and ZnNi_{2}O_{5} retain their crystal structure during this, albeit with less crystallinity, the MXene Mn_{2}H_{2}CO_{2} starts to melt away just after 3 ps, rendering it unsuitable for practical applications. Supplementary Figs 7–10 show the structural differences resulted from the MD runs.
Discussion
In this study, we predict a total of 26 2D materials to have T_{C} beyond 400 K. Many of these could be easily synthesizable, either by straight exfoliation from their bulk counterparts or by bottomup chemical methods^{24,25}. It is worth noting that, low thermodynamic stability does not necessarily mean the material would be unusable for practical applications. For instance, according to C2DB, Silicene is classified as a material with low thermodynamic stability. However, Silicene transistors have been demonstrated to work in room temperature^{35}. Some of these materials screened by us even show dynamic instability. Again, many commonly used 2D materials, such as Tphase MoS_{2} showing dynamic instability^{36} in freestanding form stabilizes themselves on substrates through possible substrate interaction and even finds application in roomtemperature devices^{37}. Also, these materials can show charge density wave (CDW) characteristics and stabilize in a larger supercell with slight structural distortion^{7}. However, CDW distortions can sometimes impact the magnetic order adversely^{38}.
To summarize, using highthroughput automated codes and datadriven models, we thoroughly screen 2D materials databases and predict a host of 2DFM materials with high Curie point. With the emergence of novel synthesis techniques, these materials could indeed be of interest to experimentalists and engineers in terms of practical application in various devices. The ML model and the automated code developed in this work could find use in the community for rapid magnetic property prediction. The model complements the rigorous DFTMC based code and if trained with sufficiently large datasets the model could eventually replace the code^{32}. Stateoftheart software engineering enables us to achieve an optimal balance between rigor and computational efficiency, which is very important for reliable highthroughput material screening. As a result, we discover many important magnetic materials involving metals like Mo, W and Ti which have so far been ignored by heuristicbased formulascreening^{10}.
Methods
Spin configurations generation
As mentioned before, the python library pymatgen^{21} has been extensively used in this study for generating spin configurations, managing, and parsing inputoutput files and performing the neighbour mappings. The python module ASE^{39} has also been used to parse the input structure files. To ensure the reliability and coverage of the pymatgen generated spin configurations, we manually verify that the codegenerated spin configuration set almost always includes all heuristicsbased configurations reported in the literature^{9,10,11,12,13,14}. Often, the code generates even more unexplored but symmetrically valid configurations that we leverage to include a higher number of neighbour interactions for a more accurate prediction of the T_{C}. At the same time, we ignore ferrimagnetic configurations since these are usually energetically highly unstable as well as asymmetric. For instance, in contrast to the previous report^{10}, our T_{C} prediction of Cr_{3}Te_{4} is based on a larger number of symmetric FM and AFM configurations and thus seems to be much more accurate_{.}
It is worth mentioning that since we cover such a huge variety of materials using an automated workflow, we use such values of various DFT and numerical parameters [see Supplementary Readme (readme file at OSF repository)], which would yield reasonable results for all materials. For instance, in case of CrI_{3}, the default value of the parameter enum_prec = 0.001 (enum_precision_parameter in pymatgen) generates 3 configurations, where the Néel AFM configuration gets excluded (Supplementary Fig 1). But with enum_prec = 1e7, all 4 configurations can be generated. For the first case, we obtain J_{1} = 2.78 and J_{2} = 0.43 meV/link, while the second set gives us almost identical values of J_{1} = 2.82, J_{2} = 0.41 and J_{3} = 0.009 meV/link and in both cases, we obtain the same T_{C}. The Néel AFM solution turns out to be the most energetically unfavourable state and gets truncated^{18} by pymatgenenumlib with default settings. For this highthroughput study, we find the default values to be accurate enough for our onefitsall scheme. However, for focused studies on some specific material one might want to obtain extremely accurate results and thus might need to tune the parameters a little.
DFT parameters
Because the energy differences between various configurations for magnetism calculations could be as low as ≈µeV, the computations need to be performed with high accuracy. We use heavily modified versions of predefined configurations ‘MPRelaxSet’, ‘MPStaticSet’ and ‘MPSOCSet’ available in pymatgen for relaxations, static runs and MAE calculations. These modifications, as well as other details of DFT, are highlighted below.
Spinpolarized DFT calculations are carried out using generalized gradient approximation (GGA) as implemented in the code VASP^{40} with projector augmentedwave (PAW)^{41} method using the Perdew–Burke–Ernzenhof (PBE)^{42} exchangecorrelation functional. Along with the CPU version, The GPU port^{43} of VASP has been used extensively. For all calculations, a correction on the strongly correlated dshell electrons (GGA + U) is applied using the Dudarev^{44} formulation. The default value of the cutoff energy (520 eV) is used which proved to be sufficiently large. For relaxations, the default reciprocal density of 64 Å^{−3} is employed whereas for all collinear and noncollinear static runs a much denser reciprocal density of 300 Å^{−3} has been used. Electronic convergence is set to be attained when the difference in energy of successive electronic steps becomes less than 10^{−6} eV, whereas the structural geometry is optimized until the maximum Hellmann–Feynman force on every atom falls below 0.01 eV/Å. For the highprecision MAE calculations, a stricter electronic convergence criterion of 10^{−8} eV is imposed. A large vacuum space of >25 Å in the direction of c is applied to avoid any spurious interaction between periodically repeated layers. The Bader charge and magnetization analysis are performed using the code developed by the Henkelman group^{45}, where charge densities generated from DFT static runs are used as inputs. These Bader partitioned magnetic moments have been used as the local magnetic moments of the magnetic elements. All crystal structure images are generated using the tool VESTA^{46}.
For metals, Co, Cr, Fe, Mn, Mo, Ni, V and W, the effective U values have been taken from the Materials Project (https://wiki.materialsproject.org/GGA%2BU_calculations#Calibration_of_U_values) where these effective U values have been calibrated by performing a fitting to experimental binary formation enthalpies^{47}. This is an established practice and has been used in similar highthroughput screening studies before^{10}. We also find that the application of an effective U is essential to perform accurate DFT calculations on materials with “low” magnetization (materials containing Nb, Sc, Ru, Rh, Pd, Cu, Os, Ti, Zr, Re, Hf, Pt, and La) as the AFM solutions become difficult to obtain for these materials without a proper effective U. In these cases, for a specific metal, first the effective U values are obtained using the linearresponse approach^{30} using 3 × 3 supercells for a few materials containing the element and then the average value of these effective U is taken as the final effective U. A complete list of effective U values (in eV) used in this study, as well as the DFT parameters imposed can be found in Supplementary Code 1 (e2e.py at OSF repository).
It is worth noting that, these effective U values depend on the element type, charge state and coordination mode of magnetic species in a certain material, which implies that the value of U is quite materialspecific and should be determined carefully for accurate predictions. Thus, we also calculate the materialspecific U values for the most promising 26 materials with predicted highT_{C} using the linearresponse approach^{30}. We observe that the U values obtained from the linearresponse method can be quite different from the highthroughput U values we have used so far (see Sheet 3 of Supplementary Spreadsheet 2). However, the computational budget of the linearresponse method is excessively high to be adopted for highthroughput material screening.
Hamiltonian fitting
The coupling constant (J) values are fitted using the collinear energy values of different FM and AFM spin configurations. However, for a lot of cases, when the energies of all configurations are taken, the determinant of the system of equations becomes zero which makes the set of equations unsolvable. As an automated remedy for these problems, again a fitting is tried omitting the most unstable AFM configuration and using one less neighbour than before. This process is repeated until a set of physically meaningful solutions is found, or the code runs out of configurations to fit. The anisotropic constant (k) values are fitted using the noncollinear energy values with spins oriented in different directions ([100], [010], [110], [001]).
Despite our best efforts, calculations for 261 materials (out of the 786 materials classified as FM in C2DB) had to be cancelled because of the following reasons: (1) pymatgen could not recognize the symmetry and could generate only one configuration, (2) material turned out to be nonmagnetic after DFT calculations, (3) severe convergence issues occurred during DFT calculations, (4) AFM configurations could not be retained even after application of proper U and manual tuning of parameters, (5) the Hamiltonian could not be fitted properly, (6) phase change of crystal structure after relaxation.
Monte Carlo simulation
To study ferromagnetic (FM) to paramagnetic (PM) transition in these monolayer materials, Monte Carlo (MC) simulations of the Heisenberg model have been performed using the Metropolis algorithm with singlespin update scheme^{48}. To eliminate the size effects, a 50 × 50 supercell containing 25008000 sites has been used to simulate the system. Total 10^{5} Monte Carlo steps have been performed for each temperature, while the results from the first 10^{4} steps have been discarded, as the system is allowed to equilibrate (thermalize) during this time. The final values of magnetization and susceptibility are calculated as the average over the last 9 × 10^{4} MC steps for each temperature.
Software engineering
We develop the complete endtoend code in python to take advantage of pymatgen. However, python being an interpreted language, the MC simulations turned out to be excessively slow, especially with high coordination numbers and inclusion of higher neighbours, which made the code unsuitable for the highthroughput study. As a remedy to this problem, we decide to use pythonbased justintime (JIT) compiler numba^{49} which compiles specific decorated python modules at the first encounter to lowlevel instructions, and when these modules are repeatedly called, the compiled version is used which makes the code extremely fast. However, the tradeoff is, a lot of powerful functions and the coding flexibility offered by python (like heterogeneous data structures, appending to a list) cannot be successfully compiled and significant software engineering, as well as timing and costbenefit analyses, are required to achieve an optimal code. GPU acceleration for neighbourmapping of large lattices (2500–8000 sites) has also been implemented, which on a CPU must be done serially and takes a lot more time.
The engineered code was optimized to such an extent that the whole study could be performed using even a workstationgrade machine, albeit with GPU acceleration. A video of realtime execution of the code, where the T_{C} of four materials (CrI_{3}, Cr_{2}Ge_{2}Te_{6}, MnSe_{2,} and MoC_{3}) are being calculated in parallel in a singleCPU (18 cores), threeGPU enabled workstation within ≈10 hour, can be found at https://youtu.be/HJkR03OzBI. At the same time excellent scalability is observed, when the code is executed on a high performance computing node (https://youtu.be/GQaFfm29LR4).
Machine learning
The python libraries automatminer and matminer^{33} have been used to featurize the datasets and search for optimal ML pipeline for the FM/AFM classification problem as well as the T_{C} predicting regression problem. Dataset cleaning and feature reduction are handled by automatminer. Then, various preprocessing algorithms along with a host of commonly used ML models employed for materials science problems^{31} for small to moderate datasets have been searched by the autoML library TPOT^{34}, such as naive Bayesian, decision tree, extra trees, random forest, gradient boosting, kneighbors, linear SVC and logistic regression for classification and elastic net CV, decision tree, extra trees, random forest, gradient boosting, kneighbors, lasso lars CV, ridge CV and linear SVR for regression. Also, the hyperparameters are tuned at the same time. A full list of preprocessing algorithms, ML models and hyperparameters searched can be found at https://github.com/hackingmaterials/automatminer/blob/master/automatminer/automl/config/tpot_configs.py. It is worth noting that TPOT uses the ML library scikitlearn^{50} for the ML as well as the featureengineering models. For each autoML search, more than 60,000 pipelines have been explored.
The generic python codes using scikitlearn for the best pipelines for both FM/AFM classification and T_{C} regression have been given in Supplementary Codes 2 and 3 (TPOT_*.py at OSF repository). For the former, a combination of SelectPercentile, MaxAbsScaler, and ExtraTreesClassifier turns out to be the best pipeline, whereas for the latter it is a combination of SelectPercentile, ZeroCount, and GradientBoostingRegressor. The list of features used for these problems can be found in Supplementary Logs 1 and 2 (*_digest at OSF repository). Moreover, the pickled pipelines have also been provided in Supplementary ML Pipes 1 and 2 (*.pipe at OSF repository) which can be loaded into automatminer to make predictions on any dataset. The Supplementary Readme file provides detailed instructions on how to use the Supplementary codes and pipes.
AIMD simulations
For the chosen materials, to minimize the temperature oscillations a large supercell containing ≥144 atoms has been constructed to run the AIMD on. Because of severe convergence issues, nonspinpolarized DFT calculations are performed with Gammapoint only sampling which can be sufficient to determine structural stability. A canonical ensemble (NVT) is used and a NoséHoover thermostat^{51,52} at 400 K is employed. The simulations run for 6.5 ps with a 2 fs time step. The crystal structures of the tested materials after the simulation can be seen in Supplementary Figs 7–10.
Data availability
The authors declare that the main data supporting the findings of this study are available within the paper and its Supplementary files, the OSF repository (https://osf.io/6ebjp/) and other open online resources. Other relevant data are available from the corresponding author upon reasonable request.
Code availability
All relevant python codes and pickled ML pipelines are provided in the OSF Repository (https://osf.io/6ebjp/). Also, the endtoend code for Curie point determination (e2e.py) is available at GitHub (https://github.com/NSDRLIISc/e2e), which we plan to update periodically.
References
Huang, B. et al. Layerdependent ferromagnetism in a van der Waals crystal down to the monolayer limit. Nature 546, 270–273 (2017).
Gong, C. et al. Discovery of intrinsic ferromagnetism in twodimensional van der Waals crystals. Nature 546, 265 (2017).
Feng, Y. P. et al. Prospects of spintronics based on 2D materials. Wiley Interdiscip. Rev. Comput. Mol. Sci. 7, e1313 (2017).
Farooq, M. U. & Hong, J. Switchable valley splitting by external electric field effect in graphene/CrI_{3} heterostructures. npj 2D Mater. Appl. 3, 3 (2019).
Soumyanarayanan, A., Reyren, N., Fert, A. & Panagopoulos, C. Emergent phenomena induced by spinorbit coupling at surfaces and interfaces. Nature 539, 509–517 (2016).
Mermin, N. D. & Wagner, H. Absence of ferromagnetism or antiferromagnetism in one or twodimensional isotropic Heisenberg models. Phys. Rev. Lett. 17, 1133–1136 (1966).
Bonilla, M. et al. Strong roomtemperature ferromagnetism in VSe_{2} monolayers on van der Waals substrates. Nat. Nanotechnol. 13, 289–293 (2018).
O’Hara, D. J. et al. Room temperature intrinsic ferromagnetism in epitaxial manganese selenide films in the monolayer limit. Nano Lett. 18, 3125–3131 (2018).
van Gog, H. et al. Thermal stability and electronic and magnetic properties of atomically thin 2D transition metal oxides. npj 2D Mater. Appl. 3, 18 (2019).
Zhu, Y., Kong, X., Rhone, T. D. & Guo, H. Systematic search for twodimensional ferromagnetic materials. Phys. Rev. Mater. 2, 81001 (2018).
Kumar, H. et al. Tunable magnetism and transport properties in nitride MXenes. ACS Nano 11, 7648–7655 (2017).
Jiang, Z., Wang, P., Xing, J., Jiang, X. & Zhao, J. Screening and design of Novel 2D ferromagnetic materials with high curie temperature above room temperature. ACS Appl. Mater. Interfaces 10, 39032–39039 (2018).
Zhuang, H. L., Xie, Y., Kent, P. R. C. & Ganesh, P. Computational discovery of ferromagnetic semiconducting singlelayer CrSnTe_{3}. Phys. Rev. B 92, 35407 (2015).
Miao, N., Xu, B., Zhu, L., Zhou, J. & Sun, Z. 2D intrinsic ferromagnets from van der Waals antiferromagnets. J. Am. Chem. Soc. 140, 2417–2420 (2018).
Mounet, N. et al. Twodimensional materials from highthroughput computational exfoliation of experimentally known compounds. Nat. Nanotechnol. 13, 246–252 (2018).
Haastrup, S. et al. The Computational 2D Materials Database: highthroughput modeling and discovery of atomically thin crystals. 2D Mater. 5, 42002 (2018).
Zhou, J. et al. 2DMatPedia, an open computational database of twodimensional materials from topdown and bottomup approaches. Sci. Data 6, 86 (2019).
Horton, M. K., Montoya, J. H., Liu, M. & Persson, K. A. Highthroughput prediction of the groundstate collinear magnetic order of inorganic materials using density functional theory. npj Comput. Mater. 5, 64 (2019).
Goodenough, J. B. An interpretation of the magnetic properties of the perovskitetype mixed crystals La_{1−x}Sr_{x}CoO_{3−λ}. J. Phys. Chem. Solids 6, 287–297 (1958).
Kanamori, J. Superexchange interaction and symmetry properties of electron orbitals. J. Phys. Chem. Solids 10, 87–98 (1959).
Ong, S. P. et al. Python Materials Genomics (pymatgen): a robust, opensource python library for materials analysis. Comput. Mater. Sci. 68, 314–319 (2013).
Torelli, D. & Olsen, T. Calculating critical temperatures for ferromagnetic order in twodimensional materials. 2D Mater. 6, 15028 (2018).
Ashton, M. et al. Twodimensional intrinsic halfmetals with large spin gaps. Nano Lett. 17, 5251–5257 (2017).
Zhang, J. et al. Janus monolayer transitionmetal dichalcogenides. ACS Nano 11, 8192–8198 (2017).
Sreepal, V. et al. Twodimensional covalent crystals by chemical conversion of thin van der Waals materials. Nano Lett. 19, 6475–6481 (2019).
Fei, Z. et al. Twodimensional itinerant ferromagnetism in atomically thin Fe_{3}GeTe_{2}. Nat. Mater. 17, 778–782 (2018).
Zhuang, H. L., Kent, P. R. C. & Hennig, R. G. Strong anisotropy and magnetostriction in the twodimensional Stoner ferromagnet Fe_{3}GeTe_{2}. Phys. Rev. B 93, 134407 (2016).
Wang, X. et al. Raman spectroscopy of atomically thin twodimensional magnetic iron phosphorus trisulfide (FePS_{3}). Cryst. 2D Mater. 3, 31009 (2016).
Kim, K. et al. Suppression of magnetic ordering in XXZtype antiferromagnetic monolayer NiPS_{3}. Nat. Commun. 10, 345 (2019).
Cococcioni, M. & de Gironcoli, S. Linear response approach to the calculation of the effective interaction parameters in the LDA+U method. Phys. Rev. B 71, 35105 (2005).
Schmidt, J., Marques, M. R. G., Botti, S. & Marques, M. A. L. Recent advances and applications of machine learning in solidstate materials science. npj Comput. Mater. 5, 83 (2019).
Nelson, J. & Sanvito, S. Predicting the Curie temperature of ferromagnets using machine learning. Phys. Rev. Mater. 3, 104405 (2019).
Ward, L. et al. Matminer: an open source toolkit for materials data mining. Comput. Mater. Sci. 152, 60–69 (2018).
Olson, R. S., Bartley, N., Urbanowicz, R. J. & Moore, J. H. Evaluation of a treebased pipeline optimization tool for automating data science. In Proc Genetic and Evolutionary Computation Conference 2016 (ed. Friedrich, T.) 485–492 (ACM, 2016).
Tao, L. et al. Silicene fieldeffect transistors operating at room temperature. Nat. Nanotechnol. 10, 227–231 (2015).
Saha, D. & Mahapatra, S. Atomistic modeling of the metallictosemiconducting phase boundaries in monolayer MoS_{2}. Appl. Phys. Lett. 108, 253106 (2016).
Kappera, R. et al. Phaseengineered lowresistance contacts for ultrathin MoS_{2} transistors. Nat. Mater. 13, 1128 (2014).
Zheng, F., Zhou, Z., Liu, X. & Feng, J. Firstprinciples study of charge and magnetic ordering in monolayer NbSe_{2}. Phys. Rev. B 97, 81101 (2018).
Hjorth Larsen, A. et al. The atomic simulation environment—a Python library for working with atoms. J. Phys. Condens. Matter 29, 273002 (2017).
Kresse, G. & Furthmüller, J. Efficiency of abinitio total energy calculations for metals and semiconductors using a planewave basis set. Comput. Mater. Sci. 6, 15–50 (1996).
Kresse, G. & Joubert, D. From ultrasoft pseudopotentials to the projector augmentedwave method. Phys. Rev. B 59, 1758–1775 (1999).
Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized Gradient Approximation Made Simple. Phys. Rev. Lett. 77, 3865–3868 (1996).
Hacene, M. et al. Accelerating VASP electronic structure calculations using graphic processing units. J. Comput. Chem. 33, 2581–2589 (2012).
Dudarev, S. L., Botton, G. A., Savrasov, S. Y., Humphreys, C. J. & Sutton, A. P. Electronenergyloss spectra and the structural stability of nickel oxide: An LSDA+U study. Phys. Rev. B 57, 1505–1509 (1998).
Hu, L. et al. Twodimensional materials from highthroughput computational exfoliation of experimentally known compounds. Phys. Rev. B 6, 1–7 (2017).
Momma, K. & Izumi, F. VESTA3 for threedimensional visualization of crystal, volumetric and morphology data. J. Appl. Crystallogr 44, 1272–1276 (2011).
Wang, L., Maxisch, T. & Ceder, G. Oxidation energies of transition metal oxides within the GGA+U framework. Phys. Rev. B 73, 195107 (2006).
Vatansever, E. et al. Strain effects on electronic and magnetic properties of the monolayer α  RuCl_{3}: A firstprinciples and Monte Carlo study. J. Appl. Phys. 125, 83903 (2019).
Lam, S. K., Pitrou, A. & Seibert, S. Numba: a LLVMbased Python JIT Compiler. In Proc. 2nd Workshop on the LLVM Compiler Infrastructure in HPC (ed. Finkel, H.) 7:1–7:6 (ACM, 2015).
Pedregosa, F. et al. Scikitlearn: Machine Learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011).
Nosé, S. A unified formulation of the constant temperature molecular dynamics methods. J. Chem. Phys. 81, 511–519 (1984).
Hoover, W. G. Canonical dynamics: equilibrium phasespace distributions. Phys. Rev. A 31, 1695–1697 (1985).
Acknowledgements
The authors are grateful to Dr. Naihua Miao, Beihang University and Prof. Ümit Akıncı, Dokuz Eylül University for several insightful discussions. Authors also thank Abdul Samadh, Manipal University, and Kota Hemanth, IIT Delhi, for their help with software engineering and machine learning. The authors thank Supercomputer Education and Research Centre (SERC), Indian Institute of Science (IISc) Bangalore, for CPU based computations. The GPU computing nodes used in this study were procured through the research funding from the Department of Science and Technology (DST), Government of India, under grant number DST/TMD/MES/2K18/28.
Author information
Authors and Affiliations
Contributions
A.K. developed the automated endtoend code, implemented it on CPU and GPU based architectures, developed the ML models and analysed the data. M.K. scanned the materials using the automated code, collected and verified the results. S.M. conceived the problem statement and overall supervised the work. All authors contributed to the writing.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Kabiraj, A., Kumar, M. & Mahapatra, S. Highthroughput discovery of high Curie point twodimensional ferromagnetic materials. npj Comput Mater 6, 35 (2020). https://doi.org/10.1038/s4152402003002
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s4152402003002
This article is cited by

Realizing unipolar and bipolar intrinsic skyrmions in MXenes from highfidelity firstprinciples calculations
npj Computational Materials (2023)

Highthroughput computation and structure prototype analysis for twodimensional ferromagnetic materials
npj Computational Materials (2022)

Highthroughput design of functionalengineered MXene transistors with lowresistive contacts
npj Computational Materials (2022)

Highthroughput inverse design and Bayesian optimization of functionalities: spin splitting in twodimensional compounds
Scientific Data (2022)

Describing chainlike assembly of ethoxygroupfunctionalized organic molecules on Au(111) using highthroughput simulations
Scientific Reports (2021)