betaFeedback

WorksPhysics and Astronomy

Active Learning Guided Computational Discovery of 2D Materials with Large Spin Hall Conductivity

Abstract

Two-dimensional (2D) materials are promising candidates for next-generation spintronic devices due to their tunable properties and potential for efficient spin-charge interconversion. However, discovering materials with intrinsically high spin Hall conductivity (SHC) is hindered by the vast chemical space and expensive nature of conventional experimental and first-principles methods. In this work, we employ an active learning framework to accelerate the discovery of high-SHC 2D materials. Machine learning (ML) models were trained on SHC values computed from density functional theory calculations, incorporating the Kubo formalism via tight-binding Hamiltonians constructed from maximally localized Wannier functions, with explicit treatment of spin-orbit coupling. Starting from random but chemically diverse 24 2D systems, the dataset was expanded to 41 cases (from an overall pool of around 2000 materials) over three active learning loops using an expected improvement acquisition strategy. The ML technique successfully identified several high SHC candidates with the best candidate exhibiting a SHC of 271.52 (hbar/e) Ohm^-1, nearly 23 times higher than the top performer in the initial round. Beyond candidate discovery, several features such as orbital symmetry near the Fermi energy, types of atomic species, material composition, covalent radii, and electronegativity of constituent atoms were found to play critical role in shaping the spin Hall response in 2D systems. The data generated is made publicly available to facilitate further advances in 2D spintronics.

Manuscript.pdf

Introduction

Spintronics devices manipulate spin degree of freedom of electrons to process, transport, compute, store information, and generate signals. The deterministic factor for many spintronic technologies is the spin Hall effect (SHE), wherein longitudinal charge current (IcI_{\mathrm{c}}) is converted into transverse spin current (IsI_{\mathrm{s}}) via intrinsic spin-orbit coupling (SOC)[1]. This paves the way for easier utilization of non-magnetic materials in the spintronic industry. Importantly, the generated pure spin flow, IsI_{\mathrm{s}}, does not involve flow of net charge and thus energy efficient information processing can be realized without Joule heating losses[2]. Moreover, upon diffusion of IsI_{\mathrm{s}} into an adjacent ferromagnetic (FM) layer, its angular momentum exerts a torque to either oscillate or switch magnetization of the FM layer, enabling spin-orbit torque functionalities such as non-volatile data storage and processing, spin logic devices, and neuromorphic computing, among others[3].

Extensive computational and experimental efforts have been dedicated to discovering materials with high SHC. Chronologically, this search has so far focused on 3D bulk materials such as heavy 5d transition metals[4, 5, 6], GaAs[7], and bulk transition metal dichalcogenides (TMDs)[8, 9]. Over the last decade, the emerging class of 2D materials has demonstrated promising spin-driven quantum phenomena and hence form a natural platform to investigate SHE[10, 11, 12]. Furthermore, for efficient and scalable spintronic devices, materials are expected to show significant spin-related properties such as spin diffusion length, spin relaxation time, and spin coherence. The flagship 2D material - graphene - is theoretically estimated to exhibit spin diffusion length and relaxation time a few orders of magnitude higher than traditional conductors[13]. However, its direct usage as spin channel or Hall layer is limited due to weak SOC and as a poor host of FM state. Thus, efforts have been made to introduce FM state in graphene via doping[14, 15], introducing vacancies[16, 17, 18, 19], and through magnetic proximity[20, 21]. Another widely adopted research direction in 2D spintronics is the coupling of 2D TMDs (with high SOC and broken inversion symmetry) with FM alloy[22, 23] or FM van der Waals materials[24] in the form of heterostructures. Huge chemical diversity, large SOC due to more confined and correlated electrons[25, 26, 27], bandgap tunability via external factors (electric field, strain, etc.), and easier integration with present industrial infrastructure[28] are some reasons why 2D materials are generally considered promising for spintronic applications.

The SHC is influenced by both intrinsic and extrinsic mechanisms. The intrinsic contribution arises from the SOC-influenced electronic band structure (EBS), while the extrinsic contribution originates from scattering processes due to impurities, defects, and other forms of disorder. Notably, the intrinsic SHC can be estimated reliably using DFT with incorporation of SOC[29, 30]. The SOC splits and shifts the band to give non-zero Berry curvature, which serves as a proxy magnetic field in the momentum space. Such split can differentiate the spin population leading to a net spin current[9, 23]. Motivated by the search for materials with large SHC, several high-throughput computational studies have evaluated the intrinsic SHC across a broad range of bulk materials[31, 32, 33]. More recently, similar efforts have been extended to 2D systems, including a dataset of 426 rare-earth-free monolayers[34]. While such high-throughput approaches are invaluable for large-scale screening, their accuracy may be compromised due to algorithmic simplifications and coarse approximations made to accelerate the calculations. For instance, a crucial aspect of SHC computations is the construction of tight-binding (TB) models using maximally localized Wannier functions (MLWFs) derived from ab initio calculations[35]. In high-throughput workflows, the generation of MLWFs is typically automated,[36, 37, 38] a process which may be prone to errors particularly for dispersive band structures, where it is important to correctly identify orbitals to be considered for projections. Thus, the quality of wannierization can significantly affect the accuracy of SHC computations.

To overcome the limitations of the existing computational approaches and to efficiently screen 2D materials with high SHC values, in this work, we develop an active learning framework that combines DFT calculations, TB modeling and ML methods. The ML models uses the concept of expected improvement to recommend candidates from a large pool of nearly 2000 2D materials with potential for high SHC. Selected candidates are then further evaluated using combined DFT and TB modeling approach, incorporating user-guided construction of MLWF to ensure accurate SHC evaluation. The complete process is carried iteratively in an active learning framework with around 5-7 candidates examined in each round using the aforementioned electronic structure methods. Overall, across three active learning cycles, we investigated 41 materials and observed a consistent improvement in SHC values in each round, culminating in the discovery of several novel compounds with SHC significantly exceeding the known materials. Notably, the top-performing material, Fe2TeSe\mathrm{Fe}_2\mathrm{TeSe}, was found to have SHC nearly 23 times greater than the best candidate in the initial round. The SHC data and the associated ML models were also analyzed to extract key chemical trends that result in high SHC. Contribution of dd- electrons near to the Fermi energy (EFE_{\mathrm{F}}), difference in the covalent radius, and the number of mirror symmetry planes are some example factors that were found to significantly impact the SHC. Overall, this work not only led to successful identification of 2D materials with high computed SHC, but it also shows how materials informatics approaches can be combined with accurate but expensive physics-based models to efficiently search for materials from large chemical spaces.

RESULTS and DISCUSSION

Figure 1 shows the schema of our workflow that incorporates DFT, empirical TB modeling and data-driven ML method to discover 2D materials with high SHC. Overall, the role of ML methods is to learn the complex relationship between SHC and fundamental features of materials emerging from their crystal symmetry and electronic structure of individual elements, and recommend potential candidates expected to exhibit large SHC from a large 2D materials database. The overall search for 2D spintronic materials with high SHC is partitioned into three parts. First, the EBS of a 2D material taken from cleaned version of Materials Cloud Two-Dimensional Structure Database (MC2D)[39, 40] is calculated using DFT. As the second step, the minimal basis set based TB modeling is then adopted using MLWF formalism to compute the Berry curvature and thereby the SHC using the Kubo formalism. The third part comprises of leveraging ML techniques to train models on the dataset of computed SHC values, and screening potential candidates expected to exhibit large SHC from a vast pool of 2D systems. To achieve this, we first created an initial training dataset of 24 diverse 2D systems and computed their SHC values. A range of hand-crafted symmetry features as well as elemental features from standard matminer[41] library were then used to train a series of traditional regression models on this dataset. Since the range of SHC values span multiple orders of magnitude, its symmetric log transformation was used as the target ML property. The top performing ML model (ridge regression) was then used to make predictions on the candidate dataset of ∼ 2000 2D material systems from the cleaned MC2D database. More details on DFT, TB and ML modeling as well as cleaning process for the MC2D database are included in the Methods section.

Figure 1. Active machine learning framework to identify 2D materials with high SHC. The workflow primarily consists of DFT (for EBS computation), MLWF (for minimal basis set formation), TB modeling (for computation of Berry curvature and thereby SHC) and ML techniques (for SHC prediction and material selection). First, 2D structures are obtained from MC2D database for which EBS is determined. This is followed by MLWF construction using projection matrices and overlap matrices generated from EBS using Wannier90 code[42]. The best-fit TB model Hamiltonian is then set up using MLWF basis to obtain Berry curvature and SHC. These values are used to create training dataset which also comprises of 26 hand-crafted symmetry and electronic features as well as 132 matminer library-based elemental features. The ML models are then trained using 5-fold cross validation (CV), recursive feature elimination and symmetric log scaling of the target property. The model with lowest CV error is used to screen 5-7 best candidates using expected improvement acquisition strategy for the next round of DFT–SOC+TB computations. The freshly computed SHC data is are added to the training dataset and next active learning loop is initiated. The final ML model is later subjected to SHAP feature analysis to extract important chemical insights.

Active learning framework was adopted to efficiently find 2D materials with high SHC. For this, expected improvement (EI) acquisition function was adopted which is known to balance the trade-off between exploration (materials with high prediction uncertainty) and exploitation (materials with high predicted SHC value). Once the ML predictions are available with their corresponding EI values, a few systems (5-7) with high EI values were shortlisted for further SHC computations using the DFT–SOC+TB scheme. Besides EI, the candidates were also selected based on their diversity, i.e., based on number of elements (binary, ternary or quaternary systems) or type of elements (e.g. metal, non-metal). The computed SHC values were again added to the initial training dataset to retrain next version of ML model with better prediction accuracy. This active learning loop were performed for three iterations. Furthermore, to interpret the final trained ML model and the factors that contribute to high SHC in 2D materials, SHapley Additive exPlanations (SHAP)[43] framework was utilized.

Figure 2 presents the full set of 2D materials investigated across the three active learning rounds, along with their DFT–SOC+TB computed SHC values. In Round 1 (Figure 2(a)), 24 diverse systems were selected based on domain knowledge and prior literature to capture chemical, structural and property diversity. Chemical diversity was introduced by selecting material systems from diverse families of metal halides (4), halonitrides (2), chalcogenides (6), transition metal dichalcogenides (7) and others (5) such as oxychloride, azide, etc.; the numbers in parentheses indicate the count per category. Diversity in terms of various crystal symmetries (12 different space groups, 18 symmorphic, 6 non-symmorphic, 16 centrosymmetric), symmetry elements (inversion, roto-inversion, etc.) as well as polymorphism (two phases of ZrNI) were also considered. Systems spanning low to high SOC and metallic to semiconducting behavior were also included. The t-SNE map (Figure S1) further illustrates the compositional diversity of the Round 1 training data. Figure 2(a) also highlights 2D materials selected in subsequent rounds based on EI values for detailed DFT–SOC+TB computations. These candidates were selected from a large candidate pool of ∼2000 2D materials, encompassing diverse range of elements (heavy/light or ss, pp, dd-block), structures from 58 different space groups (SGs), and of compositions ranging from binary to hexanary. Notably, the ML model trained on Round 1 data, which only contains binary materials systems, was successfully used to make predictions for ternary to hexanary materials. Since DFT–SOC+TB computations become exceedingly expensive for pentanary and hexanary systems[44], we restricted our ML selection to binary, ternary and quaternary systems.

Figure 2. Performance of active learning rounds: (a) 2D systems selected across different rounds for SHC (or THC) computations using active learning approach. While diverse candidates were manually selected in Round 1, EI was used for selection in Rounds 2.1, 2.2 and 3. The number in parenthesis next to the system name indicates its rank when ordered based on absolute SHC value. (b) Variation in joint kernel density and multi-class scatter plot of THC and SHC with different active learning rounds in absolute symmetric log space. (c) The box-and-whisker plot of modulus of SHC in symmetric log space with horizontal lines in the box showing median, first and third quartiles whereas whiskers (range of the data) are 1.5 times inter-quartile range. The data points beyond the whiskers are outliers marked as circles.

Figure S1. The t-SNE map showing distribution of various material systems selected during different rounds of active learning loop. The Round 1 materials (gray points) are well spread across the feature space, indicating good initial training diversity.

Generally, it is difficult to distinguish between Hall conductivity originating from spin or orbital interactions in experimental studies[45]. However, for the purpose of preliminary screening, we defined total Hall conductivity (THC) as sum of SHC and OHC as a quantitative metric. Accordingly, in our first active learning loop we attempted to build ML models to predict both THC as well as SHC. Round 2.1 and Round 2.2 contain candidates that were selected based on EI values for THC and SHC, respectively, obtained using the same initial training dataset (Round 1) of 24 systems. The actual THC and SHC values for these candidates were subsequently computed from the electronic structure. As can be observed from Figure 2(b), Round 2.2 successfully resulted in identification of candidates with high SHC, much higher than the those in the initial training data. This is evident from the right shift of the orange colored region (Round 2.2) as compared to the black colored region (Round 1). However, this is not the case for THC where green colored region (Round 2.1) is seen to marginally shift up compared to black colored region (Round 1). While it is unclear why the performance of the ML model is superior for the case of SHC, a potential cause could be the feature set used in this work. Thus, for the subsequent rounds we exclusively focused on design of materials with high SHC only. The SHC data from 36 candidates of Round 1 and 2 (i.e., both Rounds 2.1 and 2.2) was used to train next round of ML model and screen candidates for Round 3. Again, the right shift of blue region (Round 3) compared to orange region (Round 2) indicates the success of the active learning approach to find candidates with higher SHC. The success of the active learning approach to continually find candidates with higher and higher SHC with each subsequent loop is also illustrated using the box-and-whiskers plot in Figure 2(c). This clearly demonstrates the power of combining active ML approaches with traditional computational methods to accelerate discovery of materials. More details on the ML model performance and the identified top candidates is presented next.

ML model performance

In each active learning loop, three types of ML models were trained, namely, ridge regression, support vector regression and Gaussian process regression. Ridge regression consistently yielded the lowest 5-fold cross-validation errors across all rounds (see Figure S2 in Supplementary Information-1) and was thus selected for candidate screening in each cycle. Model uncertainty was estimated by training 100 model variants using different random 80% subsamples of the training data.

Figure S2. Performance of different machine learning (ML) models. Parity plots indicating performance of different ML models trained on SHC data across different active learning rounds. Results for 3 different regression models, namely, ridge regression (Ridge), Gaussian process regression (GPR) and support vector regression (SVR) are shown, with SHC values plotted in symmetric log space. Colored markers represent particular CV fold, while red dashed line indicate the ideal parity line. Performance metrics for the overall 5-fold CV, including root mean square error (RMSE) and coefficient of determination (R2R^2), are reported. Ridge regression can be seen to consistently outperform other models, achieving lower RMSE and higher R2R^2 values.

A more stringent and pertinent measure of performance of the ML model is to compare the ML model predictions for screened candidates against their subsequent DFT–SOC+TB computed values. As captured in Figure 3, the ML model generally perform well, with many predictions lying close to the parity line. Two key observations should be noted about this plot. First, it is inherently difficult for the model to obtain good accuracy on such candidates as these are the cases with extreme SHC values and thus are expected to lie in regions where less amount of training data is available. Second, the candidates were screened using EI acquisition function which also involves uncertainty in model prediction. As can be seen, many candidates (Fe2_2TeSe and BaFe8_8As2_2) lying far from the parity line shows high level of uncertainty in model prediction. Nonetheless, there are also cases (MnC6_6N4_4Se2_2) where the ML model predictions are significantly different than the computed SHC values, highlighting limitations in model generalization.

Predictive capability of ML Models

Figure 3. Predictive capability of ML Models. Parity plot comparing the ML predictions against true (computed) SHC values in modulus symmetric log space. Results are shown for candidates screened in Round 2.2 (yellow squares) and Round 3 (blue triangles). The error bars indicate uncertainty (±1σ\pm1\sigma) in model predictions. The diagonal and vertical dashed line denote the parity line and the SHC value for best candidate (AuSe) from Round 1, respectively. A general trend of discovery of 2D systems with increasing SHC magnitude with each active learning loop can be observed.

More importantly, Figure 3 also captures the progression of candidates with high SHC values identified across different active learning loops. In Round 1, AuSe exhibited the highest absolute SHC of 11.7 ℏ\hbar/e Ω−1\Omega^{-1}. The best performing candidate of Round 2.2, K2_2PtTe2_2, shows an approximately 16-fold increase over the best candidate from Round 1 with SHC of 195.56 ℏ\hbar/e Ω−1\Omega^{-1}. Further, in Round 3, the best performing candidate Fe2_2TeSe has absolute SHC approximately 1.4 times higher (271.52 ℏ\hbar/e Ω−1\Omega^{-1}) than that of K2_2PtTe2_2. Overall, the candidate from Round 3 (i.e., Fe2_2TeSe) shows highest SHC among 41 samples and is 23 times higher than the best of Round 1. This consistent improvement across rounds demonstrates the effectiveness of the active learning strategy in discovering 2D materials with better SHC.

It is to be noted that the primary objective of this work is efficient materials discovery rather than precise property prediction. Thus, we mostly concentrate on validating if the model iteratively discovers candidates with higher SHC than in the training data set in each round. This is conceptually different from a traditional ML setting wherein the accuracy of the model is validated via a hold-out test set. This process is commonly adopted in materials discovery tasks[46, 47] owing to limited training data size and exorbitant cost associated with ground truth generation.

Chemical insights from ML models and Data

To understand the impact of top features on the model predictions, we performed SHAP analysis using the final ML model trained on the full dataset of 41 systems with the results presented in Figure 4(a). A high SHAP value signifies a strong contribution of a given feature to the predicted SHC. Interestingly, both hand-crafted and Magpie features were found to be important, indicating contribution to SHC is dominated by a diverse category of features. Notably, MagpieData range CovalentRadius, defined as the difference between the maximum and minimum elemental covalent radius in a compound, was found to be the topmost feature. This is consistent with the observation that top ranked SHC candidates comprise of elements from the far right and the far left of the periodic table. We observe that majority of them have heavy elements as constituent atoms which lead to high SOC influence on electronic structure leading to high SHC. Similarly, Magpie features such as minimum atomic number and minimum atomic weight among the constituent elements of the compounds are among the top features. If these numbers are high, it is natural that the compound is made by heavy elements and as said before in such cases the SOC is very high. Further, several features (Magpie and hand-crafted) capturing statistics associated with electronic structure, such as mean, mode or range of unfilled ss, pp and dd orbitals were also found to be important. It is consistent with the understanding that partially filled orbitals are involved in strong covalent interactions which form dispersive bandstructures.

Chemical insights from ML model: (a) SHAP summary plot and (b) Pearson correlation coefficients

Figure 4. Chemical insights from ML model (a) SHAP summary plot highlighting the most influential features for the final ML model trained on 41 SHC computations (in symlog space). (b) The Pearson correlation coefficients between features retained after RFE for the final model and Symlog(SHC). The variation in saturation of color (green and orange) indicates variation in intensity of correlation (positive and negative) respectively.

When these bands are SOC-split they produce large spin Berry curvature, which is conducive for SHC. The presence or absence of inversion symmetry in the candidate structure was also found to be an important feature. This is in line with the observation that non-centrosymmetric materials show strong Rashba SOC and often result in enhanced SHC. Overall, the convergence of SHAP-identified important features with well-established physical principles provides a strong validation of the predictive capacity and physical relevance of the trained ML models.

Similar to SHAP analysis, Pearson correlation plots of features (with SHC in symmetric log space) also shows the important role played by electronic structure of the constituting elements and the crystal symmetry of the 2D materials system (see Figure 4(b)). For instance, statistics (mean, mode, range) associated with unfilled dd and pp electrons of the elemental systems showed high magnitude of correlation with the computed SHC. Similarly, presence of different symmetry features such as inversion center (sp_invsp\_inv), different number of rotational axes (sp_n_rotsp\_n\_rot), a half unit translation in screw axis or glide plane (sp_1/2tran_screwsp\_1/2tran\_screw) displayed negative correlation with SHC, which is understandable since with lowering symmetry the bands tend to be flatter leading to weak quantum transport. Also, lower symmetry can enable more non-zero components of the SHC tensor[48]. Further, in line with the results obtained in this work, a recent theoretical study also found that the absence of inversion symmetry could favor large OHC, which ultimately result in high SHC owing to SOC[49]. Other important features that showed large magnitude of correlation with SHC include mean and mode of magnetic moment, mode of electronegativity, and average deviation of ground state volume of constituting elements.

Beyond SHAP and correlation analysis, several other interesting physical trends emerged from the computed SHC data. For instance, all of the top 10 identified 2D systems, except for K8_8ZnTl10_{10} (ranked 8th), contained at least one metalloid (Ge, As, Sb, or Te) or chalcogen (O, S, Se, or Te). Furthermore, with the exception of NiTe (ranked 9th), all of the top 10 ranked systems were ternary or quaternary, indicating that complex bandstructures may be a better playground for achieving higher SHC. It was also observed that 17 out of the 41 candidates studied in this work were metallic (i.e., zero band gap), and, notably, 9 out of top 10 ranked SHC candidates exhibit metallic character. This is in-line with the observation that conventional and highest SHC bulk materials, e.g., Pt, Pd and Au are also metals[5]. Additionally, the metallic character enables rapid variation in properties such as Hall conductivity with even minor shifts in EFE_{\mathrm{F}}, further impacting SHC. This behavior is illustrated in Figure 5, where sharp fluctuations in Hall conductivity near EFE_{\mathrm{F}} can be observed for the best candidates identified from each active learning loop, all of which are metallic.

Variation of Hall conductivity with energy for Fe₂TeSe, K₂PtTe₂, K₈ZnTl₁₀, and AuSe

Figure 5. Variation of Hall conductivity with energy. The plot shows energy-dependent variation of orbital (OHC) and spin (SHC) Hall conductivity for best candidate systems from each round computed using TB model HTB\mathcal{H}_{\mathrm{TB}} and Kubo formalism. The abscissa of point on curves at which zero energy (E-EF_{\mathrm{F}}) horizontal dashed line intersects are taken as OHC and SHC values.

In continuation of previous paragraph, we further elaborate, through Figure 5, the variation of OHC and SHC as a function of change in chemical potential μ\mu which is measured with respect to EFE_{\mathrm{F}}. While OHC and SHC values at EFE_{\mathrm{F}} are sufficient to describe the orbital and spin transport, it is also important to consider that μ\mu is subject to change in real systems based on electron or hole doping, which can arise due to several factors such as interfacial charge transfer, gate voltage tuning, etc. As μ\mu changes, it alters the relative occupations of the bands and, in turn, affects the net orbital and spin magnetic moments in the Brillouin zone. These moments play an important role in controlling the magnitude of the conductivities. We found that some of the metallic systems indicate a sharp change in conductivities near EFE_{\mathrm{F}}, as can be seen in the case of K2_2PtTe2_2 (see Figure 5(b)). Such behavior can be excellent in device fabrication, as it points toward a high tunability of the transport properties with a minimal charge doping. In contrast, there are also compounds such as K8_8ZnTl10_{10} that show a much more stable behavior, which can also be useful in generating a constant spin / orbital current that remains unaffected by doping. Furthermore, comparing the orbital and spin transport, we observe that the OHC always dominates over the SHC throughout all occupations. This is expected, as for the non-magnetic compounds, the SHC can only arise from OHC through the spin orbit coupling. However, one must keep in mind that the experimental detection usually involves a ferromagnetic overlayer, where both the orbital and spin currents are passed. While the spin current can directly exert torque, the orbital current must convert to an effective spin current to produce spin-orbit torque[50]. Although the variation of SHC with the Fermi energy is of the academic interest to examine various possible doping scenarios, in this work, we consider its values at EFE_{\mathrm{F}} (set to zero) for the ML analysis, as this is typically considered as the standard reference value when reporting the SHC of a material.

The dominant presence of dd-orbitals near EFE_{\mathrm{F}} was also found to be an important factor influencing SHC. For example, EBS of 7 out of top 10 candidates, including rank-1 Fe2_2TeSe, is heavily dominated by dd orbitals of transition elements (Fe, Pt, Cu, Mn, Ta, Zn and Ni) near EFE_{\mathrm{F}} (see Figure S3-S4 and Table S1 in Supplementary Information-1 and 2 resp.). Previous studies have shown that transition metal oxides with considerable SOC and dominant dd orbital character near EFE_{\mathrm{F}} exhibit large SHC owing to the SOC enabled mixing of dd orbitals[26]. Evidently, it can be observed that all top-10 ranked candidates except Ba2_2GaBiTe5_5 exhibits the combination of metallicity (zero band gap) and dominance dd-orbital near EFE_{\mathrm{F}}. Another interesting observation we made in the context of presence of heavy elements. Generally, heavy elements (atomic number >> 30) are expected to show high SOC and thus high SHC, but instead we found 7 of the top 10 candidates having dominant dd orbital contribution from non-heavy elements near EFE_{\mathrm{F}} in our dataset. These systems include Fe2_2TeSe (Fe-dd), CuOHF (Cu-dd), BaFe8_8As2_2 (Fe-dd), K2_2MnS2_2 (Mn-dd), TaNi2_2TeSe (Ta-dd and Ni-dd), NiTe (Ni-dd) and MnC6_6N4_4Se2_2 (Mn-dd). Among these, it is noteworthy that CuOHF (rank-3) and K2_2MnS2_2 (rank-5) does not contain any heavy element at all. On the other hand, there were also candidates that contained heavy elements, e.g., AuI (2 eV band gap), In2_2Cl6_6 (3.8 eV), and Na2_2PtO6_6H6_6 (3 eV) but showed negligible SHC. This is attributed the fact that in this compounds the orbital Hall effect is estimated to be very small. As a result, the converted SHE is negligible.

Although the absence of inversion symmetry is generally considered favorable for SHC as it creates an intrinsic electric field, from our top 10 candidates we find that it is possible to create large SHC even without breaking inversion symmetry. Similarly, the presence of multiple mirror planes, typically associated with high SHC in bulk materials[31], did not consistently correlate with enhanced SHC in our dataset. For instance, while the best candidate Fe2_2TeSe exhibits 4 mirror planes, other systems with 4 mirror planes showed low SHC (<∼10ℏ/e Ω−1< \sim10 \hbar/e\ \Omega^{-1}). Importantly, a more prominent trend emerged regarding rotoinversion symmetry (i.e., symmetry operation involving rotation followed by inversion). None of the top 10 candidates exhibited rotoinversion symmetry, with the sole exception of NiTe (ranked 9th). Additionally, materials belonging to non-symmorphic space groups, characterized by fractional translation symmetries such as screw axes and glide planes, generally showed low SHC (<10ℏ/e Ω−1< 10 \hbar/e\ \Omega^{-1}), with K2_2MnS2_2 (rank-5, SHC = 58.26 ℏ/e Ω−1\hbar/e\ \Omega^{-1}) being a notable exception. It is also interesting to note that several non-symmorphic systems (AuI, ZrNI, FeOCl, AgBr, GeSe) contain a heavy element, but still exhibit low SHC. It is generally known that lesser symmetry in multilayer structures favors SHC[51, 52], but we did not observe such trend based on the 41 2D systems we studied. A particularly illustrative case is ZrNI, which was studied in two space groups, a higher symmetry, centrosymmetric orthorhombic P3ˉm1P\bar{3}m1 (rank-18) and low symmetry, non-symmorphic trigonal PmmnPmmn (rank-35). Notably, the high-symmetry phase, featuring additional rotational axes, mirror planes, and rotoinversion symmetry, exhibited a higher SHC.

Contrary to the general expectations of heavy elements (Z>30Z > 30) being crucial for high SHC, we found 2 systems, i.e., CuOHF (rank-3) and K2_2MnS2_2 (rank-5) (among top 5) that do not contain any heavy elements. Similarly, 7 of the top 10 systems contain contributions from 3dd transition elements rather than generally expected 5dd systems. Absence of inversion center or presence of multiple mirror planes is generally considered favorable for high SHC, however, no such trend was observed. Overall, the triple combination of metallic nature, dominant dd character (from heavy or non-heavy elements) near EF_\mathrm{F} and absence of roto-inversion symmetry was found to show high SHC. The 7 systems – Fe2_2TeSe, K2_2PtTe2_2, CuOHF, BaFe8_8As2_2, K2_2MnS2_2, TaNi2_2TeSe, and MnC6_6N4_4Se2_2 – from top 10 exhibited this combination. Thus, our analysis suggests that there are multiple underlying mechanisms involved in determining the SHC and they are distinct for each class of compounds. Such an observation opens up further scope for fundamental research in this domain. Another point to be noted is that all of the identified top 10 candidates fall in the category of ‘easily’ or ‘potentially’ exfoliable, as defined in the MC2D work[39]. Further, we found the rank-1 system Fe2_2TeSe to show dynamic stability (Figure S5 in Supplementary Information-1), and thus top candidates identified in this work are recommended for further detailed computational or experimental investigation.

Phonon band structure for rank-1 candidate Fe2TeSe

Figure S5. Phonon band structure for rank-1 candidate Fe2_2TeSe showing negligible negative frequency of around 0.16 THz indicating the system is dynamically stable.

To conclude, we developed the ML algorithm to scan around 2000 2D materials and predict the ones with high SHC. The trained and validation data for the ML algorithm were obtained from DFT calculations. The designed code, which is available on GitHub, is capable of predicting SHC of any other 2D materials. Our analysis emphasizes that the crystal symmetry and SOC are not the only decisive factor for a material to generate spin current via SHE. In addition, other EBS details such as orbital symmetry (pp or dd) near EF_\mathrm{F}, types of atomic species (metal or non-metal) and composition (binary or multinary), covalent radius or electronegativity of these atomic species, etc., can dictate the phenomenon. It should also be noted that the final ML model uses all of the above properties to make an accurate prediction, and no single dominant property alone can be used to arrive at a good estimate of SHC in 2D systems. Further, the active learning strategy employed in this work successfully identified high-performing 2D systems with manifold increase in SHC as compared to initial training data. We anticipate that present work shall help both the spintronic and the broader material informatics community. The set of high SHC candidates identified in this work can be further studied. At the same time, this work presents a successful example of ML guided accelerated discovery wherein the ML models are able to successfully find superior systems with desired properties from a large candidate pool.

Methods

Computational details

DFT and TB workflow. The DFT computations were performed using Quantum ESPRESSO[53], and FINDSYM[54] was used for symmetry analysis. The exchange and correlation behavior of electron-electron interactions were approximated under the generalized gradient approximation in the Perdew-Burke-Ernzerhof (GGA-PBE)[55] method. Whereas for nucleus-electron interactions, the norm-conserving pseudopotentials[56] (from PseudoDojo library[57]) were used owing to their reputation of exact eigenvalues and node-less eigenfunctions to better compute the Berry curvature. The 2D crystal structures, imported from the Materials Cloud Two-Dimensional Structure Database (MC2D)[39, 40], were modeled using a slab approach with sufficient vacuum along zz direction (>12> 12 Å) to avoid periodic interactions. Although MC2D database already contains relaxed structures, in the initial round a few structures were re-relaxed. Since only negligible changes were observed, the re-relaxation step was omitted in the subsequent rounds. A sufficiently high and plane-wave basis cut-off for wavefunction expansion (in similar range of PseudoDojo library recommended values) and charge-density was used along with relevant total energy convergence cut-off.

(i) We first perform DFT computations to obtain EBS without SOC. For this, we performed self-consistent field (SCF) calculations using an automatic kk-mesh with varied kk-spacing in a range of 0.03-0.17 Å−1^{-1}. This is followed by non-SCF calculations to analyze partial density of states (Figure S3 in Supplementary Information-1 and Table S1 in Supplementary Information-2) to choose appropriate projections and compute wavefunctions on full uniform kk-mesh with varied kk-spacing 0.02-0.10 Å−1^{-1} as inputs to construction of MLWFs using Wannier90[42] package. MLWFs are constructed in the absence of SOC to avoid a mix-up of wavefunctions and the error in the eigenstates. High-quality MLWFs were ensured by achieving spread parameters up to 10−1210^{-12} Å2^2. After post-processing accurately fit MLWFs (Figure S4 in Supplementary Information-1 and Table S1 S1 in Supplementary Information-2), the hopping parameters were obtained for step-iii.

Projected density of states (PDOS) for Fe₂TeSe, K₂PtTe₂, CuOHF, BaFe₈As₂, K₂MnS₂, and Ba₂GaBiTe₅

Figure S3. Projected density of states (PDOS) for top 6 candidates showing contribution of their constituent atomic orbitals providing information on appropriate projectors for Wannierization of bands.

MLWF fitted DFT electronic band structures for Fe₂TeSe, K₂PtTe₂, CuOHF, BaFe₈As₂, K₂MnS₂, and Ba₂GaBiTe₅

Figure S4. MLWF fitted DFT electronic band structures (EBS). The ab initio EBS along high symmetry kk-path for some top candidates from round 3 (blue) and round 2.2 (orange) fitted with strongly converged MLWFs without spin-orbit coupling (SOC). The SOC corrections are later added to the TB part.

MaterialSOC parameters (eV)
AuI\mathrm{AuI}λAu=0.584\lambda_{Au}=0.584, λI=0.622\lambda_{I}=0.622
ZrS2\mathrm{ZrS}_{2}λZr=0.058\lambda_{Zr}=0.058, λS=0.0795\lambda_{S}=0.0795
LiO\mathrm{LiO}λLi=0\lambda_{Li}=0, λO=0.0351\lambda_{O}=0.0351
In2Cl6\mathrm{In}_{2}\mathrm{Cl}_{6}λIn=0\lambda_{In}=0, λCl=0.112\lambda_{Cl}=0.112
ZrCl2\mathrm{ZrCl}_{2}λZr=0.058\lambda_{Zr}=0.058, λCl=0\lambda_{Cl}=0
As2S3\mathrm{As}_{2}\mathrm{S}_{3}λAs=0.241\lambda_{As}=0.241, λS=0.0712\lambda_{S}=0.0712
ZrNI(Pmmn)\mathrm{ZrNI(Pmmn)}λZr=0.055\lambda_{Zr}=0.055, λI=0.814\lambda_{I}=0.814, λN=0.0181\lambda_{N}=0.0181
Na2PtO6H6\mathrm{Na}_{2}\mathrm{PtO}_{6}\mathrm{H}_{6}λPt=0.568\lambda_{Pt}=0.568, λO=0.0357\lambda_{O}=0.0357, λH=0\lambda_{H}=0, λNa=0\lambda_{Na}=0
CdI2\mathrm{CdI}_{2}λCd=0.312\lambda_{Cd}=0.312, λI=0.829\lambda_{I}=0.829
FeOCl\mathrm{FeOCl}λFe=0.0691\lambda_{Fe}=0.0691, λO=0.0356\lambda_{O}=0.0356, λCl=0.121\lambda_{Cl}=0.121
PbS2\mathrm{PbS}_{2}λPb=0\lambda_{Pb}=0, λS=0.0795\lambda_{S}=0.0795
BiTe2\mathrm{BiTe}_{2}λBi=1.65\lambda_{Bi}=1.65, λTe=0.74\lambda_{Te}=0.74
ZrTe2\mathrm{ZrTe}_{2}λZr=0.059\lambda_{Zr}=0.059, λTe=0.71\lambda_{Te}=0.71
SbAs\mathrm{SbAs}λSb=0.545\lambda_{Sb}=0.545, λAs=0.246\lambda_{As}=0.246
Sb2Te2\mathrm{Sb}_{2}\mathrm{Te}_{2}λSb=0.529\lambda_{Sb}=0.529, λTe=0.707\lambda_{Te}=0.707
MoS2\mathrm{MoS}_{2}λMo=0.0712\lambda_{Mo}=0.0712, λS=0.0715\lambda_{S}=0.0715
AgN3\mathrm{AgN}_{3}λN=0.0196\lambda_{N}=0.0196, λAg=0.271\lambda_{Ag}=0.271
SbI3\mathrm{SbI}_{3}λSb=0.536\lambda_{Sb}=0.536, λI=0.891\lambda_{I}=0.891
Sb2S3\mathrm{Sb}_{2}\mathrm{S}_{3}λSb=0.541\lambda_{Sb}=0.541, λS=0.0771\lambda_{S}=0.0771
Li2CuAs\mathrm{Li}_{2}\mathrm{CuAs}λCu=0.122\lambda_{Cu}=0.122, λS=0.0797\lambda_{S}=0.0797, λLi=0\lambda_{Li}=0
MoSe2\mathrm{MoSe}_{2}λMo=0.0718\lambda_{Mo}=0.0718, λSe=0.346\lambda_{Se}=0.346
PtS2\mathrm{PtS}_{2}λPt=0.581\lambda_{Pt}=0.581, λS=0.0788\lambda_{S}=0.0788
Li2PtH2\mathrm{Li}_{2}\mathrm{PtH}_{2}λPt=0.547\lambda_{Pt}=0.547, λLi=0\lambda_{Li}=0, λH=0\lambda_{H}=0
ZrNI\mathrm{ZrNI}λZr=0.052\lambda_{Zr}=0.052, λI=0.887\lambda_{I}=0.887, λN=0.0152\lambda_{N}=0.0152
MoTe2\mathrm{MoTe}_{2}λMo=0.0912\lambda_{Mo}=0.0912, λTe=0.77\lambda_{Te}=0.77
AgBr\mathrm{AgBr}λAg=0.242\lambda_{Ag}=0.242, λBr=0.414\lambda_{Br}=0.414
GeSe\mathrm{GeSe}λGe=0\lambda_{Ge}=0, λSe=0\lambda_{Se}=0
PtSe2\mathrm{PtSe}_{2}λPt=0.568\lambda_{Pt}=0.568, λSe=0.351\lambda_{Se}=0.351
PbTe\mathrm{PbTe}λPb=1.25\lambda_{Pb}=1.25, λTe=0.69\lambda_{Te}=0.69
WSe2\mathrm{WSe}_{2}λW=0.228\lambda_{W}=0.228, λSe=0.364\lambda_{Se}=0.364
AuSe\mathrm{AuSe}λAu=0.611\lambda_{Au}=0.611, λSe=0.328\lambda_{Se}=0.328
MnC6N4Se2\mathrm{MnC}_{6}\mathrm{N}_{4}\mathrm{Se}_{2}λMn=0.0538\lambda_{Mn}=0.0538, λC=0.00844\lambda_{C}=0.00844, λSe=0.312\lambda_{Se}=0.312, λN=0.0163\lambda_{N}=0.0163
NiTe\mathrm{NiTe}λNi=0.084\lambda_{Ni}=0.084, λTe=0.692\lambda_{Te}=0.692
K8ZnTl10\mathrm{K}_{8}\mathrm{ZnTl}_{10}λZn=0.133\lambda_{Zn}=0.133, λTl=0.856\lambda_{Tl}=0.856, λK=0\lambda_{K}=0
TaNi2TeSe\mathrm{TaNi}_{2}\mathrm{TeSe}λTa=0.239\lambda_{Ta}=0.239, λNi=0.099\lambda_{Ni}=0.099, λTe=0.679\lambda_{Te}=0.679, λSe=0.357\lambda_{Se}=0.357
Ba2GaBiTe5\mathrm{Ba}_{2}\mathrm{GaBiTe}_{5}λTe=0.712\lambda_{Te}=0.712, λBi=1.62\lambda_{Bi}=1.62, λBa=0\lambda_{Ba}=0, λGa=0\lambda_{Ga}=0
K2MnS2\mathrm{K}_{2}\mathrm{MnS}_{2}λMn=0.0546\lambda_{Mn}=0.0546, λS=0.0671\lambda_{S}=0.0671, λK=0\lambda_{K}=0
BaFe8As2\mathrm{BaFe}_{8}\mathrm{As}_{2}λFe=0.0697\lambda_{Fe}=0.0697, λAs=0.233\lambda_{As}=0.233, λBa=0\lambda_{Ba}=0
CuOHF\mathrm{CuOHF}λCu=0.121\lambda_{Cu}=0.121, λF=0.059\lambda_{F}=0.059, λO=0.0357\lambda_{O}=0.0357, λH=0\lambda_{H}=0
K2PtTe2\mathrm{K}_{2}\mathrm{PtTe}_{2}λPt=0.568\lambda_{Pt}=0.568, λTe=0.706\lambda_{Te}=0.706, λK=0\lambda_{K}=0
Fe2TeSe\mathrm{Fe}_{2}\mathrm{TeSe}λFe=0.0688\lambda_{Fe}=0.0688, λSe=0.354\lambda_{Se}=0.354, λTe=0.719\lambda_{Te}=0.719

Table S1. Spin-orbit coupling (SOC) parameters λ\lambda obtained by fitting the TB bands with DFT(SOC). Note that the parameters are listed corresponding to the orbital projections used for the given materials. One may notice that some of the lighter elements have a larger SOC than some heavier elements, because of the fact that p-orbitals usually have larger SOC as compared to the d-orbitals.

(ii) Separate EBS calculations including SOC effects are performed to estimate SOC strengths as extra inputs to step-iii.

(iii) Kubo formalism for orbital and spin transport: In this sub-section, we discuss the computation of the orbital and spin transport properties. First, a numerical TB model was constructed based on the hopping parameters obtained from step-i. Keeping the accuracy of the eigenstates in mind, the spin orbit coupling was initially turned off. Later, SOC from step-ii was introduced as an additional term to the Hamiltonian, and the SOC parameters were tuned to match DFT results. We introduced the SOC as an extra parameter to the Hamiltonian and later the band structure was fit with the DFT EBS with SOC (Figure S6 in Supplementary Information-1), by tuning the atomistic SOC parameter λ\lambda. This fitting validates not just the accuracy of the eigenvalues (bands), but also the eigenstates. This method is commonly used in the literature in the context of theoretical tight-binding models[9, 58, 59]. Table S1 of Supplementary Information-1 lists λ\lambda obtained for all the materials. The matrix elements of the Hamiltonian are given by,

Comparison between bands of DFT+SOC and TB (with SOC) for AuSe

Figure S6. Comparison between bands of DFT+SOC and TB (with SOC) for AuSe. It shows a perfect match indicating correctness of both eigenvalues and eigenstates. The corresponding value of SOC parameters obtained by fitting are λAu,d=0.61\lambda_{\mathrm{Au},d}=0.61 eV and λSe,p=0.33\lambda_{\mathrm{Se},p}=0.33 eV.

Hmn=∑jtmnjeik⃗⋅d⃗j+λL⃗⋅S⃗,(1)H_{mn}=\sum_j t_{mn}^{j}e^{i\vec{k}\cdot\vec{d}_j}+\lambda\vec{L}\cdot\vec{S}, \tag*{(1)}

where mm and nn are the basis indices, jj runs over all the neighbors and tmnjt_{mn}^{j} corresponds to the hopping parameter. In the second term, the parameter λ\lambda represents the SOC strength. The eigenvalues and eigenstates of this Hamiltonian were further used to calculate the orbital (OHC) and spin Hall conductivity (SHC) on a dense kk-grid of 800×800×1800\times800\times1. The Kubo formula[60] for the conductivities can be expressed as,

σαβγ,orb/spin=−eNkVc∑k⃗Ωαβγ,orb/spin(k⃗),(2)\sigma_{\alpha\beta}^{\gamma,\mathrm{orb/spin}}=-\frac{e}{N_kV_c}\sum_{\vec{k}}\Omega_{\alpha\beta}^{\gamma,\mathrm{orb/spin}}(\vec{k}), \tag*{(2)}

where Ωαβγ,orb/spin(k⃗)=∑noccΩn,αβγ,orb/spin(k⃗)\Omega_{\alpha\beta}^{\gamma,\mathrm{orb/spin}}(\vec{k})=\sum_n^{\mathrm{occ}}\Omega_{n,\alpha\beta}^{\gamma,\mathrm{orb/spin}}(\vec{k}) is the sum over occupied bands nn at a specific momentum point k⃗\vec{k}. Here, α\alpha, β\beta, and γ\gamma can be xx, yy or zz such that α\alpha is the direction of the orbital/spin current, β\beta is the direction of the applied electric field, while the orbital and spin angular momentum points towards γ\gamma, and VcV_c corresponds to the volume of the unit cell. Importantly, for 2D systems the volume term is replaced by the area of the unit cell. Therefore, the unit of SHC for 2D systems is usually expressed in ℏ/e Ω−1\hbar/e\ \Omega^{-1}[61]. The orbital Berry curvature for the Bloch state nk⃗n\vec{k} is computed using the Kubo formula,

Ωn,αβγ,orb(k⃗)=2ℏ∑n′≠nIm⁡[⟨unk⃗∣Jαγ,orb∣un′k⃗⟩⟨un′k⃗∣vβ∣unk⃗⟩](εn′k⃗−εnk⃗)2,(3)\Omega_{n,\alpha\beta}^{\gamma,\mathrm{orb}}(\vec{k})=2\hbar\sum_{n'\ne n}\frac{\operatorname{Im}\left[\langle u_{n\vec{k}}|\mathcal{J}_{\alpha}^{\gamma,\mathrm{orb}}|u_{n'\vec{k}}\rangle\langle u_{n'\vec{k}}|v_\beta|u_{n\vec{k}}\rangle\right]}{(\varepsilon_{n'\vec{k}}-\varepsilon_{n\vec{k}})^2}, \tag*{(3)}

where vα=1ℏ∂H∂kαv_\alpha=\frac{1}{\hbar}\frac{\partial\mathcal{H}}{\partial k_\alpha} is the velocity operator as defined earlier, and the orbital current operator is Jαγ,orb=12{vα,Lγ}\mathcal{J}_{\alpha}^{\gamma,\mathrm{orb}}=\frac{1}{2}\{v_\alpha,L_\gamma\}, with LγL_\gamma being the orbital angular momentum. For the spin Berry curvature, the orbital current operator is replaced by the spin current operator Jαγ,spin=12{vα,Sγ}\mathcal{J}_{\alpha}^{\gamma,\mathrm{spin}}=\frac{1}{2}\{v_\alpha,S_\gamma\} in Eq. (3). In the present case, for the 2D systems, Jyxγ\mathcal{J}_{yx}^{\gamma} is the only conventional component of the OHC and SHC tensor that survives, which is used throughout this work. It is also important to note that both the orbital and spin Hall effects can produce spin orbit torque in the experiments, which is measured by a non-vanishing spin Hall angle. To validate our computational approach, we compared the calculated SHC values of 2D TMDs with a prior work[59]. As summarized in Table S2 of Supplementary Information-1, our results show excellent agreement with the literature.

SystemsSHC (*)SHC (This work)
MoS2_21.01.0
MoSe2_21.81.8
MoTe2_23.03.0
WSe2_27.18.5

Table S2. A brief comparison between the SHC for few TMD systems from the literature and the current work, which shows an excellent agreement. All are in the units of ℏ/e Ω−1\hbar/e\ \Omega^{-1}.

Machine learning workflow. An initial dataset comprising 24 randomly selected 2D materials was extracted from the MC2D database, a widely utilized resource in computational materials science that includes approximately 2,000 2D material compositions along with their associated crystallographic and electronic structure data. Prior to ML model development, the dataset was systematically curated by excluding entries that lacked band gap information, consist only of a single element, or included elements for which norm-conserving pseudopotentials were unavailable. This pre-processing yielded a final, cleaned candidate pool of 1,947 2D material systems.

Two categories of input features were employed to train the ML models. The first consisted of Magpie features, generated using the matminer library, which encode elemental properties such as electronegativity, covalent radius, atomic volume, and valence electron counts across the ss, pp, and dd orbitals, along with their statistical descriptors (e.g., mean, mode, range). The second category comprised of hand-crafted features designed to capture crystallographic symmetry of the 2D material system, including the presence of rotational, inversion, or roto-inversion symmetries. In total, 158 features, 26 hand-crafted (Table S3 in Supplementary Information-1) and 132 Magpie-derived, were used in the initial phase of model training.

AcronymDescription
SG numberSpace group number
sp_db_egMC2D database bandgap
sp_rotRotational symmetry present?- Yes (1) or No (0)
sp_n_rotHighest order of rotational symmetry
sp_invInversion symmetry present? - Yes (1) or No (0)
sp_rotoinvRotoinversion symmetry present? - Yes (1) or No (0)
sp_n_rotoinvHigher order of axis of rotoinversion symmetry
sp_mirMirror symmetry present? - Yes (1) or No (0)
sp_n_mirrorNumber of mirror planes
sp_screwScrew axis symmetry present? - Yes (1) or No (0)
sp_n_rot_screwHighest order of rotation in screw symmetry
sp_1/2tran_screwHalf translation in screw symmetry present? - Yes (1) or No (0)
sp_1/4tran_screwQuarter translation in screw symmetry present? - Yes (1) or No (0)
sp_glideGlide plane symmetry present? - Yes (1) or No (0)
sp_1/2tran_glideHalf translation in glide symmetry present? - Yes (1) or No (0)
sp_1/4tran_glideQuarter translation in glide symmetry present? - Yes (1) or No (0)
sp_val_metal_per_atomNumber of valence electrons in PBE norm-conserving pseudopotential (NCPP) for metal divided by total atoms present in formula unit
sp_val_s_metal_per_atomNumber of ss-orbital valence electrons in PBE NCPP for metal divided by total atoms present in formula unit
sp_val_p_metal_per_atomNumber of pp-orbital valence electrons in PBE NCPP for metal divided by total atoms present in formula unit
sp_val_d_metal_per_atomNumber of dd-orbital valence electrons in PBE NCPP for metal divided by total atoms present in formula unit
sp_val_f_metal_per_atomNumber of ff-orbital valence electrons in PBE NCPP for metal divided by total atoms present in formula unit
sp_val_anion_per_atomNumber of valence electrons in PBE NCPP for anions divided by total atoms present in formula unit
sp_val_s_anion_per_atomNumber of ss-orbital valence electrons in PBE NCPP for anions divided by total atoms present in formula unit
sp_val_p_anion_per_atomNumber of pp-orbital valence electrons in PBE NCPP for anions divided by total atoms present in formula unit
sp_val_d_anion_per_atomNumber of dd-orbital valence electrons in PBE NCPP for anions divided by total atoms present in formula unit
sp_val_f_anion_per_atomNumber of ff-orbital valence electrons in PBE NCPP for anions divided by total atoms present in formula unit

Table S3. The 26 hand-crafted features used for ML model development. The remaining list of Magpie features can be found online at https://hackingmaterials.lbl.gov/matminer/.

To ensure generalizability with a limited dataset size, relatively simple ML models, including ridge regression, Gaussian process regression (GPR), and support vector regression (SVR), were employed for model development. Given that the SHC (and THC) values span several orders of magnitude, a symmetric logarithmic (Symlog) transformation[62] given by the equation y=sign⁡(x)⋅ln⁡(1+∣x/C∣)y=\operatorname{sign}(x)\cdot\ln(1+|x/C|) was applied (with C=1C=1) to the target property prior to training. To reduce feature dimensionality and improve model generalization, a recursive feature elimination (RFE) strategy was used. Features resulting in highest 5-fold cross-validation (CV) error were recursively removed until the CV error began to increase. The feature set (Table S4 in Supplementary Information-1) corresponding to the minimum CV error was retained for subsequent model training. Hyperparameter optimization was performed individually for each model type to further enhance performance. Model evaluation was conducted using 5-fold CV, with average root mean squared error (RMSE) and average coefficient of determination (R2R^{2}) serving as performance metrics. Among the tested models, ridge regression with feature normalization and RFE consistently yielded best performance across all active learning loops. The CV R2R^{2} score (and RMSE in symlog units) across active learning rounds 1, 2.2, 3 and 4 was 0.94 (1.58), 0.98 (0.11), 0.94 (0.39) and 0.94 (0.46), respectively, demonstrating the consistently high predictive accuracy of the trained models. The parity plots comparing the performance of all ML models across different active learning loops are included in Figure S2 in Supplementary Information-1.

This optimized ridge regression model was then used to predict SHC for materials remaining in the candidate pool, excluding those cases which are part of the training dataset. To estimate prediction uncertainty, an ensemble of 100 models was generated by randomly sampling 80% of the training data for each variant. The expected improvement (EI) acquisition function, which balances exploration and exploitation, was employed to select the most promising candidates for subsequent DFT–SOC + TB modeling. Besides EI, diversity of the candidates in terms of type (light or heavy, alkali, alkaline earth or transition metals, metalloid, chalcogens or halogens) or number (binary, ternary or quaternary) of elements was also considered to screen next round of candidates.

DATA AVAILABILITY

The SHC data generated in this work can be accessed at the GitHub Repository. The visualisation of training data using t-SNE map, list of features (before and after RFE), comparative performance of different ML models across active learning rounds, SOC parameters for all 41 systems, and a representative example of DFT+SOC band fitting with TB (including SOC), and comparison of calculated SHC values of few TMDs with literature is given in Supplementary Information-1. The MLWF fitting of ab-initio EBS for top 6 ranked systems and their PDOS are also provided in Supplementary Information-1, whereas those for rank-7 to rank-41 materials are provided in Supplementary Information-2.

CODE AVAILABILITY

The trained ML model, along with the complete dataset of this work, is publicly available at GitHub Repository.

Acknowledgements

Author A.J.K. gratefully acknowledges financial support previously received from the Center for Atomistic Modelling and Materials Design (CAMMD), IIT Madras and currently from the Anusandhan National Research Foundation (ANRF / erstwhile SERB) National Post-Doctoral Fellowship (NPDF) scheme (File No. PDF/2023/000036). We gratefully acknowledge the computing resources provided on Improv, Bebop and Swing, a high-performance computing cluster operated by the Laboratory Computing Resource Center (LCRC) at Argonne National Laboratory (ANL), USA. Authors A.J.K. and R.B. acknowledge the part of this work is performed at Carbon cluster (proposal #83855) at Center for Nanoscale Materials, a U.S. Department of Energy Office of Science User Facility, was supported by the U.S. DOE, Office of Basic Energy Sciences, under Contract No. DE-AC02-06CH11357. R.B. acknowledges ANRF for the funding (File No. SRG/2023/001055). Computational resources from Robert Bosch Centre for Data Science and AI, IIT Madras, are also acknowledged.

Author contributions statement

A.J.K.: DFT Computation and MLWF Construction, ML Data Curation, Visualization, Analysis, Writing. S.N.: ML Investigation, ML Data Curation, Visualization, Analysis, Writing. P.S.: TB Investigation and SHC Computation, Validation, Analysis, Writing. H.C.: ML Investigation. B.R.K.N.: Conceptualization, Resources, Funding Acquisition, Supervision, Validation. R.B.: Conceptualization, Resources, Funding Acquisition, Supervision, Visualization, Validation. All authors reviewed and edited the manuscript.

Supplementary_Information-1.pdf

  • Bhowal, Sayantika, and Sashi Satpathy. “Intrinsic orbital and spin Hall effects in monolayer transition metal dichalcogenides.” Physical Review B 102.3 (2020): 035409.

Round 2.2Round 3Round 4
Common in all three rounds
1sp_1/2tran_screwsp_1/2tran_screwsp_1/2tran_screw
2sp_1/2tran_glidesp_1/2tran_glidesp_1/2tran_glide
3MagpieData maximum NValenceMagpieData maximum NValenceMagpieData maximum NValence
Common in round 2.2 and 3
4MagpieData avg_dev AtomicWeightMagpieData avg_dev AtomicWeight
5MagpieData mean MeltingTMagpieData mean MeltingT
6MagpieData minimum NValenceMagpieData minimum NValence
7MagpieData mean NdUnfilledMagpieData mean NdUnfilled
Common in round 2.2 and 4
8MagpieData minimum MeltingT- - - - - - - - - - - - - -MagpieData minimum MeltingT
9MagpieData mean CovalentRadius- - - - - - - - - - - - - -MagpieData mean CovalentRadius
10MagpieData avg_dev NpValence- - - - - - - - - - - - - -MagpieData avg_dev NpValence
11MagpieData maximum GSvolume-pa- - - - - - - - - - - - - -MagpieData maximum GSvolume-pa
Common in round 3 and 4
12- - - - - - - - - - - - - -sp_val_p_metal_per_atomsp_val_p_metal_per_atom
13- - - - - - - - - - - - - -sp_val_d_anion_per_atomsp_val_d_anion_per_atom
14- - - - - - - - - - - - - -MagpieData minimum NumberMagpieData minimum Number
15- - - - - - - - - - - - - -MagpieData mean NumberMagpieData mean Number
16- - - - - - - - - - - - - -MagpieData range CovalentRadiusMagpieData range CovalentRadius
17- - - - - - - - - - - - - -MagpieData minimum ElectronegativityMagpieData minimum Electronegativity
18- - - - - - - - - - - - - -MagpieData mode ElectronegativityMagpieData mode Electronegativity
19- - - - - - - - - - - - - -MagpieData range NdValenceMagpieData range NdValence
20- - - - - - - - - - - - - -MagpieData mode NpUnfilledMagpieData mode NpUnfilled
21- - - - - - - - - - - - - -MagpieData mode NdUnfilledMagpieData mode NdUnfilled
22- - - - - - - - - - - - - -MagpieData mean GSmagmomMagpieData mean GSmagmom
23- - - - - - - - - - - - - -MagpieData minimum Space-GroupNumberMagpieData minimum Space-GroupNumber
24- - - - - - - - - - - - - -MagpieData maximum Space-GroupNumberMagpieData maximum Space-GroupNumber
25- - - - - - - - - - - - - -sp_invsp_inv

Table S4. Table S4: List of features retained after recursive feature elimination (RFE) across active ML rounds.

Round 2.2Round 3Round 4
Not Common in any rounds
26sp_rotoinvsp_val_metal_per_atomsp_n_rot
27sp_n_mirrorsp_val_d_metal_per_atomMagpieData range Number
28MagpieData mode NumberMagpieData mode MeltingTMagpieData avg-dev Number
29MagpieData mode AtomicWeightMagpieData range ColumnMagpieData avg-dev Number
30MagpieData maximum CovalentRadiusMagpieData minimum NsValenceMagpieData maximum Column
31MagpieData mode CovalentRadiusMagpieData maximum NsValenceMagpieData avg-dev Row
32MagpieData mean NsValenceMagpieData avg-dev NdValenceMagpieData mode NpValence
33MagpieData mean NdValenceMagpieData maximum NsUnfilledMagpieData minimum NdValence
34MagpieData mode NdValenceMagpieData minimum NdUnfilledMagpieData range NsUnfilled
35MagpieData range NValenceMagpieData minimum NfUnfilledMagpieData mode NsUnfilled
36MagpieData mode NValenceMagpieData maximum NfUnfilledMagpieData avg-dev NpUnfilled
37MagpieData mean NpUnfilledMagpieData range NfUnfilledMagpieData range NdUnfilled
38MagpieData avg-dev Space-GroupNumberMagpieData mean NfUnfilledMagpieData avg-dev NdUnfilled
39- - - - - - - - - - - - - -MagpieData avg-dev NfUnfilledMagpieData mode NUnfilled
40- - - - - - - - - - - - - -MagpieData mode NfUnfilledMagpieData avg-dev GSvolume-pa
41- - - - - - - - - - - - - -MagpieData range GSvolume_paMagpieData maximum GS-bandgap
42- - - - - - - - - - - - - -MagpieData range GSbandgapMagpieData mode GSbandgap
43- - - - - - - - - - - - - -MagpieData minimum GSmag-momMagpieData mode GSmagmom
44- - - - - - - - - - - - - -MagpieData maximum GSmag-momMagpieData mean SpaceGroup-Number
45- - - - - - - - - - - - - -MagpieData range GSmagmom- - - - - - - - - - - - - -
46- - - - - - - - - - - - - -sp_1/4tran_screw- - - - - - - - - - - - - -
47- - - - - - - - - - - - - -sp_1/4tran_glide- - - - - - - - - - - - - -

Table 1.

References

  1. [1]Sinova, J., Valenzuela, S. O., Wunderlich, J., Back, C. H. & Jungwirth, T. Spin Hall effects. Rev. Mod. Phys. 87, 1213–1260, DOI: 10.1103/RevModPhys.87.1213 (2015).DOI
  2. [2]Li, X. & Yang, J. First-principles design of spintronics materials. Natl. Sci. Rev. 3, 365–381, DOI: 10.1093/nsr/nww026 (2016). _eprint: https://academic.oup.com/nsr/article-pdf/3/3/365/31566317/nww026.pdf.DOI
  3. [3]Ren, H. et al. Hybrid spin Hall nano-oscillators based on ferromagnetic metal/ferrimagnetic insulator heterostructures. Nat. Commun. 14, 1406, DOI: 10.1038/s41467-023-37028-4 (2023).DOI
  4. [4]Kimura, T., Otani, Y., Sato, T., Takahashi, S. & Maekawa, S. Room-Temperature Reversible Spin Hall Effect. Phys. Rev. Lett. 98, 156601, DOI: 10.1103/PhysRevLett.98.156601 (2007).DOI
  5. [5]Guo, G. Y. Ab initio calculation of intrinsic spin Hall conductivity of Pd and Au. J. Appl. Phys. 105, 07C701, DOI: 10.1063/1.3054362 (2009).DOI
  6. [6]Pai, C.-F. et al. Spin transfer torque devices utilizing the giant spin Hall effect of tungsten. Appl. Phys. Lett. 101, 122404, DOI: 10.1063/1.4753947 (2012).DOI
  7. [7]Chernyshov, A. et al. Evidence for reversible control of magnetization in a ferromagnetic material by means of spin–orbit magnetic field. Nat. Phys. 5, 656–659, DOI: 10.1038/nphys1362 (2009).DOI
  8. [8]Xu, H. et al. High Spin Hall Conductivity in Large-Area Type-II Dirac Semimetal PtTe2. Adv. Mater. 32, 2000513, DOI: https://doi.org/10.1002/adma.202000513 (2020). _eprint: https://advanced.onlinelibrary.wiley.com/doi/pdf/10.1002/adma.202000513.
  9. [9]Sahu, P., Bidika, J. K., Biswal, B., Satpathy, S. & Nanda, B. R. K. Emergence of giant orbital Hall and tunable spin Hall effects in centrosymmetric transition metal dichalcogenides. Phys. Rev. B 110, 054403, DOI: 10.1103/PhysRevB.110.054403 (2024).DOI
  10. [10]Ahn, E. C. 2D materials for spintronic devices. npj 2D Mater. Appl. 4, 17, DOI: 10.1038/s41699-020-0152-0 (2020).DOI
  11. [11]Wunderlich, J., Kaestner, B., Sinova, J. & Jungwirth, T. Experimental Observation of the Spin-Hall Effect in a Two-Dimensional Spin-Orbit Coupled Semiconductor System. Phys. Rev. Lett. 94, 047204, DOI: 10.1103/PhysRevLett.94.047204 (2005).DOI
  12. [12]Song, P. et al. Coexistence of large conventional and planar spin Hall effect with long spin diffusion length in a low-symmetry semimetal at room temperature. Nat. Mater. 19, 292–298, DOI: 10.1038/s41563-019-0600-4 (2020).DOI
  13. [13]Han, W., Kawakami, R. K., Gmitra, M. & Fabian, J. Graphene spintronics. Nat. Nanotechnol. 9, 794–807, DOI: 10.1038/nnano.2014.214 (2014).DOI
  14. [14]Hong, J. et al. Room-temperature Magnetic Ordering in Functionalized Graphene. Sci. Reports 2, 624, DOI: 10.1038/srep00624 (2012).DOI
  15. [15]Nair, R. et al. Dual origin of defect magnetism in graphene and its reversible switching by molecular doping. Nat. Commun. 4, 2010, DOI: 10.1038/ncomms3010 (2013).DOI
  16. [16]Yazyev, O. V. & Helm, L. Defect-induced magnetism in graphene. Phys. Rev. B 75, 125408, DOI: 10.1103/PhysRevB.75.125408 (2007).DOI
  17. [17]Červenka, J., Katsnelson, M. I. & Flipse, C. F. J. Room-temperature ferromagnetism in graphite driven by two-dimensional networks of point defects. Nat. Phys. 5, 840–844, DOI: 10.1038/nphys1399 (2009).DOI
  18. [18]Nanda, B. R. K., Sherafati, M., Popović, Z. S. & Satpathy, S. Electronic structure of the substitutional vacancy in graphene: density-functional and Green’s function studies. New J. Phys. 14, 083004, DOI: 10.1088/1367-2630/14/8/083004 (2012).DOI
  19. [19]Padmanabhan, H. & Nanda, B. R. K. Intertwined lattice deformation and magnetism in monovacancy graphene. Phys. Rev. B 93, 165403, DOI: 10.1103/PhysRevB.93.165403 (2016).DOI
  20. [20]Haugen, H., Huertas-Hernando, D. & Brataas, A. Spin transport in proximity-induced ferromagnetic graphene. Phys. Rev. B 77, 115406, DOI: 10.1103/PhysRevB.77.115406 (2008).DOI
  21. [21]Yang, H. X. et al. Proximity Effects Induced in Graphene by Magnetic Insulators: First-Principles Calculations on Spin Filtering and Exchange-Splitting Gaps. Phys. Rev. Lett. 110, 046603, DOI: 10.1103/PhysRevLett.110.046603 (2013).DOI
  22. [22]Shao, Q. et al. Strong Rashba-Edelstein Effect-Induced Spin–Orbit Torques in Monolayer Transition Metal Dichalcogenide/Ferromagnet Bilayers. Nano Lett. 16, 7514–7520, DOI: 10.1021/acs.nanolett.6b03300 (2016).DOI
  23. [23]Mudgal, R. et al. Magnetic-Proximity-Induced Efficient Charge-to-Spin Conversion in Large-Area PtSe2/Ni80Fe20 Heterostructures. Nano Lett. 23, 11925–11931, DOI: 10.1021/acs.nanolett.3c04060 (2023).DOI
  24. [24]Dai, Y. et al. Interfacial magnetic spin Hall effect in van der Waals Fe3GeTe2/MoTe2 heterostructure. Nat. Commun. 15, 1129, DOI: 10.1038/s41467-024-45318-8 (2024).DOI
  25. [25]Meitei, O. R. & Van Voorhis, T. Electron Correlation in 2D Periodic Systems from Periodic Bootstrap Embedding. The J. Phys. Chem. Lett. 15, 11992–12000 (2024).DOI
  26. [26]Jadaun, P., Register, L. F. & Banerjee, S. K. Rational design principles for giant spin Hall effect in 5d-transition metal oxides. Proc. Natl. Acad. Sci. 117, 11878–11886 (2020).DOI
  27. [27]Soumyanarayanan, A., Reyren, N., Fert, A. & Panagopoulos, C. Emergent phenomena induced by spin–orbit coupling at surfaces and interfaces. Nature 539, 509–517, DOI: 10.1038/nature19820 (2016).DOI
  28. [28]Kong, W. et al. Path towards graphene commercialization from lab to market. Nat. Nanotechnol. 14, 927–938, DOI: 10.1038/s41565-019-0555-2 (2019).DOI
  29. [29]Guo, G. Y., Murakami, S., Chen, T.-W. & Nagaosa, N. Intrinsic Spin Hall Effect in Platinum: First-Principles Calculations. Phys. Rev. Lett. 100, 096401, DOI: 10.1103/PhysRevLett.100.096401 (2008).DOI
  30. [30]Sagasta, E. et al. Tuning the spin Hall effect of Pt from the moderately dirty to the superclean regime. Phys. Rev. B 94, 060412, DOI: 10.1103/PhysRevB.94.060412 (2016).DOI
  31. [31]Zhang, Y. et al. Different types of spin currents in the comprehensive materials database of nonmagnetic spin Hall effect. npj Comput. Mater. 7, 167, DOI: 10.1038/s41524-021-00635-0 (2021).DOI
  32. [32]Ji, Y., Zhang, W., Zhang, H. & Zhang, W. Spin Hall conductivity and anomalous Hall conductivity in full Heusler compounds. New J. Phys. 24, 053027, DOI: 10.1088/1367-2630/ac696c (2022).DOI
  33. [33]Xu, X. et al. High-throughput calculation of large spin Hall conductivity in heavy-metal-based antiperovskite compounds. Mater. Genome Eng. Adv. e69 (2024).DOI
  34. [34]Zhou, J., Poncé, S. & Charlier, J.-C. High-throughput calculations of spin Hall conductivity in non-magnetic 2D materials. npj 2D Mater. Appl. 9, 39, DOI: 10.1038/s41699-025-00562-4 (2025).DOI
  35. [35]Qiao, J., Zhou, J., Yuan, Z. & Zhao, W. Calculation of intrinsic spin Hall conductivity by Wannier interpolation. Phys. Rev. B 98, 214402, DOI: 10.1103/PhysRevB.98.214402 (2018).DOI
  36. [36]Vitale, V. et al. Automated high-throughput Wannierisation. npj Comput. Mater. 6, 66, DOI: 10.1038/s41524-020-0312-y (2020).DOI
  37. [37]Garrity, K. F. & Choudhary, K. Database of Wannier tight-binding Hamiltonians using high-throughput density functional theory. Sci. Data 8, 106, DOI: 10.1038/s41597-021-00885-z (2021).DOI
  38. [38]Gresch, D. et al. Automated construction of symmetrized Wannier-like tight-binding models from ab initio calculations. Phys. Rev. Mater. 2, 103805, DOI: 10.1103/PhysRevMaterials.2.103805 (2018).
  39. [39]Mounet, N. et al. Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds. Nat. Nanotechnol. 13, 246–252, DOI: 10.1038/s41565-017-0035-5 (2018).DOI
  40. [40]Campi, D., Mounet, N., Gibertini, M., Pizzi, G. & Marzari, N. Expansion of the Materials Cloud 2D Database. ACS Nano 17, 11268–11278, DOI: 10.1021/acsnano.2c11510 (2023).DOI
  41. [41]Ward, L. et al. Matminer: An open source toolkit for materials data mining. Comput. Mater. Sci. 152, 60–69, DOI: https://doi.org/10.1016/j.commatsci.2018.05.018 (2018).
  42. [42]Mostofi, A. A. et al. An updated version of wannier90: A tool for obtaining maximally-localised Wannier functions. Comput. Phys. Commun. 185, 2309–2310, DOI: 10.1016/j.cpc.2014.05.003 (2014).DOI
  43. [43]Lundberg, S. M. & Lee, S.-I. A Unified Approach to Interpreting Model Predictions. In Guyon, I. et al. (eds.) Advances in Neural Information Processing Systems, vol. 30 (Curran Associates, Inc., 2017).arxiv.org/abs/1705.07874
  44. [44]Tsirkin, S. S. High performance Wannier interpolation of Berry curvature and related quantities with WannierBerri code. npj Comput. Mater. 7, 1–9, DOI: 10.1038/s41524-021-00498-5 (2021).DOI
  45. [45]Sala, G. & Gambardella, P. Giant orbital Hall effect and orbital-to-spin conversion in 3d, 5d, and 4f metallic heterostructures. Phys. Rev. Res. 4, 033037, DOI: 10.1103/PhysRevResearch.4.033037 (2022).DOI
  46. [46]Xue, D. et al. Accelerated search for materials with targeted properties by adaptive design. Nat. Commun. 7, 11241, DOI: 10.1038/ncomms11241 (2016).DOI
  47. [47]Khalak, Y., Tresadern, G., Hahn, D. F., de Groot, B. L. & Gapsys, V. Chemical Space Exploration with Active Learning and Alchemical Free Energies. J. Chem. Theory Comput. 18, 6259–6270, DOI: 10.1021/acs.jctc.2c00752 (2022).DOI
  48. [48]Roy, A., Guimarães, M. H. D. & Šafarik, J. Conventional Free Energies. Phys. Rev. Mater. 6, 045004, DOI: 10.1103/PhysRevMaterials.6.045004 (2022).
  49. [49]Sahu, P., Bhowal, S. & Satpathy, S. Effect of the inversion symmetry breaking on the orbital Hall effect: A model study. Phys. Rev. B 103, 085113, DOI: 10.1103/PhysRevB.103.085113 (2021).DOI
  50. [50]Go, D. & Lee, H.-W. Orbital torque: Torque generation by orbital current injection. Phys. Rev. Res. 2, 013177, DOI: 10.1103/PhysRevResearch.2.013177 (2020).DOI
  51. [51]Zhu, Y. et al. Large spin hall conductivity in low-symmetry semiconductor ZrSe3. J. Alloy. Compd. 918, 165579, DOI: https://doi.org/10.1016/j.jallcom.2022.165579 (2022).
  52. [52]MacNeill, D. et al. Control of spin–orbit torques through crystal symmetry in WTe2/ferromagnet bilayers. Nat. Phys. 13, 300–305, DOI: 10.1038/nphys3933 (2017).DOI
  53. [53]Giannozzi, P. et al. QUANTUM ESPRESSO: a modular and open-source software project for quantum simulations of materials. J. Phys.: Condens. Matter 21, 395502, DOI: 10.1088/0953-8984/21/39/395502 (2009).DOI
  54. [54]Stokes, H. T. & Hatch, D. M. FINDSYM: program for identifying the space-group symmetry of a crystal. J. Appl. Crystallogr. 38, 237–238, DOI: 10.1107/S0021889804031528 (2005).DOI
  55. [55]Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized Gradient Approximation Made Simple. Phys. Rev. Lett. 77, 3865–3868, DOI: 10.1103/PhysRevLett.77.3865 (1996).DOI
  56. [56]Hamann, D. R. Optimized norm-conserving Vanderbilt pseudopotentials. Phys. Rev. B 88, 085117, DOI: 10.1103/PhysRevB.88.085117 (2013).DOI
  57. [57]van Setten, M. J. et al. The PseudoDojo: Training and grading a 85 element optimized norm-conserving pseudopotential table (MMC S2. Notebook providing the analysis of the test results of the ONCVSP PBE v0.4 PseudoDojo tables.). Comput. Phys. Commun. 226, 39–54, DOI: 10.1016/j.cpc.2018.01.012 (2018).
  58. [58]Go, D., Jo, D., Kim, C. & Lee, H.-W. Intrinsic Spin and Orbital Hall Effects from Orbital Texture. Phys. Rev. Lett. 121, 086602, DOI: 10.1103/PhysRevLett.121.086602 (2018).DOI
  59. [59]Bhowal, S. & Satpathy, S. Intrinsic orbital and spin Hall effects in monolayer transition metal dichalcogenides. Phys. Rev. B 102, 035409, DOI: 10.1103/PhysRevB.102.035409 (2020).DOI
  60. [60]Sinova, J. et al. Universal Intrinsic Spin Hall Effect. Phys. Rev. Lett. 92, 126603, DOI: 10.1103/PhysRevLett.92.126603 (2004).DOI
  61. [61]Matsusalem, F. et al. Quantization of spin Hall conductivity in two-dimensional topological insulators versus symmetry and spin-orbit interaction. Phys. Rev. B 100, 245430, DOI: 10.1103/PhysRevB.100.245430 (2019).
  62. [62]Webber, J. B. W. A bi-symmetric log transformation for wide-range data. Meas. Sci. Technol. 24, 027001, DOI: 10.1088/0957-0233/24/2/027001 (2012).DOI

Paper details

vlnc:W_6ak3npmr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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