betaFeedback

WorksPhysics and Astronomy

Machine-learning Guided Search for Phonon-mediated Superconductivity in Boron and Carbon Compounds

Abstract

We present a workflow that iteratively combines ab-initio calculations with a machine-learning (ML) guided search for superconducting compounds with both dynamical stability and instability from imaginary phonon modes, the latter of which have been largely overlooked in previous studies. Electron-phonon coupling (EPC) properties and critical temperature (Tc_c) of 417 boron, carbon, and borocarbide compounds have been calculated with density functional perturbation theory (DFPT) and isotropic Eliashberg approximation. Our study addresses Tc_c convergence of Brillouin zone sampling with an ansatz test, stabilizing imaginary phonon modes for significant EPC contributions and comparing performance of two ML models especially when including compounds of dynamical instability. We predict a few promising superconducting compounds with formation energy just above the ground state convex hull, such as Ca5_5B3_3N6_6 (35 K), TaNbC2_2 (28.4 K), Nb3_3B3_3C (16.4 K), Y2_2B3_3C2_2 (4.0 K), Pd3_3CaB (7.0 K), MoRuB2_2 (15.6 K), RuVB2_2 (15.0 K), RuSc3_3C4_4 (6.6 K) among others.

Introduction

The pursuit of high-temperature superconductivity (SC) is a challenging and active research area. The recent discovery of superconducting temperature (TcT_{\mathrm{c}}) near 200 K in H3_3S at the high pressure of 150 GPa[1] has re-energized the field to focus on phonon-mediated SC. For phonon-mediated SC, high phonon frequency from light atomic mass with strong electron-phonon coupling (EPC) is beneficial, as shown by the recent discovery of near room temperature TcT_{\mathrm{c}} for metal hydrides under high pressure[2–10]. But making metallic bonds with hydrogen requires extremely high pressure. B and C compounds are known for their diverse structures and due to their light atomic mass, sizable TcT_{\mathrm{c}} has been found in compounds [11–15] ranging from MgB2_2 to metal intercalated graphite and fullerenes. Therefore, metallic B and C containing compounds with diverse structures are promising pools to search for new phonon-mediated superconductors. However, at ambient pressure, increasing TcT_{\mathrm{c}} even just higher than the 40 K of MgB2_2 [20] [11, 16, 17, 19] has been difficult [21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33]. While the recent report of a TcT_{\mathrm{c}}=32 K for MoB2_2 under the pressure of 110 GPa is a promising development [34], the large temperature and pressure gaps between MgB2_2 and H3_3S still remain, which motivates intensive search for new phonon-mediated SC compounds. Despite the challenges faced in experiments, theoretical studies continue to provide valuable insights into potential SC materials and their properties, such as SC in FeB4_4 was first predicted[35] and then verified[36]. Density functional perturbation theory (DFPT) [37, 38] is one of the most robust ab-initio methods [10] to compute the EPC matrices over the full Brillouin zone (BZ). One can then employ either isotropic Eliashberg approximation or Green function-based anisotropic Migdal-Eliashberg equations to compute TcT_{\mathrm{c}} [40, 41, 42, 44, 45, 46, 47] [1]. Therefore, computational exploration of compounds containing atoms slightly heavier than H, such as B and C[11–15, 48], with strong EPC is a promising venue to search for high-temperature SC at ambient pressure and also fill the large materials gap between MgB2_2 and H3_3S.

Machine learning (ML) and Artificial intelligence (AI) are increasingly taking important roles for predicting materials properties including SC. ML models usually offer a computationally inexpensive alternative to DFPT calculations for evaluating superconducting properties, completing predictions in minutes compared to the months, typically required for high throughput DFPT calculations of hundreds of compounds. Previous studies performing ML predictions based on random forest [49], regression [50–53], classification [50], natural language processing (NLP) [54], and deep learning [55] models have trained on experimental data, mostly from SuperCon database [36]. Comparing to the more expensive and time consuming experiments to explore many new compounds to generate SC data for ML, the high throughput (HT) ab-initio first-principles approaches are valuable tools to obtain the data that can be trained to predict potential SC compounds. Several recent studies have utilized ab-initio computed data for training ML model and predicting SC. One approach involves performing BCS-inspired screening of materials to identify potential candidates based on certain key properties such as the Debye temperature and the density of states at the Fermi level (N(EF)N(E_{\mathrm{F}})) [9]. A study similar to that described in Ref. [9] has been conducted on a vast range of materials, but restricting the size of the compounds to eight or fewer atoms, as reported in Ref. [58]. More recently, a new ML model suitable for the limited size of EPC dataset has been developed to predict Eliashberg spectral functions α2F(ω)\alpha^{2}F(\omega) rather than the overall EPC strength, which represents an important step to account for the correlation between phonon density of state and EPC spectral function [59]. In a recent study [60], a ML approach was used to predict the maximum TcT_{\mathrm{c}} and corresponding pressure of binary metal hydrides. The input layer consisted of atomic properties of the heavier metallic atom, while the output layer had two nodes representing TcT_{\mathrm{c}} and pressure. The ab-initio data utilized in this study has been collected from literature. Recently, ML-driven search with experimental feedback was also performed to discover a novel superconductor in Zr-In-Ni systems [61].

Notably in these previous HT and ML studies, compounds of dynamical instability with imaginary phonon modes have been largely discarded. However, as shown by our recent EPC study [62] on Y2_{2}C3_{3} with experimentally known TcT_{\mathrm{c}}=18K[32], imaginary phonon of C dimer wobbly motion once stabilized can carry significant EPC contributions, which explains well the observed sizable TcT_{\mathrm{c}}. In recent model Hamiltonian studies, phonon softening and anharmonicity have also been found to enhance TcT_{\mathrm{c}} [63, 64]. Here we present a workflow that iteratively combines ab-initio calculations with an ML-guided search across the dataset of compounds with both dynamical stability and instability from imaginary phonon modes by focusing on boron/carbon/borocarbide (B/C/B+C) compounds. Ab-initio calculations were performed to compute the EPC strength (λ\lambda), the logarithmic average phonon frequency (ωlog⁡\omega_{\log}) and TcT_{\mathrm{c}} of 417 compounds employing DFPT and isotropic Eliashberg approximation. Two major issues arise during DFPT calculations: choosing appropriate BZ sampling (k and q-mesh) for convergence and the problem of calculated dynamic instability. To address the convergence problem, we developed an ansatz test to check the convergence of EPC properties, particularly the TcT_{\mathrm{c}}. For dynamically unstable compounds, we employed large electronic smearing, lattice distortion, and pressure to stabilize them. We then calculated their EPC properties, which were included in building the ML models. We evaluated ML models, specifically the crystal graph convolutional neural network (CGCNN) and the atomistic line graph neural network (ALIGNN), trained utilizing ab initio computed data to predict SC properties. The key new findings of this study are two-fold. First is the inclusion of the previously overlooked dynamical unstable compounds in the EPC training dataset, accounting for a sizable 20% of all compounds being studied, which would otherwise be omitted. Second, by including compounds with imaginary phonon modes, we reveal a clear performance gap between the two ML architectures, reinforcing the widely accepted principle that the completeness of a training dataset is as crucial as the ML model itself. Among the two models, especially when including the dynamically unstable compounds, ALIGNN consistently outperforms the CGCNN in predicting EPC properties. We predict a few promising SC compounds with formation energy just above the ground state convex hull. For dynamically stable systems, we predict TaNbC2_{2} (28.4 K), Nb3_{3}B3_{3}C (16.4 K), Y2_{2}B3_{3}C2_{2} (4.0 K) among others. For systems with dynamical instability and imaginary phonon modes, we predict Ca5_{5}B3_{3}Nb with a TcT_{\mathrm{c}} as high as 35-42 K, besides Pd3_{3}CaB (7.0 K), and a few Ru compounds of MoRuB2_{2} (15.6 K), RuVB2_{2} (15.0 K), and RuSc3_{3}C4_{4} (6.6 K).

Results

Machine learning guided workflow

A ML-guided search workflow is employed in this study. It can be divided into three parts: data extraction, DFPT calculations, and training ML models. We will discuss each part in the following three sections (A, B and C) before presenting the main results in the last three sections (D, E and F). As illustrated in Figure 1 for obtaining the crystal structures of all the known B/C/B+C compounds, we utilized the Materials Project (MP) database [65–67], which offers a diverse range of compounds, including experimentally synthesized and theoretically predicted ones. We selected those B and C compounds in the MP database that meet certain criteria: being metallic with negative formation energy, excluding oxides, C60_{60} and Lanthanides except for La. Approximately 1500 compounds fall within this category, out of which 400 exhibit magnetic moments, and these magnetic cases are not considered in the present study. Consequently, our focus narrows down to around 1100 nonmagnetic compounds, forming the pool for investigating phonon-mediated superconductivity. To manage computational cost, we set a further criterion that considers only systems with primitive cells containing 40 atoms or less and composed of up to four different elements (Ntype≤4N_{\mathrm{type}} \leq4). This refinement narrows our selection to approximately 700 compounds. These 700 include 121 compounds with known TcT_{\mathrm{c}} from experimental measurement as in SuperCon database (113 being dynamically stable and 8 dynamically unstable) and the other 579 compounds of unknown TcT_{\mathrm{c}}. We will first discuss the 113 compounds with known TcT_{\mathrm{c}} and also dynamical stability (no imaginary phonon modes), while the remaining 8 compounds with dynamical instability will be discussed later.

Figure 1. Machine learning workflow. The numbers are the counts of compounds involved in each step. See main text for the detailed summary.

Figure 2 provides a statistical description of the 113 B/C/B+C compounds with known TcT_{\mathrm{c}} and dynamic stability, which are 53 superconductors (SC) and 60 non-superconductors (NSC). We reviewed and corrected any inaccuracies or discrepancies found in the SuperCon database through an extensive literature review [29, 33, 34, 36, 37, 41, 42, 44, 46] [16, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115, 116, 117, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152].

Figure 2. Statistical description of dynamically stable 113 compounds, whose (non-) superconductivity have been experimentally measured. (a) Number of compounds according to elements, each bar is partitioned into two color segments, with red representing the number of superconductors (SC), while blue denote non-superconductors (NSC).; (b) Proportion of boron and carbon compounds (c) Histogram for energy above the convex hull in (eV/atom), (d) Distribution according to space group, (e) Crystal structure of MgB2\mathrm{MgB}_2. (f)-(m): Crystal structures of known superconductors. (f) NbC\mathrm{NbC} (10.03 K)[70], (g) Mo2BC\mathrm{Mo}_2\mathrm{BC} (7.5 K) [124], (h) MgCNi3\mathrm{MgCNi}_3 (8 K)[154], (i) Mo2GaC\mathrm{Mo}_2\mathrm{GaC} (3.9 K) [155], (j) SrC6\mathrm{SrC}_6 (1.65 K)[156], (k) YC2\mathrm{YC}_2 (3.89K)[121], (l) YB6\mathrm{YB}_6 (7.2 K) [104], and (m) LaPt2B2C\mathrm{LaPt}_2\mathrm{B}_2\mathrm{C} (10 K)[82].

Figure 2 (a)-(c) illustrate the distribution of different elements in these systems, allocation of B/C/B+C compounds and thermodynamic stability, respectively. The thermodynamic stability of these compounds, indicated by the energy above the ground state convex hull (ΔEh\Delta E_{\mathrm{h}}), are obtained from the MP database[65]. Approximately 71% of the compounds are on the ground state convex hull with ΔEh∼0\Delta E_{\mathrm{h}} \sim0. Around 18% of the cases have an ΔEh\Delta E_{\mathrm{h}} within 0.05 eV/atom, while the remaining 11% have ΔEh\Delta E_{\mathrm{h}} larger than 0.05 eV/atom. Figure 2(d) shows the distributions of these compounds based on their space groups (SGs). In Figure 2(a) and (d), each bar is partitioned into two segments with the red segment representing the number of SC, while blue segment for NSC. From Figure 2(a), many known SC compounds are associated with transition metals (TM) such as Y, La, Ni, Rh, Mo, Nb and others. Figure 2(e)-(m) depict the crystal structures of the representative SC compounds in the top 6 SGs among the experimentally known SC. In these structures, B and C atoms form various structural motifs: honeycomb lattice of B in MgB2_2 (Figure 2(e)); monomers in NbC (Figure 2(f)), MgCNi3_3 (Figure 2(h)) and Mo2_2GaC (Figure 2(i)); dimers in YC2_2 (Figure 2(k)); graphene sheets in SrC6_6 (Figure 2(j)); chains of lighter elements in Mo2_2BC (Figure 2(g)) and LaPt2_2B2_2C (Figure 2(m)); octahedral cage structures in YB6_6 (Figure 2(l)). In terms of the lattice types of these known SC compounds, highly symmetric structures with hexagonal, tetragonal and cubic SGs have the most compounds, as shown in Figure 2(d). Similarly, statistical description for the SC compounds with predicted TcT_{\mathrm{c}} that have not been measured, akin to Figure 2, is illustrated in Supplementary Materials (SM) Figure S1.

Statistical description of dynamically stable materials whose superconductivity has not been experimentally measured

Figure S1. Statistical description of dynamically stable materials, whose (non-) superconductivity hasn’t been experimentally measured. Each bar is partitioned into segments, with the segments colored in “red” representing the number of superconductors (SC), while segments in “blue” denote nonsuperconductors (NSC).; (a) Number of compounds according to elements, (b) Proportion of boron and carbon compounds (c) Statistics in terms of energy above the convex hull, (d) Distribution according to spacegroups, (e)-(g): Crystal structures of some potential superconductors with their respective spacegroups.

Overview of DFPT calculations

After the crystal structures of B and C compounds have been collected, the next step as shown in Figure 1 is to do HT calculations with DFPT on EPC properties for compounds with both known and unknown TcT_{\mathrm{c}} using our recently developed high-throughput electronic structure package (HTESP) [157]. We performed DFPT calculations and computed EPC properties using the isotropic Eliashberg approximation. The accuracy of the EPC data is crucial for building reliable ML models. We have encountered two major challenges in the HT calculations with DFPT. One is the convergence with respect to BZ sampling and the other is dynamic instability. The first obstacle involved determining appropriate BZ samplings (k- and q-meshes) to compute the EPC properties, as these calculations become computationally expensive with dense meshes. To reduce the computational cost, we initiated an efficient screening by using a k-point mesh that accurately describes the ground-state structures and energetics. EPC quantities are then interpolated to fine k-mesh only twice the size of the coarse k-mesh. But we noticed that calculations with such grid combination can lead to inaccurate predictions, with discrepancies as high as approximately 10% of the total compounds with known TcT_{\mathrm{c}}, giving NSC for known SC and vice versa. To address this problem, we developed an ansatz test to assess the convergence of TcT_{\mathrm{c}} with respect to the k-point mesh. This approach leverages the decaying behavior of TcT_{\mathrm{c}} with respect to Gaussian broadening width (σ\sigma), which is used in the double-delta integration. Initially, we acquired results using the DFPT method with the k-point mesh size from the MP database. Subsequently, we assessed the convergence of these results based on the convergence ansatz test. To ensure convergence, we repeated calculations with denser k-mesh for the cases where results did not pass this test. We applied this technique to the 113 dynamically stable compounds, and obtained reasonable accuracy for the calculated TcT_{\mathrm{c}} with a mean absolute error (MAE) of 2.21 K compared to experimental data, as presented in Supplementary Figure S2 (b). Moreover, incorporating dynamically unstable compounds further improves the dataset quality, as evidenced in supplementary Figure S2(c)–(d) by the enhanced Rlog⁡2R^{2}_{\log} score in the log⁡(Tc)\log(T_{\mathrm{c}}) comparison with experiment. Further details regarding the convergence with respect to the k- and q-point meshes are discussed in the Method section and SM. In addition to the 113 dynamically stable compounds with known TcT_{\mathrm{c}}, we also computed the EPC properties of 268 compounds of unknown TcT_{\mathrm{c}} with dynamical stability, resulting from the ML-guided search as summarized in Figure 1.

Four-panel comparison of predicted and experimental critical temperatures under k-grid refinement and dataset expansion

Figure S2. Comparison of Predicted and Experimental Critical Temperatures Under k-Grid Refinement and Dataset Expansion. (a) Comparing calculated and experimental critical temperature obtained from efficient calculations; Some largely deviated results are not enclosed within red rectangular box. (b) Same plot with improved results with denser k-grids for compounds with A>AMgB2A > A_{MgB_2}. Red rectangular box on the right plot, enclosed more data points along the y=xy = x axis (black solid line), compared to plot on the left. (c) Predicted log⁡(Tc)\log(T_c) vs. measured log⁡(Tcexp)\log(T_c^{exp}) for dynamically stable superconductors. (d) Same comparison after incorporating additional data obtained by stabilizing dynamically unstable compounds. The inclusion of these superconductors improves the coefficient of determination Rlog⁡2R_{\log}^{2} score from 0.838 to 0.867.

The second obstacle encountered in EPC calculations was the presence of imaginary phonon modes and dynamical instability in almost 146 compounds. Among them, the imaginary phonon modes in 36 compounds show large EPC contribution. Stabilizing these imaginary modes is crucial for calculating EPC properties in such systems. These imaginary modes can be stabilized through lattice distortion, pressure, and electronic smearing, with the later method being particularly effective in HT screening[62]. A comprehensive analysis of dynamical instability and its implications for superconductivity will be presented in the “Imaginary phonon modes and superconductivity” section.

In addition to these instances, there are cases up to 173 compounds, where EPC calculations are not complete due to either numerical problems or too large size of unit cells. For now, we will set these cases aside and will revisit them in the future.

Training and testing ML models

Besides the data extraction and DFPT calculations, training and testing ML models also play crucial roles in the ML-guided search workflow in Figure 1. We utilized two different ML models: CGCNN[158] and ALIGNN[159]. CGCNN maps 3D crystal structures to 2D graphs by using chemical element information and neighbor bonding distance to encode them into local chemical environments through convolution operations and updating the node features based on these descriptors. The updated node features are then aggregated to represent the entire crystal, which is connected to the output via a neural network. ALIGNN includes extra local chemical information, such as bonding angles, in addition to the crystal graphs in CGCNN with another auxiliary graph of bonding distances and angles. The parameters, including weights and biases, of the neural network connections are learned through training with available DFPT data.

As shown in Figure 1, we trained the initial ML models in Run 1 using a dataset of 250 dynamically stable compounds with 109 SC (calculated Tc>1T_{\mathrm{c}} > 1 K) and 141 NSC (calculated Tc<1T_{\mathrm{c}} < 1 K) including those 113 from SuperCon database that were already measured in experiments. This dataset encompasses 45 distinct SGs. For the purpose of evaluation, a separate collection of 58 stable compounds (18 SC and 40 NSC), belonging to 27 unique SGs (6 of which are not part of the training 45), was reserved exclusively for independent testing and was not involved in the ML training process. Moving from Run 1 to Run 2, we use the Run 1 ML model (trained on 250 dynamically stable compounds) to predict TcT_{\mathrm{c}} for the remaining B–C compounds, which is fewer than 450, because some compounds were already found to be dynamically unstable but without EPC data during Run 1 ML training. We then sort the predictions and select the top candidates of both SC and NSC for more DFPT calculations to close the ML-guided loop. In Run2, the original 250 stable systems were augmented with an additional 73 dynamically stable systems to refine the ML models. In the concluding stage of Run 3, we incorporated the results derived from the compounds with dynamic instability. Notably, this stage included results from 2 new SGs. Among the 36 results in this category, we incorporated 28 into the training set and added 8 compounds into the independent test set. In the overall count of 417 compounds with converged EPC properties, 181 were classified as SCs, whereas 236 were categorized as NSCs. The progress from Run 1 to Run 3 constitutes a loop, as illustrated in Figure 1. In summary, our methodology has a series of iterations involving training and testing ML models with increasingly comprehensive datasets from dynamically stable to unstable compounds.

Comparison of ML models

Next, we will present the main results of the ML models and guided search, highlight the notable compounds, including cases of dynamical stability and instability. Figure 3(a-Figure 3d) depict the ML-predicted vs. DFPT-calculated λ\lambda, ωlog\omega_{\mathrm{log}}, TcT_{\mathrm{c}}, and Tc′T_{\mathrm{c}}^{\prime} using the dynamically stable systems in Run 1, respectively. Here, TcT_{\mathrm{c}} represents the critical temperature computed using DFPT-computed λ\lambda and ωlog\omega_{\mathrm{log}} or predicted directly from ML models, while Tc′T_{\mathrm{c}}^{\prime} is calculated from the ML-predicted λML\lambda^{\mathrm{ML}} and ωlogML\omega_{\mathrm{log}}^{\mathrm{ML}} using Eq. (1) in a postprocessing manner.

Figure 3. Graph neural network-based regression models for dynamically stable compounds. Predicting λ\lambda, ωlog⁡\omega_{\log}, TcT_{\mathrm{c}}, and Tc′T'_{\mathrm{c}} using CGCNN (red circles) and ALIGNN (blue upper triangle) models for the independent test set in comparison to the DFPT-calculated results in Run 1 (left panel, (a)–(d)) and Run 2 (right panel, (e)–(h)). Run 1 uses 250 training samples (45 space groups) and 58 test samples (27 space groups). Run 2 employs 323 training samples (54 space groups), with the test set unchanged. Similarly, Mean absolute error (MAE) are presented in inset. Red and blue arrows show the clustering pattern, discussed in the main text.

Tc′=ωlogML1.2exp⁡[−1.04(1+λML)λML−μc∗(1+0.62λML)],(1)T_{\mathrm{c}}^{\prime} = \frac{\omega_{\mathrm{log}}^{\mathrm{ML}}}{1.2} \exp\left[-\frac{1.04(1+\lambda^{\mathrm{ML}})}{\lambda^{\mathrm{ML}}-\mu_{\mathrm{c}}^{*}(1+0.62\lambda^{\mathrm{ML}})}\right], \tag*{(1)}

Here, It is important to note that the data first used do not include dynamically unstable cases. For Run 1, we trained ML models using 250 compounds, divided into training, validation, and testing sets in a ratio of 0.8:0.1:0.1. The mean absolute errors (MAEs) for the 10% test set in each training iteration are documented in Supplementary Table S1. An additional independent test set of 58 compounds was used to assess the predictability of the ML models and their MAE and predictions are plotted in Figure 3. Apart from MAE, we also computed the R2R^{2}-score for the ML parity plot, which is presented in Supplementary Table S2. As expected, accurately predicting the exact EPC properties is challenging. However, the performance of the ML models improves when superconductors with stabilized phonons are included. The training process consisted of 3000 epochs using default settings provided by the ML packages, which have been thoroughly investigated in the original work [158, 159]. In both CGCNN and ALIGNN, training and validation procedures are employed. Checkpoints are established at regular intervals to store crucial parameters such as model weights and architecture. The ML packages operate automatically to retain and update the model exhibiting the best performance, determined by the lowest validation error. This iterative process also acts to mitigate overfitting concerns. Our Supporting Figure S3 and Supporting Table S2 demonstrate that training up to 1500 epochs is sufficient to achieve the lowest validation error for both ALIGNN and CGCNN model. The performance of the models was evaluated by computing the MAE between the predicted and target quantities for the independent test set. The MAEs for CGCNN-predicted λ\lambda and ωlog⁡\omega_{\log} stand at 0.23 and 105 K, respectively, while the corresponding numbers for ALIGNN are slightly higher at 0.28 and 113 K (Figure 3(a) and (Figure 3b)). In terms of predicting TcT_{\mathrm{c}}, the CGCNN and ALIGNN models yield MAEs of 2.4 K and 3.8 K, respectively. Despite the relatively small magnitude of MAEs, the ML outcomes exhibit distinct clustering patterns, as shown in Figure 3(c). Specifically, for the CGCNN model, the results tend to cluster closely along the “DFPT” axis (red arrow in Figure 3(c)), whereas for ALIGNN, the clustering is pronounced along the “ML prediction” axis (blue arrow in Figure 3(c)). An alternative approach, rather than directly training and predicting TcT_{\mathrm{c}}, is to utilize ML-predicted values, specifically λML\lambda^{\mathrm{ML}} and ωlog⁡ML\omega_{\log}^{\mathrm{ML}}, to estimate Tc′T'_{\mathrm{c}} using Eq. (1). This modification not only enhances predictive performance for both models but also slightly ameliorates the issue of clustering [Figure 3(d)]. A comparable approach has been employed in a recent study [160], wherein λ\lambda and ωlog⁡\omega_{\log} can be directly acquired from first principles calculations.

Then we use the predicted Tc′T'_{\mathrm{c}} to rank the remaining compounds to pick the ones with high and low Tc′T'_{\mathrm{c}} for additional DFPT calculations. In this next stage, we utilized ML-guided search to expand the dataset to 323 systems characterized by 54 distinct SGs, the ML models underwent further training. Subsequently, the improved ML models were subjected to the same independent testing, employing the same set of 58 system test cases. The outcomes of the Run 2 are presented in Figure 3(e)-(Figure 3h). The results demonstrate a notable improvement in addressing the issue of prediction clustering with the expansion of the training dataset, all while maintaining reasonable accuracy. ALIGNN improves the prediction of λ\lambda and ωlog⁡\omega_{\log} with MAEs of 0.24 and 93 K, respectively, compared to CGCNN’s MAEs of 0.26 and 97 K. However, Tc′T'_{\mathrm{c}} calculated from ML-predicted λML\lambda^{\mathrm{ML}} and ωlog⁡ML\omega_{\log}^{\mathrm{ML}} shows a significant improvement for ALIGNN, with an MAE of 2.7 K. Notably, ωlog⁡\omega_{\log} is more accurately predicted than λ\lambda and then TcT_{\mathrm{c}}, because the former is directly related to the overall bonding strength and cohesive energy, while the later depends on the details of the Fermi surface and the EPC matrix elements.

As mentioned above, from the ML-guided search, we find quite some number of compounds with imaginary phonon and dynamical instability. In all the previous high throughput phonon-mediated SC studies, these dynamically unstable compounds are simply discarded, but as shown by our recent study [62] on Y2_2C3_3, some imaginary phonon modes once stabilized can carry a large EPC and give rise to a large TcT_{\mathrm{c}}. Thus, specifically here we also include dynamically unstable compounds in ML. As far as we know, this is the first ML trained with both dynamically stable and unstable compounds for phonon-mediated SC. We carried out a distinct training and testing phase for ML models, incorporating dynamically unstable cases after stabilization, denoted as Run 3. These results predominantly included nonzero TcT_{\mathrm{c}}. Among these, 28 results were added to the training dataset, while the remaining 8 were added for independent testing. In total, the independent test set now consists of the original 58 dynamically stable and additional 8 stabilized compounds with imaginary phonon modes. Both the training and test datasets were balanced across various SGs. The outcomes of Run 3 are illustrated in Figure 4(a)-(Figure 4d). In Run 3, the ALIGNN model consistently outperforms the CGCNN model across different superconducting properties. The ALIGNN model achieves MAEs of 0.27 for λ\lambda, 86 K for ωlog⁡\omega_{\log}, 4.3 K for TcT_{\mathrm{c}}, and 3.5 K for Tc′T'_{\mathrm{c}}. In contrast, the CGCNN model records MAEs of 0.34, 104 K, 4.5 K, and 5.3 K for the respective properties. The reason is that ALIGNN includes the bonding angles as part of training parameters, which better describe the dynamically unstable compounds, because imaginary phonon modes often involves lower-energy bond rotation with changing angle than higher-energy bond stretching vibration. Based on the information presented in Figure 3 and Figure 4, it is apparent that the ML model performs better in learning the parameters ωlog⁡\omega_{\log} than λ\lambda and then TcT_{\mathrm{c}}. As a result, it is justifiable to utilize Eq. (1) to estimate the ML-predicted critical temperature TcT_{\mathrm{c}} rather than relying solely on the directly predicted TcT_{\mathrm{c}} values. Furthermore, we separately computed the MAE for SC and NSC, as shown in Supplementary Table S3. ALIGNN demonstrates better accuracy in predicting ωlog⁡\omega_{\log} compared to CGCNN for both categories. However, it predicts λ\lambda more accurately than CGCNN for NSC, while performing slightly worse for SC. Additionally, a classification model was employed, with the results presented in the Supplementary Table S4.

Figure 4. Graph neural network-based regression models similar to Figure 3, but including both stable and stabilized unstable compounds. Predicting λ\lambda, ωlog⁡\omega_{\log}, TcT_{\mathrm{c}}, and Tc′T'_{\mathrm{c}} using CGCNN (red circles) and ALIGNN (blue upper triangle) models for the independent test set including both stable and dynamically unstable cases in comparison to the DFPT-calculated results in Run3. Run 3 employs 351 training samples (56 space groups) and 66 test samples (27 space groups). MAEs are presented in inset.

Dynamically stable systems

In Table I, we present the results of our ML-guided search for dynamically stable compounds with DFPT-calculated TcT_{\mathrm{c}} exceeding 10 K, while the complete list of EPC properties of the 381 dynamically stable compounds are presented in SM from Table S10 to S17. It is worth noting that the compounds with high TcT_{\mathrm{c}} tend to be thermodynamically metastable, as indicated by their formation energy significantly above the convex hull. B2_2CN in different phases, with ΔEh\Delta E_{\mathrm{h}} up to 0.35 eV/atom, has been predicted to exhibit SC, consistent with earlier theoretical findings [161]. TaC and NbC, both have sizable TcT_{\mathrm{c}} in the metastable hexagonal structure, while their more stable cubic structures in Fm3ˉmFm\bar{3}m have already been observed with SC [162, 163]. In Table I, the first compound close to the ground state

Imaginary phonon modes and superconductivity

TABLE I: DFPT calculated EPC results for experimentally unknown systems with Tc>10T_c > 10 K, whose superconductivity has not yet been measured; SG is spacegroup, and ΔEh\Delta E_h is taken from materials project. [65]

CompoundSGΔEh\Delta E_h (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_c (K)
B2_2CNR3m0.341.6657860.7
B2_2CNP3m10.351.1156734.9
Mo7_7B24_{24}P-6m20.151.2838629.6
TaNbC2_2R-3m0.021.4132628.4
TcBP63_3/mmc0.261.3233326.5
TaCP-6m20.411.5423522.8
ZrBCP63_3/mmc0.451.1232120.0
Ta2_2CNI41_1/amd0.131.9416219.8
Ta2_2CNP4/mmm0.152.7012719.5
B2_2CNP-4m20.320.8064418.8
TaB2_2P6/mmm0.01.2225418.1
NbFeBP-6m20.421.6417318.0
Nb3_3B3_3CCmcm0.021.2522916.4
V2_2CNR-3m0.121.0032316.2
NbCP63_3/mmc0.150.8445515.3
ZrMoB4_4P6/mmm0.091.3019315.1
NbVCNR3m0.160.9927413.4
Ta4_4C3_3Pm-3m0.131.4613812.6
NbVC2_2R-3m0.110.8336511.9
Nb2_2CNR-3m0.080.8831011.7
Nb4_4B3_3C2_2Cmcm0.051.0121911.3
TaVC2_2R-3m0.080.8034310.3

Table 1.

convex hull (ΔEh∼0.02\Delta E_h \sim0.02 eV/atom) is TaNbC2_2, whose crystal structure and EPC properties are plotted in Figure 5(a)-(Figure 5c). It crystallizes in the trigonal structure of R−3mR - 3m. The calculated EPC properties are λ=1.4\lambda= 1.4, ωlog⁡=326\omega_{\log} = 326 K, and Tc=28.4T_c = 28.4 K, with the majority of the contribution coming from phonons within the 3-6 THz range (Figure 5(b) and (c)), as well as a significant contribution from the 16-20 THz range of C-dominated modes. Other compounds close to the ground state convex hull with ΔEh≤0.02\Delta E_h \leq0.02 is TaB2_2. The presence of SC in TaB2_2 remains a subject of debate as discussed in the previous studies [95].

Figure 5. DFPT calculated EPC properties of a few dynamically stable compounds. (a) Crystal structure, (b) phonon dispersion, and (c) Eliashberg isotropic spectral function of TaNbC2_2, respectively. (d)-(k) X-B-C compounds (X=Nb,Y): Comparison between DFPT computed TcT_{\mathrm{c}} values with experiments for (d) Nb-B-C and (e) Y-B-C systems. Experimentally known results are represented by blue circle, while red circles are not reported ones shown along the y=xy=x dashed line as labeled. Crystal structures respectively for (f) Nb3_3B3_3C, and (g) Y2_2B3_3C2_2. Atoms are highlighted by the colorized symbols. Phonon dispersion projected with mode-resolved λ\lambda (green open circles) for (h) Nb3_3B3_3C, and (i) Y2_2B3_3C2_2 respectively. Eliashberg spectral functions for (j) Nb3_3B3_3C, and (k) Y2_2B3_3C2_2 respectively.

Moreover, we predict other ternary superconductors, such as Nb3_3B3_3C (ΔEh\Delta E_h of 0.018 eV/atom [65]), with TcT_c of 16.4 K (Figure 5(d), (Figure 5f), (Figure 5h) and (Figure 5j)). Comparison between the calculated and experimental TcT_c are also plotted (blue circles) for the Nb-B-C superconductors with known TcT_c. These ternary metallic borocarbides were all experimentally synthesized [164], however, some of their SC (red circles) have not been reported yet, with Nb3_3B3_3C has a predicted TcT_c as high as 16.4 K. To differentiate them from experimentally measured systems, we color-coded them in red and presented them along the y=xy = x line. The crystal structure of Nb3_3B3_3C exhibits an interesting layer-like arrangement as in Figure 5(f), where B atoms form strips of honeycomb lattice, while C being monomers. The λ\lambda-projected phonon dispersion (Figure 5(h)) analysis indicates that the SC is attributed to the presence of low-frequency soft phonon modes in the vicinity of the ZZ and TT points of the BZ. For ternary Nb-B-C systems, EPC calculations with isotropic approximation tend to overestimate TcT_c, while for Y-B-C systems, EPC calculations have also notable underestimations as plotted in Figure 5(e). Therefore, we also show the EPC properties of Y2_2B3_3C2_2 with a predicted TcT_c of 4 K, but having ΔEh∼0\Delta E_h \sim0 [65], along with other experimentally measured Y-B-C systems in Figure 5(e). Unlike Nb3_3B3_3C, the crystal structure of Y2_2B3_3C2_2 shows a layer of mixed B and C network sandwiching the Y layer (Figure 5(g)). The λ\lambda-projected phonon dispersion shows that the EPC properties are mostly contributed by phonon within the 4-10 THz energy range around the Γ\Gamma and ZZ points (Figure 5(i) and (Figure 5k)). Other compounds with calculated Tc<10T_c < 10 K and their EPC properties are listed in SM.

Imaginary phonon modes and superconductivity

In this section, we address the issue of dynamical instability observed in the DFPT-computed phonon dispersion (146 out of 700 compounds). Out of these 146 dynamical unstable compounds, 34 exhibit large EPC in their imaginary phonon modes. Figure 6 (a) and (b) provide the breakdown of these 34 compounds based on their constituent elements, while (c) and (d) show the distribution of systems in terms of formation energy and SG, respectively. As expected, a considerable portion of these compounds with imaginary phonon have formation energies above the convex hull. Superconductivity in Sc2_2C3_3, shown in Figure 6 (d), are very likely because SC has been observed in the same structure with larger cations of the same group as in Y2_2C3_3 [165] and La2_2C3_3 [166], which are closer to the convex hull. We have categorized these compounds into two groups based on whether the imaginary phonon modes occur at the Γ-point or elsewhere. The instabilities at Γ-point are represented by Sc2_2C3_3, Ta2_2B, and La3_3InB (Figure 6(i-l)) which constitutes 11 cases, whereas the other 23 compounds including MoB2_2 represents the case of dynamic instability outside of Γ-point, as illustrated in Figure 6(l) for MoB2_2. The unstable phonon modes possess a large mode-resolved λ\lambda in the vicinity of instability, λqν=γqνπN(EF)ωqν2\lambda_{q\nu} = \frac{\gamma_{q\nu}}{\pi N(E_F)\omega_{q\nu}^{2}}, where γqν\gamma_{q\nu} is the change in phonon linewidth due to EPC (See Figure 6). The systems with dynamical instability at Γ-point can be stabilized by obtaining a low-symmetry ground state structure through lattice distortion along the direction of the imaginary eigenmodes at Γ and performing a full ionic relaxation. The analysis presented here is expanded from our earlier work on Y2_2C3_3 [62]. The application of smearing to stabilize imaginary phonon modes has been explored in other systems, including β-phase Nix_xAl(1−x)_{(1-x)} [167], NiTi [168] and EuAl4_4 [169]. In these cases, instead of optical modes, the instability arises from acoustic modes. It was proposed that the dynamic instability observed in these materials is the result of strong EPC between nested electronic states near the Fermi level [167–169]. Previously as exemplified with Y2_2C3_3 [62] these imaginary phonon modes can significantly contribute to λ\lambda after stabilization. Although these cases represent only a small fraction of the compounds, disregarding them would exclude potential SC compounds with a sizable TcT_{\mathrm{c}}.

Figure 6. Imaginary frequency modes with large EPC on the soft modes. Fig. (a) and (b) represent the count of different elements and allocation of B/C/B+CB/C/B+C compounds in dynamically unstable systems respectively. Statistics according to spacegroup(c) and (d) energy above the convex hull. (e)-(h) Crystal structures of Sc2C3\mathrm{Sc}_{2}\mathrm{C}_{3}, Ta2B\mathrm{Ta}_{2}\mathrm{B}, La3InB\mathrm{La}_{3}\mathrm{InB}, and MoB2\mathrm{MoB}_{2} are shown, respectively. (i)-(l) display the phonon dispersion plots of the dynamically unstable systems corresponding to Sc2C3\mathrm{Sc}_{2}\mathrm{C}_{3}, Ta2B\mathrm{Ta}_{2}\mathrm{B}, La3InB\mathrm{La}_{3}\mathrm{InB}, and MoB2\mathrm{MoB}_{2}. First 3 plots represent the instability of soft phonon modes at Γ\Gamma-point, while the last one corresponds to the instability outside of Γ\Gamma. (m)-(p): EPC results for compounds with stabilized imaginary phonon modes. EPC projected (highlighted by green circles) phonon band dispersions for Sc2C3\mathrm{Sc}_{2}\mathrm{C}_{3} (P-30), Ta2B\mathrm{Ta}_{2}\mathrm{B} (D), La3InB\mathrm{La}_{3}\mathrm{InB} (D), and MoB2\mathrm{MoB}_{2} (S-0.1) respectively.

Performing EPC calculations on low-symmetry structures can be computationally expensive, especially when the system contains a large number of atoms like Sc2_2C3_3. Hence, it is recommended to first try stabilization using pressure and smearing. The former option of stabilizing through pressure can give interesting pressure-dependent properties, whereas the latter with increasing electronic smearing can be beneficial for HT computations. To stabilize the dynamically unstable systems, we applied pressure (ranging from 5 to 60 GPa) or used a larger electronic smearing (0.05, 0.06, 0.08, 0.1 Ry). In each case, we fully relaxed the structures. We computed the EPC properties for the stabilized systems around Γ-point and presented the results in Table II together with their SG and ΔEh\Delta E_h. Most of these compounds crystallize in high-symmetry structures such as cubic, hexagonal and tetragonal lattices. The relative ground state energy of relaxed low-symmetry structures compared to the original high-symmetry ones are listed as Distorted GS. We utilize pressure and the electronic smearing to stabilize the imaginary phonons in Y2_2C3_3, La2_2C3_3 and Sc2_2C3_3. However, for Ta2_2B and La3_3InB systems, pressure and smearing were insufficient for stabilization, so lattice distortion was employed. The calculated TcT_{\mathrm{c}} using DFPT for the stabilized systems agrees well with experimental data. For example, Al2_2Mo3_3C exhibits an instability at Γ akin to that observed in Sc2_2C3_3. After stabilization with electronic smearing, the calculated TcT_{\mathrm{c}} is 12.05 K, which is comparable to the experimental TcT_{\mathrm{c}} of 9.2 K [170]. We predict the EPC properties for Sc2_2C3_3, YBC, and MoB4_4 with a sizable TcT_{\mathrm{c}} of 27.9 K, 10.15 K and 7.58 K under ambient or moderate pressure. These compounds have formation energy higher (0.11, 0.42, and 0.26 eV/atom respectively) than the ground state convex hull, indicating they are metastable.

Figure 6(m)-(p) displays the λ\lambda-projected phonon dispersion for four different stabilized compounds: Sc2_2C3_3 (P-30), Ta2_2B (D), La3_3InB (D), and MoB2_2 (S-0.1), utilizing pressure of 30 GPa, distortion (D), distortion (D), and smearing of 0.1 Ry, respectively. By comparing these plots with Figure 6(i)-(Figure 6l), we can see that the soft optical phonon modes are stabilized and contribute significantly to λ\lambda near the Γ-point, represented by green open circles. Despite the slight lifting of phonon band degeneracy caused by distortion, the large contribution to EPC remains, which give rise to SC. The discovery of such systems is interesting as it presents opportunities for stabilization through pressure and alloying, leading to potentially high TcT_{\mathrm{c}} metastable compounds that may be synthesizable in experiment.

For the compounds with imaginary phonon modes away from the Γ point, i.e. MoB2_2-type, we also stabilized these dynamically unstable compounds with larger electronic smearing of 0.1 Ry and tabulated the results in Table. Table III. For example, compounds like MoB2_2 in the MgB2_2 structure have shown experimentally measured TcT_{\mathrm{c}} of 32 K under high pressure [34]. Moreover, we demonstrate that large electronic smearing can play a role similar to pressure in stabilizing imaginary phonon modes. The pressure-induced stabilization of MoB2_2 is presented in Supplementary Figure S4. Our calculated structural phase transition from R-3m to P6/mmm phase at 75 GPa agrees well with the experimentally reported critical pressure of 70 GPa [34]. The calculated electron–phonon coupling (EPC) at 75 and 90 GPa yields TcT_{\mathrm{c}} values of 28.4 K and 28.6 K, respectively, which are consistent with the experimental value of 32 K, as well as with results obtained from the electronic smearing approach used for high-throughput calculations. Notably, both approaches give comparable EPC parameters for the same stabilized low-lying phonon modes: at 75 GPa we obtain λ=1.85\lambda= 1.85, ωlog⁡=244\omega_{\log} = 244 K, and Tc=28.4T_{\mathrm{c}} = 28.4 K, while the smearing-based approach (Table III) gives λ=2.12\lambda= 2.12, $\omega_{\log} = 209 K, and Tc=27.3T_c = 27.3 K. Another notable example in Table III is Ca5_5B3_3N6_6 with ΔEh\Delta E_h of 0.03 eV/atom, which exhibits dynamical instability at the HH point and displayed a significant mode-resolved λ\lambda near its unstable phonon modes, as depicted in Figure 7 (d)-(e). After stabilizing the system with electronic smearing of 0.06 Ry, Ca5_5B3_3N6_6 exhibits a λ\lambda of 1.5. The ωlog⁡\omega_{\log} is found to be 372 K. Moreover, when computing λ\lambda (with broadening parameter σ=0.01\sigma= 0.01 Ry), the TcT_c is determined to be 35 K with Coulomb potential μc∗=0.16\mu_c^*=0.16 and 42.4 K with μc∗=0.10\mu_c^*=0.10, much larger than that of MgB2_2 (16-20 K) computed from isotropic approximation. It should be noted that Ca5_5B3_3N6_6 has been synthesized in cubic structure and Im3′m (229) SG with partial occupancy in the 8c site of Ca [172]. The crystal structure of this boronitride (Figure 7(a)-(c)) has a cage-like structure, similar to XB3_3C3_3 borocarbides (X = Ca,Ba,Sr,Y,La). For the stoichiometric Ca5_5B3_3N6_6, Figure 7(f) and (g) present the isotropic Eliashberg spectral function and electronic band structure, respectively. It shows an electron-doped band structure that connects to strong EPC as also found in other predicted SC compounds and studies. As listed in Table III, interestingly some of the trigonal compounds of NbMoC2_2 are in the same structure as the stable TaNbC2_2 in Table I. This shows that in this particular structure, with the substitution of neighboring group of early TMs, although bringing dynamical instability, the phonons once stabilized can still provide a large EPC contribution for a sizable TcT_c. As also listed in Table III, other compounds with sizable predicted TcT_c after stabilization that are also near GS hull are some notable ternary Ru compounds, MoRuB2_2 at 15.6 K, RuVB2_2 at 15.0 K, Pd3_3CaB at 7.0 K and RuSc3_3C4_4 at 6.6 K.

CompoundSGΔEh\Delta E_h (eV/atom) [65]λ\lambdaωlog⁡\omega_{\log} (K)TcT_c (K)TcExptT_c^{Expt} (K)
Ca5_5B3_3N6_6Im3′m0.031.537235.0a^{a}
MoB2_2P6/mmm0.1562.1220927.332[34]
NbMoC2_2R-3m0.1721.7221523.5
TaMo2_2C3_3P-3m10.1972.0117121.4
TaMoC2_2R-3m0.151.6918920.3
WVC2_2R-3m0.2251.3124919.8
TaTiWC3_3P3m10.1191.4221318.9
ScCP63_3/mmc0.60.9541018.5
TaWC2_2R-3m0.2112.0412916.3
MoRuB2_2Pmc21_10.091.2920215.6
RuCP-6m20.6491.0228815.1
RuVB2_2Pmc21_10.0731.0925515.0
Ni3_3AlCPm-3m0.1661.8611613.6
Nb4_4C3_3Pm-3m0.1670.9926713.2
Ta4_4C3_3Pm-3m0.1340.9920310.1
RhCF-43m0.5610.912289.3
MoCF-43m0.5850.971919.0
Pd3_3CaBPm-3m0.01.23967.0
RuSc3_3C4_4C2/m0.00.723216.6
Ta2_2S2_2CP-3m10.0070.692685.1
Mo3_3ZrB2_2C2_2Amm20.0460.642824.0
HfCP-6m20.7631.06563.2

Table III. DFPT calculated EPC results for MoB2_2-type instability (outside of Γ\Gamma), stabilized with electronic smearing of 0.10 Ry; SG is spacegroup, and ΔEh\Delta E_h is ΔEh\Delta E_h taken from MP database. Experimental results for some high-pressure TcT_c are also reported.

Figure 7. Structures and EPC properties of Ca5B3N6\mathrm{Ca}_{5}\mathrm{B}_{3}\mathrm{N}_{6}. (a)-(c): Crystal structures of Ca5B3N6\mathrm{Ca}_{5}\mathrm{B}_{3}\mathrm{N}_{6}, viewed from different directions (d)-(e): EPC results for compounds with stabilized imaginary phonon modes for Ca5B3N6\mathrm{Ca}_{5}\mathrm{B}_{3}\mathrm{N}_{6}. (d): EPC projected (highlighted by green circles) phonon band dispersions for dynamically unstable systems represented by small well at H-point (shown by red arrow), (e): Similar results for systems, stabilized with large electronic smearing of 0.06 Ry. (f) Eliashberg spectral function (g) Electronic band structure

Pressure-induced stabilization of phonons in MoB₂ and electron–phonon coupling calculations

Figure S4. Pressure-induced stabilization of phonons in MoB2_2 and electron–phonon coupling (EPC) calculations. (a) Energy (EE) versus volume (VV) equation of state for the two phases of MoB2_2: the high-pressure hP3 (P6/mmm) phase and the ambient pressure hR18 (R-3m) phase. (b) Enthalpy (H=E+PVH = E + PV) as a function of pressure (PP, GPa) for two phases referenced to the hP3 phase. The crossover between the two phases occurs near 75 GPa (indicated by the arrow). (c) Phonon dispersion of hP3 MoB2_2 at 75 GPa showing stabilization with EPC strength projected in green shading and (d) the corresponding Eliashberg spectral function. (e), (f) Same as (c), (d), but at 90 GPa.

Table II: Stabilization of imaginary phonon modes (around Γ\Gamma) utilizing Distortion (D), Pressure (P), and Smearing (S). The values of pressure and smearing are presented as P-value (GPa) and S-value (Ry) respectively. Crystal is distorted along the direction represented by eigenmode at Γ\Gamma and relaxed to obtain low-symmetry ground-state structures. Distorted GS is the difference between ground-state total energy of an undistorted and distorted structures at equilibrium. SG is spacegroup and ΔEh\Delta E_h represents the energy above the ground state hull.

CompoundSGΔEh\Delta E_h (eV/atom)[65]Distorted GS (meV/atom)EPC (λ\lambda)ωlog⁡\omega_{\log} (K)TcT_c (K)TcExptT_c^{Expt} (K)
Y2_2C3_3I-43d0.040.91.94 (P-10) 1.14 (S-0.1)175.23 (P-10) 227.97 (S-0.1)21.27 (P-10) 14.50 (S-0.1)18 [165]
La2_2C3_3I-43d0.02.21.15 (S-1)219.4 (S-1)14.3 (S-0.1)13.4 [166]
Sc2_2C3_3I-43d0.113.61.435 (P-30) 1.99 (S-0.1)303.18 (P-30) 205.6 (S-0.1)27.90 (P-30) 25.5 (S-0.1)- -
Al2_2Mo3_3CP41_1320.050.61.19 (S-0.1)174.95 (S-0.1)12.05 (S-0.1)9.2 [170]
YBCCmmm0.421500.73 (P-20)454.42 (P-20)10.15 (P-20)-
W2_2BI4/mcm0.00.60.81 (D)215.68 (D)6.61 (D)3.10 [29]
Mo2_2BI4/mcm0.033.60.79 (D)284.65 (D)8.12 (D)4.74 [29]
Ta2_2BI4/mcm0.038.90.44 (D)236.84 (D)0.34 (D)3.12 [29]
MoB4_4P6/mmm0.260.70.69 (D)418.42 (D)7.58 (D)-
La3_3InCPm-3m0.00.61.04 (D)108.63 (D)5.844 (D)2.6 [171]
La3_3InBPm-3m0.202.31.18 (D)83.03 (D)5.60 (D)10 [171]

Table 2.

Discussion

In this work, we have employed the machine learning (ML) models that utilize the data generated from ab-initio calculations using the DFPT and the isotropic Eliashberg approximation to iteratively guide the search for new phonon-mediated superconductors among B and C compounds. Our study also focuses on addressing the challenges encountered during DFPT calculations, such as convergence of Brillouin zone (BZ) sampling and the problem of calculated dynamic instability. To address the convergence issue, we developed an ansatz test to verify the convergence of superconductivity (SC) critical temperature TcT_c. This test uses the variation of TcT_c with respect to Gaussian broadening to compute the double delta summation. For dynamically unstable compounds, we applied large electronic smearing, lattice distortion, and pressure to stabilize imaginary phonon modes. We then calculated their EPC properties and incorporated these into the ML models. Between the two ML models, ALIGNN consistently outperforms CGCNN in predicting EPC properties especially after including the stabilized compounds with imaginary phonon and dynamical instability. Our ML-guided search demonstrates promising predictability for TcT_c values. For example, we predict SC in compounds with calculated dynamically stability such as TaNbC2_2 (28.4 K), Nb3_3B3_3C (16.4 K), Y2_2B3_3C2_2 (4.0 K), among others. In addition to studying dynamically stable compounds, we also focus on compounds with calculated dynamic instability, an area that, to our knowledge, has not been systematically explored before. We predicted SC in compounds showing dynamic instability such as Ca5_5B3_3N6_6 (35 K), Pd3_3CaB (7.0 K), some ternary Ru compounds,

TABLE III: DFPT calculated EPC results for MoB2_2-type instability (outside of Γ\Gamma), stabilized with electronic smearing of 0.10 Ry; SG is spacegroup, and ΔEh\Delta E_h is ΔEh\Delta E_h taken from MP database. Experimental results for some high-pressure TcT_c are also reported.

MoRuB2_2 (15.6 K), RuVB2_2 (15.0 K), Pd3_3CaB (7.0 K) and RuSc3_3C4_4 (6.6 K) with ΔEh\Delta E_h mostly below 0.1 eV/atom. With further refinement and larger dataset, our workflow can be improved in accurately predicting more SC compounds. In this regard, identifying metastable compounds with calculated dynamic instability, where soft phonon exhibit significant EPC contribution, plays a crucial role.

METHODS

Convergence with respect to Brillouin zone sampling (k-point mesh)

To identify cases of convergence failure (i.e., incorrect prediction of SC/NSC) related to Brillouin zone sampling, we analyzed the variation of TcT_c with respect to Gaussian broadening (σ\sigma) in the double delta integration and developed a simple ansatz based on the converged results of MgB2_2, as described in the “Convergence tests for MgB2_2 and AlB2_2 (Supplementary Figures S5–S10, Supplementary Tables S5–S6)” and ”Convergence ansatz (Supplementary Figures S11–S14, Supplementary Tables S7–S9)” sections of the Supplemental Material (SM). This ansatz involves extracting TcT_c, similar to MgB2_2, and estimating the decay parameter (A) in the exponential variation of TcT_c with σ\sigma,

Tc=exp⁡(−Aσ1/3+B)(2)T_c = \exp(-A\sigma^{1/3} + B) \tag*{(2)}

where A is the variable that quantifies the rate of exponential decay, and B is the constant associated with Debye temperature. Unconverged results show larger values of A, which decrease with denser k-meshes. With larger k- or qq-grids, the TcT_{\mathrm{c}} vs σ\sigma curve becomes less steep. Our convergence analysis of MgB2\mathrm{MgB}_{2} suggests that AMgB2=12A_{\mathrm{MgB}_{2}} = 12–1313 can be used as a threshold for this study: calculations are considered as unconverged for A>AMgB2A > A_{\mathrm{MgB}_{2}}, requiring a denser kk-mesh, while those with A<AMgB2A < A_{\mathrm{MgB}_{2}} can be regarded as converged for accurate prediction.

Computational details

Data extraction from the MP database, as well as input preparation for ground state calculations, calculation submission, result extraction and input preparation for ML studies, and plotting, were performed using the high-throughput electronic structure package (HTESP) [157]. A script, ‘fitting elph smearing.py‘, is included in the HTESP package to compute the decay parameter from TcT_{\mathrm{c}} vs. broadening σ\sigma data. Ground-state DFT and EPC calculations were performed using the QE code [173, 174]. Ultrasoft or norm-conserving pseudopotentials (PP) from the efficiency standard solid-state pseudopotentials (SSSP) dataset [175] were employed, with the replacement of projector-augmented wave (PAW) PPs by GBRV ultrasoft norm-conserving high-throughput PPs [176]. The exchange-correlation energy was approximated using the Perdew-Burke-Ernzerhof (PBE) generalized-gradient approximation (GGA) [177]. The Brillouin zone was sampled using a kk-point mesh from the MP database to compute the ground-state charge density for EPC calculations. The qq-point mesh required for EPC calculations was obtained by halving the kk-point mesh (q=k/2q = k/2), with odd kk-points changed to even by adding 1 to it. The structures were fully relaxed using Broyden–Fletcher–Goldfarb–Shanno (BFGS) minimization [178] until the total energy and forces for ionic minimization converged within 10−510^{-5} Ry and 10−410^{-4} Ry/Bohr, respectively. The self-consistent electronic energy and charge density were minimized with a convergence threshold of 10−1210^{-12} Ry. Similarly, the SCF convergence for phonon calculations is achieved with an energy cutoff of 10−1410^{-14}. Starting from the default αmix\alpha_{\mathrm{mix}} value of 0.7, it is reduced to 0.3 if the EPC calculation does not converge. A fine kk-point grid, twice the size of the kk-point mesh used for charge density convergence, was used for interpolating EPC matrix elements to compute double-delta integration and the λ\lambda. Gaussian smearing of 0.02 Ry was applied for charge-density optimization. To compute λ\lambda for various σ\sigma values, we computed double-delta integration using 10 broadening (σ\sigma) values ranging from 0.005 to 0.05 Ry for the kk-mesh. For the qq-mesh integration, a fixed smearing of 0.5 meV was employed. The reported results were obtained for σ=0.01\sigma= 0.01 Ry with a Coulomb potential μc∗\mu_{\mathrm{c}}^{*} of 0.16.

Limitations

When evaluating the significance of machine-learning (ML) models trained on ab-initio data, it is essential to recognize both their capabilities and inherent limitations. First, the performance of any ML model is strongly dependent on the quality, diversity, and representativeness of its training dataset. Limited or biased datasets, which are common in computational materials science, can restrict a model’s predictive accuracy and transferability. Second, ML models are susceptible to overfitting, particularly when trained on small or noisy datasets, although careful validation procedures during model development can mitigate this issue to some extent. Third, the fidelity of ML predictions is ultimately bounded by the accuracy of the underlying ab-initio methods used to generate the training data. For example, the widely used isotropic Eliashberg approximation in superconductivity calculations neglects the inherently anisotropic and multi-band character of electron–phonon coupling. Although fully solving the anisotropic Migdal-Eliashberg equations (e.g., using EPW) yield more accurate estimates of superconducting properties, they are not computationally practical for large scale data generation. Consequently, ML models trained on data obtained from the isotropic Eliashberg approximation can function as efficient screening tools to identify promising candidate materials, which may then be examined with more accurate, anisotropic, and computationally demanding methods.

Data availability

Data computed and utilized in this study are tabulated in Supplementary Tables S10–S17, Tables. II and III. The raw data can be downloaded from the Figshare platform [179].

Acknowledgements

We thank Dr. Paul C. Canfield for the funding support, initiating the idea of searching for new phonon-mediated superconductors among boron and carbon compounds, and the helpful discussion throughout the project.

Funding

This work was supported by Ames National Laboratory LDRD and U.S. Department of Energy, Office of Basic Energy Science, Division of Materials Sciences and Engineering. Ames National Laboratory is operated for the U.S. Department of Energy by Iowa State University under Contract No. DE-AC02-07CH11358.

Author information

Authors and Affiliations

Ames National Laboratory, Ames, Iowa 50011, USA

Niraj K. Nepal, Lin-Lin Wang

Department of Physics and Astronomy, Iowa State University, Ames, Iowa 50011, USA

Lin-Lin Wang

Contributions

L.-L.W. conceived and supervised the work. N.K.N. and L.-L.W. designed and performed the high throughput calculations with the machine learning guided approach. N.K.N. developed the ansatz to test the convergence of TcT_c calculation. All authors discussed the results and contributed to the final manuscript.

Corresponding author

Correspondence to Niraj K. Nepal ([email protected])

Competing interests

The authors declare no competing interests.

References

Statistics for compounds with unknown TcT_c

In Figure S1, we present the descriptions of 268 dynamically stable compounds for which the experimental TcT_c is not known. Figure S1-(Figure S1) illustrate the distribution of different elements in these systems, the allocation of B/C/B+C compounds, and the deviation of their formation energy (ΔEh\Delta E_h) from stable structures, respectively. Most unknown compounds are associated with transition metals (TM) such as Nb, Ta, Mo, V, Y, and others. Approximately 80% of the structures are close to the ground state convex hull with ΔEh≤0.05\Delta E_h \leq0.05 eV/atom, while the remaining 20% have ΔEh\Delta E_h greater than 0.05 eV/atom. Panel S1(d) showcases the distributions of these compounds based on their space groups. In Figure S1 and (d), each bar is partitioned into segments, with the segments colored in red representing the number of SC, while the segments in blue denote the NSC. The overall description is similar to that of known compounds, as shown in Figure 2 of the main text.

Learning curves for various training and validation during the ALIGNN model training process.

Figure S3. Learning curves for various training and validation during the ALIGNN model training process.

Training sizeCGCNNALIGNN
λ\lambdaωlog\omega_{log} (K)TcT_c (K)Tc′T'_c (K)λ\lambdaωlog\omega_{log} (K)TcT_c (K)Tc′T'_c (K)
250 (Run1)0.15721.81.30.15561.72.2
323 (Run2)0.27856.45.50.16833.03.3
351 (Run3)0.31703.96.50.24603.04.4

Table S1. Mean Absolute Errors (MAEs) for the 10% test set in each training iteration with different sizes running for 3000 epochs.

ModelEPCEpochBest MAER2R^2-score
CGCNN-250λ\lambda480.1450.75
CGCNN-250ωlog⁡\omega_{\log}60490.45
CGCNN-250TcT_c132.800.56
CGCNN-323λ\lambda11900.2030.66
CGCNN-323ωlog⁡\omega_{\log}590570.5
CGCNN-323TcT_c7362.730.31
CGCNN-351λ\lambda640.2040.62
CGCNN-351ωlog⁡\omega_{\log}62510.51
CGCNN-351TcT_c463.340.263
ALIGNN-250λ\lambda6710.090.58
ALIGNN-250ωlog⁡\omega_{\log}672270.47
ALIGNN-250TcT_c7601.320.23
ALIGNN-323λ\lambda4830.1960.64
ALIGNN-323ωlog⁡\omega_{\log}293570.63
ALIGNN-323TcT_c12752.330.32
ALIGNN-351λ\lambda1860.220.72
ALIGNN-351ωlog⁡\omega_{\log}701530.77
ALIGNN-351TcT_c6172.850.38

Table S2. The best validation MAEs and their corresponding epochs during the training process are reported. CGCNN-250 represents the CGCNN model trained with 250 compounds and so on. Additionally, the coefficient of determination (R2R^2-score) is also provided for a linear-fit YML=slope×YDFPTY^{ML} = slope \times Y^{DFPT}, Y=(λ,ωlog⁡,Tc)Y = (\lambda, \omega_{\log}, T_c). A test set of 58 compounds is used for the model trained with 250 and 323 compounds, while 66 compounds are used for the model trained with 351 compounds. The R2R^2-score with higher values close to 1 indicating better accuracy, while a score of 0 suggests no better performance than predicting the mean.

Training/test sizePropertyTc≤1T_c \le1Tc>1T_c > 1
CGCNNALIGNNCGCNNALIGNN
323/58λ\lambda0.280.210.200.32
323/58ωlog⁡\omega_{\log}5654188111
323/58TcT_c3.601.982.234.30
351/66λ\lambda0.390.200.270.39
351/66ωlog⁡\omega_{\log}125996963
351/66TcT_c4.671.976.235.95

Table S3. MAE calculated separately for nonsuperconductors and superconductors. The test set of 58 includes 40 nonsuperconductors and 18 superconductors, while the test set of 66 consists of 41 nonsuperconductors and 25 superconductors, among which 7 are stabilized superconductors.

Classification models

In addition to regression, we also employ machine learning models for binary classification of materials into two categories: SC or NSC. We utilized the similar settings (batch, epochs, etc in config.json) as that of regression model training. Instead of finding a threshold TcT_c that maximizes the accuracy metrics, we simply perform the classification ML for a fix threshold TcT_c (0.5, 1.0, 2.0 K) and assess the model performance. The accuracy metrics used in our evaluation are presented in Table S4, and they are defined as follows:

TcThresholdT_c^{\mathrm{Threshold}} (K)Run 1Run 3
AccuracyPrecisionRecallF1-scoreAccuracyPrecisionRecallF1-score
0.50.67/0.590.83/0.580.57/0.500.68/0.540.71/0.670.74/0.670.68/0.640.71/0.66
1.00.66/0.620.83/0.720.47/0.430.60/0.540.64/0.670.84/0.720.51/0.550.64/0.62
2.00.72/0.620.77/0.770.43/0.340.56/0.480.69/0.610.9/0.70.5/0.410.64/0.52

Table S4. Classification accuracy: Results from ML model are presented in the form CGCNN/ALIGNN.

Accuracy=TP+TNN(3)\mathrm{Accuracy} = \frac{T_P + T_N}{N} \tag*{(3)}
Precision=TPTP+FP(4)\mathrm{Precision} = \frac{T_P}{T_P + F_P} \tag*{(4)}
Recall=TPTP+FN(5)\mathrm{Recall} = \frac{T_P}{T_P + F_N} \tag*{(5)}
F1-score=2(Precision×Recall)precision+Recall(6)\mathrm{F1\text{-}score} = \frac{2(\mathrm{Precision} \times\mathrm{Recall})}{\mathrm{precision} + \mathrm{Recall}} \tag*{(6)}

Accuracy, Precision, Recall, and F1-score are performance metrics that evaluate different aspects of a classifier’s effectiveness. Accuracy measures the overall correctness of the classifier’s predictions. Precision quantifies the correctness of positive predictions, while Recall (sensitivity) assesses the successful identification of positive instances. The F1-score is the harmonic mean of Precision and Recall, providing a balance between the two. TPT_P, TNT_N, FPF_P, and FNF_N represent true positive (correctly identifying true superconductors), true negative (correctly identifying true nonsuperconductors), false positive (misidentifying nonsuperconductors as superconductors), and false negative (misidentifying superconductors as nonsuperconductors), respectively. The accuracy metrics of classification models for both Run 1 and Run 3 are displayed in Table S4. Across all these metrics, the overall accuracies are reasonable, indicating a useful classification model despite the small dataset. By employing a classification threshold of Tc=1T_c = 1 K, both ML models exhibit comparable performance. Notably, one can observe the enhanced accuracy of Run 3 in contrast to Run 1. This improvement might be attributed to the expanded training dataset size and the balanced representation of both SC and NSC instances after adding stabilized high TcT_c results. The ML model, when using a TcT_c threshold of 1 K, achieves an overall accuracy ranging from 60% to 70%, precision ranging from 70% to 90%, and an F1-score ranging between 50% and 60%. The ML models demonstrate lower recall values, ranging from 50% to 55%. Considering the data imbalance between the number of SC instances (approximately 40%) and NSC instances (approximately 60%), these recall values are reasonable.

Theory of superconductivity: Isotropic Approximation

We have adopted the DFPT calculation with isotropic Eliashberg approximation to compute SC properties, which provides a harmony between accuracy and efficiency. The critical temperature can be calculated using the Allen-Dynes formula [1],

Tc=ωlog⁡1.2exp⁡[−1.04(1+λ)λ−μc∗(1+0.62λ)],(7)T_c = \frac{\omega_{\log}}{1.2} \exp\left[-\frac{1.04(1+\lambda)}{\lambda-\mu_c^*(1+0.62\lambda)}\right], \tag*{(7)}

where, λ=2∫0∞dωωα2F(ω)\lambda= 2\int_0^\infty\frac{d\omega}{\omega}\alpha^2F(\omega) is the EPC strength constant, ωlog⁡=exp⁡[2λ∫0∞dωωα2F(ω)log⁡ω]\omega_{\log} = \exp\left[\frac{2}{\lambda}\int_0^\infty\frac{d\omega}{\omega}\alpha^2F(\omega)\log\omega\right], α2F(ω)\alpha^2F(\omega) is frequency (ω)(\omega) resolved Eliashberg spectral function, and μc∗\mu_c^* is the Coulomb potential. The spectral function α2F(ω)\alpha^2F(\omega) is defined as,

α2F(ω)=12∑ν∫BZdqΩBZωqνλqνδ(ω−ωqν).(8)\alpha^2F(\omega) = \frac{1}{2}\sum_{\nu}\int_{BZ}\frac{dq}{\Omega_{BZ}}\omega_{q\nu}\lambda_{q\nu}\delta(\omega-\omega_{q\nu}). \tag*{(8)}

Here, ΩBZ\Omega_{BZ} is the volume over the Brillouin zone ∫BZdqΩBZ→1Nq∑q\int_{BZ}\frac{dq}{\Omega_{BZ}} \to\frac{1}{N_q}\sum_q, ωqν\omega_{q\nu} is the mode (ν)(\nu) resolved phonon frequency, and λqν\lambda_{q\nu} is the mode resolved EPC strength constant,

λqν=1N(ϵF)ωqν∑mn∫BZdkΩBZ∣gmn,ν(k,q)∣2δ(ϵnk−ϵF)δ(ϵmk+q−ϵF).(9)\lambda_{q\nu} = \frac{1}{N(\epsilon_F)\omega_{q\nu}}\sum_{mn}\int_{BZ}\frac{dk}{\Omega_{BZ}}|g_{mn,\nu}(\mathbf{k},\mathbf{q})|^2\delta(\epsilon_{nk}-\epsilon_F)\delta(\epsilon_{mk+\mathbf{q}}-\epsilon_F). \tag*{(9)}

N(ϵF)N(\epsilon_F) is the density of states at the Fermi level ϵF\epsilon_F, and gmn,ν(k,q)g_{mn,\nu}(\mathbf{k},\mathbf{q}) is the EPC matrix element which quantifies the scattering process between Kohn-Sham states k+q\mathbf{k}+\mathbf{q} and k\mathbf{k}. In Ref. [2], an effective approach to approximate double delta integration is explored, specifically tailored for situations where it is permissible to disregard the dependence on the momentum vector (q)(\mathbf{q}). This method finds utility in scenarios such as the study of extensive molecular systems like alkali fullerides, where momentum dependence can be safely overlooked. The net EPC strength constant is computed as

λ=∑qνλqν(10)\lambda= \sum_{q\nu} \lambda_{q\nu} \tag*{(10)}

For computational feasibility, these Dirac deltas can be approximated by Gaussian functions with a broadening parameter σ\sigma [3], and EPC strength is given by Eqs. (9) and (10) can be redefined as

λ≈1N(ϵF)NqNk∑nm∑q∑k∣gmn,ν(k,q)∣2ωqν12πσ2exp⁡[−(ϵnk−ϵF)2+(ϵmk+q−ϵF)2σ2].(11)\lambda\approx\frac{1}{N(\epsilon_F)N_qN_k}\sum_{nm}\sum_{\mathbf{q}}\sum_{\mathbf{k}}\frac{\lvert g_{mn,\nu}(\mathbf{k},\mathbf{q})\rvert^2}{\omega_{q\nu}}\frac{1}{2\pi\sigma^2}\exp\left[-\frac{(\epsilon_{n\mathbf{k}}-\epsilon_F)^2+(\epsilon_{m\mathbf{k}+\mathbf{q}}-\epsilon_F)^2}{\sigma^2}\right]. \tag*{(11)}

Here, NqN_q and NkN_k respectively are the total number of q\mathbf{q} and k\mathbf{k} grid points, and σ\sigma is the smearing used to broaden states at the Fermi-level ϵF\epsilon_F. With infinitely large k\mathbf{k}-grids, and σ→0\sigma\to0, the double summation changes back to double delta integration. Moreover, λ\lambda can also obtained from frequency ω\omega resolved Eliashberg spectral function as [4]

λ=2∫dωα2F(ω)ω=N(ϵF)⟨g2⟩M⟨ω2⟩,(12)\lambda= 2\int\frac{d\omega\alpha^2F(\omega)}{\omega} = \frac{N(\epsilon_F)\langle g^2\rangle}{M\langle\omega^2\rangle}, \tag*{(12)}

where, ⟨ω2⟩\langle\omega^2\rangle average of the square of the phonon frequency (∫dωω2F(ω)∫dωωα2F(ω))\left(\frac{\int d\omega\omega^2F(\omega)}{\int\frac{d\omega}{\omega}\alpha^2F(\omega)}\right), ⟨g2⟩\langle g^2\rangle is average over the Fermi surface of the square of electronic phonon coupling matrix element [4].

Convergence tests for MgB2 and AlB2

The ground-state total energy and phonon frequencies at Γ\Gamma-point are well converged for MgB2_2 with k\mathbf{k}-grids of 8×8×68\times8\times6, compared to the dense grid of 24×24×2424\times24\times24. K-mesh grid in materials project database for MgB2_2 is 8×8×78\times8\times7. Since the superconducting properties of MgB2_2 with k\mathbf{k}-grid of 8×8×88\times8\times8 doesn’t follow the trend of converged result of denser k-mesh with σ→0\sigma\to0, we chose 8×8×68\times8\times6 so that we can use sufficient q\mathbf{q}-grids of 4×4×34\times4\times3 instead of using 9×9×99\times9\times9 k\mathbf{k}-grid and 3×3×33\times3\times3 q\mathbf{q}-grid. We utilized 8×8×68\times8\times6 k-grid and its multiple to compute the decay parameter. In this work, we are taking λ=0.75\lambda= 0.75 obtained from solving anisotropic Migdal-Eliashberg equation using μ∗=0.16\mu^* = 0.16 as a reference [Comput. Phys. Commun. 209, 116 (2016)]. One can obtain this converged λ\lambda with denser k- and q- grids and σ→0\sigma\to0, as shown in Figure S5 − Figure S7. At last, we check whether the result is transferable to systems with larger unit cell where essentially smaller k-grids provide converged ground-state properties and q-grid, taken half of k-grid, sometime shrinks only to include the Γ\Gamma-point. To do that, we performed EPC calculations with 2×2×22\times2\times2 supercell of MgB2_2 which has 24 atoms per cell, and present the results in Figure S8 and Figure S9. Results confirm that q-grid as half as that of the k-grid can provide converged results even for larger systems. Instead of increasing coarse k- and q- grids, one can also utilize extremely large k- grid for interpolating EPC matrix element to achieve convergence, as shown in Figure S7 and Figure S10 respectively for MgB2_2 and AlB2_2. However, it also increases computational complexity for larger systems.

Figure S5. Convergence test for MgB2_2 results with respect to q-mesh. Unit of Tc_c and ωlog⁡\omega_{\log} is Kelvin (K); As q becomes denser, convergence can be achieved across the σ→0\sigma\to0, with less exponential decay. Fine k- grid is utilized for interpolation.

Figure S7. Convergence test for MgB2_2 results with respect to fine kk-mesh used for interpolation. Unit of TcT_c and ωlog⁡\omega_{\log} is Kelvin (K); As fine kk becomes denser, convergence can be achieved across the σ→0\sigma\to0, with less exponential decay.

Figure S8. Convergence test for 2×2×22 \times2 \times2 supercell of MgB2\mathrm{MgB}_{2} results with respect to q-mesh. Unit of TcT_{\mathrm{c}} and ωlog⁡\omega_{\log} is Kelvin (K); As q becomes denser, convergence can be achieved across the σ→0\sigma\to0, with less exponential decay. Fine k-grid is utilized for interpolation.

Figure S9. Convergence test for 2×2×22 \times2 \times2 supercell of MgB2\mathrm{MgB}_{2} results with respect to k-mesh. Unit of TcT_{\mathrm{c}} and ωlog⁡\omega_{\log} is Kelvin (K); As k becomes denser, convergence can be achieved across the σ→0\sigma\to0, with less exponential decay. Fine k-grid is utilized for interpolation.

Figure S10. Convergence test for AlB2_2 results with respect to various k\mathbf{k}- and q\mathbf{q}-grids. Unit of TcT_{\mathrm{c}} and ωlog⁡\omega_{\log} is Kelvin (K); As grids becomes denser, convergence can be achieved across the σ→0\sigma\to0, with less exponential decay. Separate results corresponding to k\mathbf{k}- and q\mathbf{q}- grids respectively of 12×12×1212 \times12 \times12 and 6×6×66 \times6 \times6 is represented by “star (*)” symbol.

K-meshTotal Energy (eV/atom)
6×6×4-615.178
8×8×6-615.179
8×8×8-615.180
9×9×9-615.180
12×12×12-615.180
16×16×16-615.180
24×24×24-615.180

Table S5. Convergence of the ground-state total energy with respect to K-point mesh

K-meshω1\omega_1ω2\omega_2ω3\omega_3ω4\omega_4ω5\omega_5ω6\omega_6ω7\omega_7ω8\omega_8ω9\omega_9
6×6×46\times6\times4-0.654-0.6540.34310.1810.1812.0720.9624.3824.38
8×8×68\times8\times6-0.49-0.490.4810.1110.1112.1116.4516.4520.75
8×8×88\times8\times8-0.63-0.630.2810.110.112.1117.4317.4320.8
9×9×99\times9\times9-0.62-0.62-0.339.999.9912.0814.3614.3620.65
12×12×1212\times12\times12-0.64-0.64-0.2310.0610.0612.1217.1317.1320.69
16×16×1616\times16\times16-0.64-0.64-0.2510.0510.0512.117.217.220.69
24×24×2424\times24\times24-0.64-0.64-0.2210.0510.0512.116.9916.9920.71

Table S6. Convergence of phonon frequency (THz) at Γ\Gamma-point with respect to K-point mesh

Figure S6. Convergence test for MgB2_2 results with respect to kk-mesh. Unit of TcT_c and ωlog⁡\omega_{\log} is Kelvin (K); As kk becomes denser, convergence can be achieved across the σ→0\sigma\to0, with less exponential decay. Fine kk-grid is utilized for interpolation.

Convergence ansatz

Estimation of SC properties requires the computation of double-delta integration, which also defines the nesting function, around the Fermi level EFE_{\mathrm{F}} over the entire Brillouin zone [Eq. 3]. However, it has a slow convergence with respect to k\mathbf{k}- and q\mathbf{q}- grids. In principle, it requires infinitely dense grids, which makes the computation exorbitantly expensive. Therefore, Gaussian broadening technique [Eq. 5] is employed to compute λ\lambda with finite k\mathbf{k}-mesh. To compute TcT_{\mathrm{c}}, we employed the DFPT calculation with the isotropic Eliashberg approximation. We used a coarse k\mathbf{k}-grid obtained from the MP database, which provides reasonably accurate ground-state properties. The q\mathbf{q}-grid was set to half the size of the k\mathbf{k}-grid, and a broadening parameter (σ\sigma) of 0.01 Ry was utilized with the coarse MP grid, which accurately captures a converged electron-phonon coupling constant (λ\lambda) of MgB2_2 computed using the anisotropic Migdal-Eliashberg equations [5]. However, in practice, ∣gmn,ν(k,q)∣2\lvert g_{mn,\nu}(\mathbf{k}, \mathbf{q}) \rvert^2 is computed for a reasonable coarse k\mathbf{k}- and q\mathbf{q}- grids and interpolated the matrix elements to fine k\mathbf{k}- for any q\mathbf{q}-point, from 32×32×3232 \times32 \times32 to as high as 60×60×6060 \times60 \times60, to achieve numerical convergence, as implemented in Quantum Espresso (QE) code[3]. Furthermore, an efficient interpolation scheme utilizing both dense k\mathbf{k}- as well as q\mathbf{q}-grids has been implemented in EPW package that uses wannier orbitals [5]. However, employing extremely fine grids for interpolation can increase the computational complexity, which is not suitable for highthroughput calculations. Therefore, we have restricted ourselves for choosing fine k\mathbf{k}-grid only twice of corresponding coarse grid for highthroughput calculations.

Our investigation revealed that a reasonable number of calculations did not converge, leading to inaccurate predictions, while utilizing a coarse k\mathbf{k}-grid from MP database and fine grid for interpolation only twice that of the k\mathbf{k}-grid [Table S7]. For example, AlB2_2 was predicted to be a superconductor (Tc∼11T_{\mathrm{c}} \sim11 K) with the coarse grid [k-grid: 8×8×88 \times8 \times8, q-grid: 4×4×44 \times4 \times4, fine-k-mesh: 16×16×1616 \times16 \times16] from MP database, whereas a denser grid twice the size of the coarse grid corrected this inaccuracy and predicted AlB2_2 to be a nonsuperconductor, consistent with experimental observations [Fig. S7]. To identify such cases of convergence failure, we examined the variation of TcT_{\mathrm{c}} with respect to σ\sigma and developed a simple ansatz based on the converged results of MgB2_2, as depicted in Figure S13. The ansatz involves extracting TcT_{\mathrm{c}} with a fixed broadening value (as in the case of MgB2_2) and estimating the decay parameter (AA) in the exponential variation of TcT_{\mathrm{c}} with σ\sigma as Tc∼exp⁡(−Aσ1/3+B)T_{\mathrm{c}} \sim\exp(-A\sigma^{1/3}+B). Unconverged results exhibit larger values of AA, which decrease with a denser k-mesh.

Figure S13. Schematics of the convergence testing procedure explained in the main text. AcA_{\mathrm{c}} is the value of the decay parameter computed for MgB2_2. First, we compute TcT_{\mathrm{c}} for different values of σ\sigma. If the variation of TcT_{\mathrm{c}} with σ\sigma decreases exponentially, we compute AA. If A<AcA < A_{\mathrm{c}}, then the results have converged. If the variation increases exponentially and A<0A < 0, satisfying A<AcA < A_{\mathrm{c}}, the results are also considered converged. If A>AcA > A_{\mathrm{c}} or if the variation follows a different pattern, the calculations are repeated with a denser k\mathbf{k}-mesh. If TcT_{\mathrm{c}} decreases extremely rapidly with increasing σ\sigma, then the material is considered a non-superconductor.

The detail theory of EPC calculations within DFPT formalism in presented in the section “Theory of superconductivity: isotropic approximation”. The variation of density of states at the Fermi level (N(ϵF)N(\epsilon_F)) with respect to smearing σ\sigma, depends on Gaussian,

G(Δ,σ)=1σ2exp⁡[−(ϵnk−ϵF)2+(ϵmk+q−ϵF)2σ2]∼1σ2exp⁡[−Δ2σ2](13)G(\Delta,\sigma)=\frac{1}{\sigma^2}\exp\left[-\frac{(\epsilon_{nk}-\epsilon_F)^2+(\epsilon_{mk+q}-\epsilon_F)^2}{\sigma^2}\right]\sim\frac{1}{\sigma^2}\exp\left[-\frac{\Delta^2}{\sigma^2}\right] \tag*{(13)}

with Δ2∼(ϵnk−ϵF)2+(ϵmk+q−ϵF)2\Delta^2\sim(\epsilon_{nk}-\epsilon_F)^2+(\epsilon_{mk+q}-\epsilon_F)^2 also depends on σ\sigma. Despite of dependency of λ\lambda on EPC matrix (⟨g2⟩\langle g^2\rangle) and phonon frequency (⟨ω2⟩\langle\omega^2\rangle) terms, the variation in λ\lambda is largely guided by the variation in N(ϵF)N(\epsilon_F) or the Gaussian term G. Figure S11(a) shows the variation of G with respect to smearing σ\sigma for different values of Δ\Delta, assuming Δ\Delta independent of σ\sigma. The effect of σ\sigma on G dictates the variation of λ\lambda and hence TcT_c, with Δ\Delta depends on the materials as well as on size of the grid to compute double delta integration. A larger Δ\Delta represents more states contributing towards the double delta summation within space around the Fermi-level spanned by σ\sigma. In other word, Δ\Delta will increase with σ\sigma if the density of states has local maximum close to the Fermi-level, and the variation of EPC properties will be similar to the Gaussian plots corresponding to Δ≥0.05\Delta\ge0.05. For Niobium, the variation of EPC properties with respect to σ\sigma follows similar to blue-dashed curve corresponding to Δ=0.01\Delta=0.01 [3], whereas for MgB2_2, the variation follows the Gaussian corresponding to Δ<0.01\Delta<0.01. For σ→0\sigma\to0 and Δ→0\Delta\to0, if Δσ≤1\frac{\Delta}{\sigma}\le1 the Gaussian function recovers the delta function, while for Δσ≥1\frac{\Delta}{\sigma}\ge1 the Gaussian drops to zero similar to Δ≥0.01\Delta\ge0.01 cases. Our primary focus is more on exponentially decaying cases.

Figure S11. Decay Constant Determination from Converged MgB2_2 data. (a) Variation of G(Δ,σ)Gmax⁡\frac{G(\Delta,\sigma)}{G_{\max}} with respect to σ\sigma for different values of Δ\Delta; Gmax⁡G_{\max} being the maximum value of Gaussian for corresponding Δ\Delta within the value of σ\sigma from 0.005 Ry to 0.05 Ry. (b) TcT_{\mathrm{c}} vs σ\sigma for various k- grids for MgB2_2. k-q-2k represents k-mesh for self-consistent calculation for charge density and EPC, q-mesh for phonon and EPC, and 2k-mesh represents fine grid used for interpolating EPC matrix obtained from k and q-grids; k and k′ respectively are 6×6×46 \times6 \times4 and 8×8×68 \times8 \times6. (c) log⁡Tc\log T_{\mathrm{c}} vs σ1/3\sigma^{1/3} plot; Linear-fit is performed for data up to σ=0.05\sigma= 0.05 Rydberg (denoted by vertical dashed line), after which lines change slope; α=1/3\alpha= 1/3 fits with an optimal coefficient of determination (R2^2-score) (Table inside plot); Fitting parameter “A” decreases from 19 to 10 as k- grid change from coarse to dense; (c) TcT_{\mathrm{c}} vs σ\sigma for various q- grids keeping k-grid fixed at 12×12×1212 \times12 \times12. (d) Spectral function α2F(ω)\alpha^2F(\omega) vs ω\omega plots with σ=0.01\sigma= 0.01 Ry with k-grid of 8×8×68 \times8 \times6.

The effect of smearing σ\sigma on ωlog⁡\omega_{\log} is insignificant compared to N(ϵF)N(\epsilon_F) or λ\lambda [2]. According to the Bardeen–Cooper–Schrieffer (BCS) theory [6],

Tc∼ΘDexp⁡(−1/λ)∼ΘDexp⁡(−1N(ϵF)U)(14)T_c\sim\Theta_D\exp(-1/\lambda)\sim\Theta_D\exp\left(-\frac{1}{N(\epsilon_F)U}\right) \tag*{(14)}

where ΘD\Theta_D is the Debye cutoff energy and U is electron-phonon coupling potential. One can established the relation between λ\lambda or TcT_c on σ\sigma as λ∼1/σα\lambda\sim1/\sigma^\alpha and Tc∼ΘDexp⁡(−1/λ)=exp⁡(−Aσα+B)T_c\sim\Theta_D\exp(-1/\lambda)=\exp(-A\sigma^\alpha+B) with ΘD∼exp⁡(B)×\Theta_D\sim\exp(B)\times constant. Here parameter α\alpha is a constant with a positive value if TcT_c decreases with increasing σ\sigma and becomes negative otherwise. Here A and B are constants to be determined with A denoting the coefficient of exponential increase (for negative A) or decrease (for positive A) of TcT_c with respect to σ\sigma. To determine the value of these constants, we perform linear-fit of log⁡Tc\log T_c vs σα\sigma^\alpha for different k-grids for MgB2_2, which is frequently studied both theoretically as well as experimentally and often challenging to obtain the converged SC properties [7–9]. Furthermore, one can capture the behavior of TcT_c vs σ\sigma for a wide range of TcT_c using MgB2_2, as shown in Figure S11(b). We utilize a smaller k-mesh of 6×6×4 and a slightly larger k′-grid of 8×8×6 for comparison [Figure S11(b) and (c)].

Fig. S11 (b) represents the variation of TcT_c with respect to σ\sigma, while Fig. S11 (c) presents log⁡Tc\log T_c vs σα\sigma^\alpha plot. It exhibits nearly linear behavior up to σ=0.05\sigma=0.05 Ry, i.e. σ1/3=0.37\sigma^{1/3}=0.37, after which it deviates from linearity. The linear-fit has optimal coefficient of determination score (R2^2-score) for α<0.5\alpha<0.5 on log⁡Tc\log T_c vs σα\sigma^\alpha data for k(charge density and EPC)-q(Phonon and EPC)-2k(interpolating EPC matrix) grids (Fig. S11 (c)) [Table within the Fig. S11(c)]. Therefore, we chose α=1/3\alpha=1/3 in this work with R2^2-score of 0.9987. Besides different choice of α\alpha leads to different values of A and B, it doesn’t have a significant role. For coarse grid, the TcT_c has a larger dependency on σ\sigma and shows larger exponential decay, compared to more converged calculations on denser grids. The exponential decay parameter “A” decreases from 19 to 10 with grids changing from coarse to dense one. This analysis suggests that AMgB2=12A_{MgB_2}=12–13 can be used as a cutoff for this work, the calculations can be considered unconverged for A>AMgB2A>A_{MgB_2}, while the calculations can be considered converged for A<AMgB2A<A_{MgB_2} for accurate predictions. Fig. S11(d) shows the convergence of the TcT_c with respect to q-point mesh, keeping k-grid fixed at 12×12×12. A q-point grid as half as that of the ground-state k-point grid is sufficient to provide the converged results for MgB2_2. Fig. S11 (e) represents the spectral functions for various grids with k=8×8×6. Denser grid (2k) slightly blue shifts the spectral function peaks at lower frequency range (15.5–16.5 THz), while peaks at higher frequency ranges (20–22.5 THz) remain unaffected. This results slight change in λ\lambda from 0.78 to 0.72 and TcT_c changing from 20 K to 15–16 K [Figs. S5-S7]. A λ\lambda of 0.78 agrees with previous theoretical value of ∼0.75\sim0.75 from anisotropic Migdal-Eliashberg calculation using μ∗=0.16\mu^*=0.16 [5], λ=0.71\lambda=0.71 from fully anisotropic SCDFT [10], and λ=0.73\lambda=0.73 from previous isotropic Eliashberg approximation [11] at slightly larger σ=0.015\sigma=0.015 Ry. This indicates that k-grid obtained from MP database with σ=0.01\sigma=0.01 Ry already provide converged results in the case of MgB2_2. However, it is not always the case for other materials. Based on these results, parameters σ=0.01\sigma=0.01 Ry with μc∗=0.16\mu_c^*=0.16 seems to be reasonable choice for smearing with 8×8×6 to 16×16×12 k-grids for MgB2_2, and for other systems for the sake of comparison. These parameters could also depend on pseudopotentials. In order to achieve converged results, it is necessary to use denser k- and q-grids. Subsequently, a double-delta integration should be performed, selecting the results that correspond to the limit of σ→0\sigma\to0. Other details convergence tests of MgB2_2 with respect to Brillouin-zone sampling are presented in Figs. S5-S10.

In Figure S12, we show two cases of YB12_{12} and C2_2I2_2Y2_2 in which the coarse k\mathbf{k}- grids from Materials Project and fine k\mathbf{k}- grid only twice of that result to qualitatively inaccurate results, unusually high TcT_{\mathrm{c}} for the former while predicting negligible TcT_{\mathrm{c}} for the latter. With coarse grids, the exponential decay parameter of YB12_{12} is around 38, much larger than the critical value estimated for MgB2_2 (AMgB2∼12A_{\mathrm{MgB}_2} \sim12). Increasing k\mathbf{k}-mesh grids two times in each direction, TcT_{\mathrm{c}} drops from 40.6 K to 1.67 K at σ=0.01\sigma= 0.01 Ry and drops to almost zero for σ>0.01\sigma> 0.01 Ry indicating non-superconductor, which is in good agreement with the experiment [12]. We found a couple of cases such as C2_2I2_2Y2_2 when TcT_{\mathrm{c}} first decreases and increases later with σ\sigma (we set A=0A = 0 in Figure S14). Such behavior hasn’t been observed in G(Δ\Delta,σ\sigma) vs σ\sigma plots for a wide range of Δ\Delta (Fig. 8(a) of the main text). The coarse grids result to a very low TcT_{\mathrm{c}} of 0.97 K for C2_2I2_2Y2_2. However, the denser mesh corrects the behavior with an exponential decay parameter of 12.5 much closer to the reference value of MgB2_2. Also, the critical temperature TcT_{\mathrm{c}} improves to 5.29 K agreeing much more strongly to the experiment [13]. Note that, this analysis only works for compounds with non-zero TcT_{\mathrm{c}} for a wide range of σ\sigma. For Tc≲1T_{\mathrm{c}} \lesssim1 K at σ=0.01\sigma= 0.01 Ry (computational details for choosing σ=0.01\sigma= 0.01 Ry), Tc<0.1T_{\mathrm{c}} < 0.1 K for σ>0.01\sigma> 0.01 Ry, and rapidly drops to zero for larger σ\sigma, we identify the compound to be non-superconducting with the value of AA as zero (A=0<AMgB2A = 0 < A_{\mathrm{MgB}_2}, already converged). To summarize the ansatz, we present a simple schematics of the process shown as in Figure S13.

Figure S12. TcT_{\mathrm{c}} vs σ\sigma for various k- grids; (a) for YB12_{12} and (b) for C2_2I2_2Y2_2.

Figure S14. Critical temperature TcT_c vs exponential decay parameter (AA) plot; Black open circle and red cross symbols represent converged (A<AMgB2A < A_{MgB_2}) true and false predictions respectively; Blue plus and green diamond symbols represent unconverged (A>AMgB2A > A_{MgB_2}) true and false predictions respectively; A vertical dashed line at A=12A = 12 represents the MgB2_2 result; When TcT_c doesn’t exponentially decay or increase with σ\sigma (parabolic in nature, first decrease and increase later), those results are simply identified as unconverged, and presented in the plot with A=0A = 0; With denser grids, these results of parabolic TcT_c vs σ\sigma nature changes to exponential decay. Here, a true prediction indicates agreement with the available experimental results, while a false prediction indicates disagreement.

TABLE S7: Comparing results between coarse and fine grids for compounds with A>AMgB2A > A_{MgB2}; TceffT_c^{eff} is the critical temperature calculated from efficient EPC calculations, using k-point mesh grid from MP database; TcCorrT_c^{Corr} is the critical temperature calculated using denser k-point mesh and fine k- mesh grids (2 times of MP database K-point mesh), while q-grid is kept fixed;

CompoundTceffT_c^{eff} (K)TcCorrT_c^{Corr} (K)TcExptT_c^{Expt} (K)
YB12 (Fm-3m)40.61.670 [12][14]
CaNiBN (P4/nmm)0.962.432.2 [15]
BaC6 (P63/mmc)6.501.700 [16]
La3PbC (Pm-3m)3.200.500 [17][18]
La3SnC (Pm-3m)5.090.810 [17] [19]
CsC8 (P6/mmm)5.970.40.13 [20]
AlB2 (P6/mmm)11.30.30 [21]
C2I2Y2 (C2/m)0.975.299.97 [13]
C2Cl2Y2 (C2/m)0.5082.732.3 [22]

Table S7.

TABLE S8: Comparing results between coarse and fine grids for compounds with A<AMgB2A < A_{MgB2}; TceffT_c^{eff} is the critical temperature calculated from efficient EPC calculations, using k-point mesh grid from MP database; TcCorrT_c^{Corr} is the critical temperature calculated using denser k-point mesh and fine k- mesh grids (2 times of MP database K-point mesh), while q-grid is kept fixed; Here results do not change much qualitatively from coarse to denser mesh. Vertical line in the table after LaB6 separates data from accurate to inaccurate qualitative predictions, compared to experimental results.

CompoundTceffT_c^{eff} (K)TcCorrT_c^{Corr} (K)TcExptT_c^{Expt} (K)
WB (Cmcm)5.694.92.8 [23]
ReB2 (P63/mmc)000 [24]
LaB6 (Pm-3m)0.4200.005 [25]
Ta2C (P-3m1)0.00.00 [26]
YIr3B2 (P6/mmm)5.55.54N/A [27]
RuB2 (Pmmn)0.590.321.5 [28]
VC (Fm-3m)31.6719.93.2 [29]
ZnNi3C (Pm-3m)15.3715.40 [30][31]
YRh3B2 (P6/mmm)5.84.20 [32] [33]
YRu3B2 (P6/mmm)3.72.120 [34][35]

Table S8.

Next, we present results obtained from EPC calculations for 113 known boron and carbon superconducting (N = 53) and non-superconducting (N = 60) compounds from SuperCon database [36]. Figure S14 represents the distribution of TcT_c with respect to exponential decay parameter AA for efficient calculations for σ=0.01\sigma= 0.01 Ry with μc∗=0.16\mu_c^* = 0.16. Based on the convergence check ansatz, we found that around 15 % (NoConv) of the results are not fully converged, while 70 % of them are qualitatively inaccurate (NoConv-False), compared to available experimental results. Similarly, we have 12 % of cases that have A<AMgB2A < A_{MgB2} with qualitatively inaccurate predictions (Conv-False), which could be attributed from either inaccuracy of approximations to compute TcT_c or inaccuracies within available experimental results. This work not only addresses the 15 % (NoConv) cases but also assists in validating true (Conv-True) results.

We present a few qualitatively inaccurate predictions with A>AMgB2A > A_{MgB_2} and improved results with a denser grid in Table S7, while Table S8 shows unaffected results (qualitatively) with denser grids for the cases with A<AMgB2A < A_{MgB_2}. For example, AlB2_2 is predicted to be superconductors with a coarse grid, while it changes to a non-superconductor with a denser grid, consistent with the experimental prediction. Similarly, TcT_c of superconductors such as CaNiBN, C2_2I2_2Y2_2, and C2_2Cl2_2Y2_2 change from below 1 K (non-superconductor) to a larger value, and show a better agreement with the experiment. On the contrary, the qualitative results remain unchanged for converged calculations represented by A<AMgB2A < A_{MgB_2}, regardless of their agreement with the experimental references.

Application of convergence ansatz on known compounds of SuperCon

In this section, we compare the computed TcT_c with available experimental results, as shown in Figure S2 for efficient EPC calculations with and without improved data for A>AMgB2A > A_{MgB_2} (left panel) and A<AMgB2A < A_{MgB_2} (right panel) cases respectively. We also highlight some significantly deviated results with red rectangles for efficient calculations. With improved data, computed critical temperatures TcT_c have much better agreement compared to the references, as indicated by the red rectangle enclosing more data points (right panel of Figure S2), compared to the plot on the left. The results obtained from the ab-initio calculations demonstrate a notable level of accuracy with the mean absolute error (MAE) of 2.21 K for the critical temperature, compared to experiments, highlighting the reliability and robustness of this ansatz test. This is in comparison to the MAE of 3.18 K observed in efficient calculations. Please note that we have exclusively incorporated DFPT TcT_c results for systems that are dynamically stable. Despite the general agreement between DFPT and experimental findings, there exist notable discrepancies. Instances like ZnNi3C (Pm-3m), with a calculated TcT_c of 15 K, deviate enormously from the experimental values of 0 K [37]. A comparable investigation of ZnNi3C by Hoffmann et al.[38], utilizing different pseudopotentials from the PSEUDODOJO project[39], revealed phonon instabilities in these compounds. Notably, these unstable phonon modes were identified to possess significant mode-resolved EPC strength[40], a finding deserving attention, as we address these concerns in the “Imaginary Phonon Modes and Superconductivity” section. Similarly, Nb2InC (P63/mmc) with a calculated TcT_c of 0.34 K, VC (Fm-3m) at 20 K, and WB (I4_1/amd) at 18.5 K, present a problematic cases. Experimental TcT_c measurements stand at 3.2 K [29], 4.3 K [41], and 7.5 K [42] for VC, WB, and Nb2InC, respectively. These discrepancies can be ascribed to the constraints of both theoretical approaches and experimental procedures. For instance, older experimental investigations initially reported critical temperatures that were later rectified by more recent experiments or vice versa. For instance, in Ref [43], superconductivity below 4.7 K was claimed in certain YB12_{12} samples, yet subsequent experiments, such as those in [44], failed to corroborate these findings. Furthermore, employing more sophisticated theoretical methodologies like solving the many-body anisotropic Migdal-Eliashberg equations or utilizing SCDFT to estimate TcT_{\mathrm{c}} can significantly enhance the accuracy of calculations compared to isotropic methods. As an illustration, when examining MgB2_2, TcT_{\mathrm{c}} calculated using isotropic Eliashberg theory falls within the 15-20 K range. However, in Ref. [45], TcT_{\mathrm{c}} values of 34-42 K were reported over a range of μ∗\mu^{*} values, demonstrating the improvement achievable with these advanced techniques.

In Table S9, we show some of the inaccurate results predicting nonsuperconductors with 2D q- grids (4×4×14 \times4 \times1). Emergence of strong el-ph coupling and TcT_{\mathrm{c}} with denser 3D q- grid (3×3×33 \times3 \times3) is unlikely for such nonsuperconducting predictions with Tc<0.01T_{\mathrm{c}} < 0.01 K for σ>0.01\sigma> 0.01 Ry [Fig. S13].

CompoundGridsλ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)TcExptT_{\mathrm{c}}^{Expt} (K)
Nb2InC12×12×1212 \times12 \times12, 3×3×33 \times3 \times30.392850.17.5 [46]
8×8×28 \times8 \times2, 4×4×14 \times4 \times10.432970.34
Ti2InC12×12×1212 \times12 \times12, 3×3×33 \times3 \times30.183550.03.1 [47]
16×16×416 \times16 \times4, 4×4×14 \times4 \times10.133520.0

Table S9. Comparison of superconducting properties predicting nonsuperconductors that do not agree with the experiments. Denser k- and q- grids do not improve the results.

DFPT results for dynamically stable compounds

References

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)
mp-2252B2Sc1191-0.840.346060.032
mp-10020C1Sc1225-0.140.614425.086
mp-29941C1Sc2164-0.420.243510.0
mp-1232372B2C2Sc2194-0.60.215650.0
mp-27693B8C8Sc455-0.460.455601.092
mp-10343B1C2Sc2139-0.570.414670.383
mp-10139B1Sc3Sn1221-0.610.13310.0
mp-12062Ag1B21910.520.8856821.563
mp-7817B12Y1225-0.240.486051.69
mp-1084B12Zr1225-0.210.613063.409
mp-345B1Hf1225-0.410.512811.225
mp-1890B4Mo4141-0.50.43640.228
mp-999198B2Mo263-0.490.633384.398
mp-451B1Zr1225-0.370.473830.946
mp-763B2Mg1191-0.130.7873420.027
mp-450B2Nb1191-0.690.723657.68
mp-1773B4Re2194-0.430.224090.0
mp-1077B4Ru259-0.290.423640.356
mp-1472B2Zr1191-0.990.145810.0
mp-2680B6La1221-0.560.263430.0
mp-2203B6Y1221-0.41.97819.999
mp-1079500B2Ca2N2Ni2129-0.970.534642.474
mp-1106180B6Re14186-0.220.721803.83
mp-15671B2Re663-0.20.722094.4
mp-2850B4Os259-0.210.612633.07
mp-21502B4Rh862-0.210.651832.692
mp-7857B4Ti462-0.830.274430.0
mp-20881B2La2N2Ni2129-1.050.643014.241
mp-9219B2C1La1Pt2139-0.640.762626.732
mp-3465B2La1Rh3191-0.670.841745.859
mp-6114B2La3N3Ni2139-1.150.613624.056
mp-20234B4Li8Pt12212-0.571.31967.615
mp-4984B4Mo10Si2140-0.40.812577.838
mp-1189073B8Os4Sc462-0.680.633304.305
mp-1105309B4Ge2Ta10140-0.590.292340.0
mp-1078866B4C4Y2127-0.440.524552.323
mp-15955B4C4Y2131-0.30.53921.531
mp-1024989B2C1Pd2Y1139-0.51.1424915.971
mp-5984B8Rh8Y2137-0.590.772235.866
mp-1190832B16Ru4Y455-0.620.285010.0
mp-980205B16Os4Y455-0.570.284040.0
mp-2091B4V6127-0.720.234730.0
mp-260B2Cr263-0.520.454120.68
mp-9973B2V263-0.850.235080.0
mp-1183252B1Ir1187-0.20.472140.539
mp-567164B2Rh2194-0.390.153130.0
mp-1063752B2Rh2194-0.180.911737.127
mp-4472B2C2Mo463-0.250.781995.414
mp-10112B2Ir3La1191-0.650.621511.828
mp-2536B2Ni4140-0.290.223160.0

Table S10. EPC properties of dynamically stable compounds

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)
mp-1491B2V1191-0.740.294940.001
mp-8431B4Ca2Rh470-0.610.263250.0
mp-1008487B2W263-0.360.682744.902
mp-10142B4Ta371-0.770.363570.046
mp-865Ca1B6221-0.410.165380.0
mp-14445B4Ca2Ir470-0.680.263380.0
mp-944Al1B2191-0.040.45640.295
mp-10852B4C4La2127-0.440.384200.113
mp-7783B4C4La2131-0.310.54591.862
mp-2967B2Co2La1139-0.480.313040.002
mp-5992B2C1Ir2La1139-0.570.362500.035
mp-568083B2C1La1Ni2139-0.440.332990.009
mp-6794B2C1La1Rh2139-0.60.352840.022
mp-571428B6Mg3Ni9181-0.380.23530.0
mp-4938B2Co3Sc1191-0.550.372750.07
mp-6939B4Ir4Sr270-0.630.283510.0
mp-7348B4Rh4Sr270-0.560.233350.0
mp-3515B2Co2Y1139-0.560.243830.0
mp-1024941B2Rh3Y1191-0.710.761624.181
mp-4382B2Ru3Y1191-0.470.621722.121
mp-9956Al2C2Cr4194-0.170.244030.0
mp-3271Al1C1Ti3221-0.590.282710.0
mp-1025497Al2C2V4194-0.520.234250.0
mp-4448Y3Al1C1221-0.420.12540.0
mp-1190604Al8C4Nb12213-0.460.352560.022
mp-1190760Al8C4Ta12213-0.450.41910.111
mp-1214417Ba2C12194-0.090.533001.709
mp-9961C2Cd2Ti4194-0.540.282330.0
mp-21075C1Hf1225-0.940.134570.0
mp-1096993C2Hf2194-0.740.464160.865
mp-1002124Hf1C1216-0.30.054240.023
mp-10611C1La3Pb1221-0.430.491510.516
mp-1206443C1La3Sn1221-0.490.531460.806
mp-5443C2Nb4Sn2194-0.420.452710.527
mp-20661C2Pb2Ti4194-0.570.313210.004
mp-631C1Ti1225-0.810.166010.0
mp-1282C1V1225-0.411.1431219.989
mp-20648C4V860-0.470.33480.001
mp-1008632C1V2164-0.450.243190.0
mp-2795C1Zr1225-0.80.154840.0
mp-1014307C2Zr2194-0.640.414800.321
mp-570112C4Cr663-0.070.553432.407
mp-28861C8Cs1191-0.050.47660.438
mp-568643C16Rb270-0.040.57552.902
mp-16290C1Ni3Zn1221-0.061.9412615.406
mp-1066566C2Ni1Y138-0.330.473951.074
mp-1079635C2Ga2Mo4194-0.120.772496.627
mp-2305C1Mo1187-0.080.185120.0
mp-1552C4Mo860-0.110.72725.298
mp-1221498C1Mo2164-0.050.62873.119

Table S11. EPC properties of dynamically stable compounds

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_c (K)
mp-1009832C1Ta1C_{1}Ta_{1}216-0.060.692134.034
mp-7088C1Ta2C_{1}Ta_{2}164-0.60.222500.0
mp-2034C4W8C_{4}W_{8}60-0.020.772125.512
mp-33065C2W4C_{2}W_{4}58-0.010.792136.119
mp-2367C2La1C_{2}La_{1}139-0.160.52580.978
mp-313C2Y1C_{2}Y_{1}139-0.190.832829.383
mp-1208630C12Sr2C_{12}Sr_{2}194-0.040.673105.291
mp-1018048C2La1Ni1C_{2}La_{1}Ni_{1}38-0.250.713466.943
mp-1206284Br2C2La2Br_{2}C_{2}La_{2}12-1.120.71743.416
mp-20315C2In2Ti4C_{2}In_{2}Ti_{4}194-0.670.133520.0
mp-643367Br2C2Y2Br_{2}C_{2}Y_{2}12-1.080.772516.614
mp-37919Br2C2Y2Br_{2}C_{2}Y_{2}59-0.081.211097.789
mp-23062C2I2Y2C_{2}I_{2}Y_{2}12-0.860.771965.177
mp-1206889C2Cl2Y2C_{2}Cl_{2}Y_{2}12-1.440.62582.725
mp-6576B2C1Ni2Y1B_{2}C_{1}Ni_{2}Y_{1}139-0.510.593153.159
mp-2192B2Pt2B_{2}Pt_{2}1940.071.2418813.786
mp-1009218C1Mo1C_{1}Mo_{1}2160.51.7110311.153
mp-632442Al4C3Al_{4}C_{3}10.310.62562.704
mp-1076KB6KB_{6}221-0.030.518233.715
mp-1078278CrB4CrB_{4}58-0.310.256220.0
mp-1079437FeB4FeB_{4}58-0.170.946818.723
mp-1080111B3MoB_{3}Mo166-0.310.295740.001
mp-1106184MnB4MnB_{4}14-0.290.145590.0
mp-1213975CaB4CaB_{4}127-0.390.255510.0
mp-1228730B24Mo7B_{24}Mo_{7}187-0.141.3335928.901
mp-2315NaB15NaB_{15}74-0.050.17630.0
mp-262Na3B20Na_{3}B_{20}65-0.060.47820.513
mp-27710CrB4CrB_{4}71-0.30.237550.0
mp-576B13C2B_{13}C_{2}166-0.070.9783539.304
mp-637YB4YB_{4}127-0.590.433880.457
mp-7283LaB4LaB_{4}127-0.560.284650.0
mp-1218188SrLaB12SrLaB_{12}123-0.510.363060.041
mp-1232339LiC12LiC_{12}63-0.00.4310551.24
mp-1227841BaLaB12BaLaB_{12}123-0.490.363040.052
mp-1001581LiC6LiC_{6}191-0.00.3910430.397
mp-1021323LiC12LiC_{12}191-0.010.339490.035
mp-1001835LiBLiB194-0.170.213280.0
mp-1002188TcBTcB187-0.280.44450.267
mp-1009695CaB2CaB_{2}191-0.140.684447.604
mp-1224328HfNbB4HfNbB_{4}191-0.860.334620.013
mp-1019317TcB2TcB_{2}194-0.440.274970.0
mp-1025170Ti3B4Ti_{3}B_{4}71-0.930.255170.0
mp-1217974TaB4W3TaB_{4}W_{3}25-0.480.62772.921
mp-1080021Nb2B3Nb_{2}B_{3}63-0.750.454090.777
mp-1217965TaB2MoTaB_{2}Mo38-0.660.573482.788
mp-1102394Nb5B6Nb_{5}B_{6}65-0.770.43980.209
mp-1108TaB2TaB_{2}191-0.651.2225318.116
mp-1217924TaNbC2TaNbC_{2}166-0.521.4132628.387
mp-1217818TaVB4TaVB_{4}191-0.690.63874.047
mp-1206441V5B6V_{5}B_{6}65-0.830.264880.0

Table S12. EPC properties of dynamically stable compounds

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)
mp-1216709TiNbB238-0.810.314720.005
mp-20857CoB62-0.40.113570.0
mp-2580NbB63-0.770.314150.005
mp-1215250ZrB4Mo191-0.621.319315.047
mp-569803B2W194-0.330.483230.999
mp-999120TcB194-0.220.423780.335
mp-1215178ZrTiB4191-1.00.096170.0
mp-9208V2B363-0.80.294880.001
mp-999118TcB194-0.121.3233226.488
mp-12054Cr2B363-0.430.563582.753
mp-10114ScB4Ir3176-0.590.182580.0
mp-9880YB2C135-0.520.325060.012
mp-995282LiAlB455-0.050.56932.659
mp-1008527B2CN115-0.440.864418.759
mp-1207086MgAlB4191-0.150.347690.059
mp-1215209ZrTaB4191-0.840.394360.177
mp-1215223ZrNbB238-0.710.354040.04
mp-1216966TiCrB4191-0.680.593393.385
mp-1217028TiB2Mo38-0.820.254610.0
mp-1220697Nb2AlB6191-0.470.544312.781
mp-1224283HfTaB4191-0.850.414230.354
mp-1224291HfTaB238-0.770.43540.199
mp-1224347HfNbB238-0.760.373870.073
mp-1226228CrB2Mo38-0.480.623313.919
mp-13341YB4Rh55-0.610.624836.008
mp-568985La2B3Br187-0.840.352120.018
mp-569002Y3ReB763-0.330.813309.953
mp-1008526B2CN156-0.421.1156634.9
mp-569121La2B3Cl174-1.020.412090.158
mp-1191641YVB455-0.70.34730.001
mp-1216398VCrB238-0.690.394600.19
mp-1216475V3ReB425-0.720.393910.169
mp-1216667TiVB4191-0.90.275160.0
mp-1217026TiB4Mo191-0.680.693486.374
mp-1217095Ti3B4Mo6-0.830.284540.0
mp-1217898TaTiB4191-0.880.294410.001
mp-1217997Ta3AlB8191-0.50.543161.858
mp-1220349NbB2W38-0.580.573412.717
mp-1220351NbVB4191-0.710.514211.948
mp-1220383NbB4Mo338-0.560.583543.101
mp-1008525B2CN160-0.431.6657760.733
mp-10057VCoB363-0.620.264380.0
mp-1226242CrB2W38-0.430.742906.877
mp-1228634B4MoIr187-0.250.52801.054
mp-1196778Y2B6Ru55-0.660.225110.0
mp-22709TaNiB262-0.630.33240.001
mp-569116Sc2B6Rh55-0.730.524142.079
mp-1224274HfTiC2166-0.830.424640.462
mp-1215222ZrNbC2166-0.660.733978.92
mp-1224334HfNbC2166-0.730.653855.753

Table S13. EPC properties of dynamically stable compounds

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)
mp-38818HfNbB471-0.860.44200.254
mp-1086667ScNiC2129-0.280.523681.864
mp-1217975Ta3TiB425-0.840.263940.0
mp-9530Y4C714-0.280.133230.0
mp-9459Y4C555-0.330.842257.678
mp-1188534ScCrC259-0.340.623894.881
mp-1224170HfZrC2123-0.870.254830.0
mp-1334Y2C166-0.290.292010.0
mp-1232379YBC194-0.290.284270.0
mp-29896Y2B3C265-0.50.63753.915
mp-612670YNiBC129-0.520.24210.0
mp-567692B4C3Ni4Y3139-0.510.313930.003
mp-12737B2C1Rh2Y1139-0.60.382850.098
mp-1087495Tc3B63-0.270.82547.66
mp-978989Tc7B3186-0.320.682274.037
mp-1068296Fe(BW)271-0.430.452790.465
mp-11750Ti6Si2B189-0.660.223330.0
mp-1238800CaBC194-0.070.425010.494
mp-27261Ba7(BIr)12166-0.480.472330.626
mp-29980Nb4B3C263-0.591.0121911.258
mp-29979Nb3B3C63-0.661.2222916.438
mp-541849Al3(BRu2)2123-0.590.263080.0
mp-1188408Zr5Sn3B193-0.680.452020.343
mp-1206909CaBPd3221-0.561.11036.187
mp-9985NbNiB63-0.560.342560.018
mp-1080829Ti6Ge2B189-0.620.213050.0
mp-1025192Ta4C3221-0.451.4613812.644
mp-1215211ZrNbB4191-0.850.314790.005
mp-1220641Nb3B4W338-0.430.462990.66
mp-10721Ti2C227-0.640.143860.0
mp-27919Ti8C5166-0.720.224550.0
mp-1218000Ta4C3160-0.550.521830.919
mp-1220752Nb10(SiB)342-0.650.243100.0
mp-1189539Hf2Al3C5194-0.130.4714543.645
mp-9958Ti2GeC194-0.810.482760.847
mp-1025524Zr2TlC194-0.630.211880.0
mp-1216707TiNbC2166-0.670.734229.623
mp-1079992Zr2PbC194-0.660.42300.146
mp-1207413Zr5Sn3C193-0.680.531801.013
mp-1217106Ti2C166-0.630.353570.037
mp-1217822TaVC2166-0.460.834310.296
mp-12990Ti2AlC194-0.70.253710.0
mp-3871Ti2SnC194-0.730.243530.0
mp-1025427Ta2GaC194-0.520.452310.385
mp-1078712Hf2TlC194-0.630.211710.0
mp-1079076Hf2PbC194-0.630.411940.15
mp-1079908Ti2SiC194-0.790.43340.205
mp-1220365NbVC2166-0.390.8336511.893
mp-21023Ti3SnC2194-0.790.184210.0
mp-22144Ta2InC194-0.380.372330.062

Table S14. EPC properties of dynamically stable compounds

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)
mp-13137Hf2_2CS194-1.510.193030.0
mp-1220725Nb2_2CN166-0.80.8831011.685
mp-1216616V2_2CN166-0.81.032216.224
mp-1214755BPd6_615-0.20.161690.0
mp-7424BPd2_258-0.270.421590.16
mp-1078540Ni6_6Ge2_2B189-0.310.341900.011
mp-1078623Zr2_2BIr6_6225-0.810.361840.03
mp-12073Ba(BIr)2_2139-0.50.372690.051
mp-7349Ba(BRh)2_2139-0.480.292950.001
mp-7705NbFeB187-0.121.7616218.063
mp-1215258ZrBeB187-0.580.454360.751
mp-1208348Ta5_5Ga3_3B193-0.420.311900.002
mp-605839Li2_2B2_2Rh3_355-0.50.273020.0
mp-8308Ca3_3Ni7_7B2_2166-0.340.133050.0
mp-1206490Nb2_2B2_2Mo127-0.640.283670.0
mp-31052LaBPt2_2180-0.920.281420.0
mp-1223681La2_2(Ni2_2B)3_344-0.330.462270.517
mp-1097B2_2Ta2_263-0.810.413660.297
mp-28930C16_{16}K2_270-0.030.5711028.897
mp-7832B4_4W4_4141-0.371.5818618.463
mp-28613B3_3Li3_3Pt9_9189-0.510.41550.088
mp-569759B4_4Rh8_8Zn5_565-0.420.262220.0
mp-571419Al4_4C5_5Zr2_2166-0.350.363870.06
mp-1207385Al8_8C8_8Zr2_2164-0.250.264820.0
mp-1189895B2_2Ge6_6Ta10_{10}193-0.450.611812.053
mp-10140B1_1Sc3_3Tl1_1221-0.360.542161.324
mp-1216165B1_1Si6_6Y10_{10}162-0.650.41660.103
mp-20175C2_2In2_2Sc4_4194-0.550.221700.0
mp-20983C2_2In2_2V4_4194-0.340.353490.03
mp-1224263B4_4Hf1_1Ti1_1191-1.020.15920.0
mp-1224184B4_4Hf1_1Zr1_147-1.00.155070.0
mp-8307B2_2Ca2_2Ni8_8191-0.30.331290.004
mp-4079Al1_1C1_1Sc3_3221-0.590.043030.048
mp-1103814C3_3K5_5N6_6229-0.520.081900.0
mp-1224285C2_2Hf1_1Ta1_1166-0.810.623594.486
mp-1215219C2_2Ta1_1Zr1_1166-0.740.653705.558
mp-570499B2_2La5_5N6_612-1.510.253510.0
mp-569935B2_2La3_3N4_471-1.50.263780.0
mp-1223086C6_6La2_2Y1_112-0.160.562712.049
mp-1221519C1_1Mo2_2N1_125-0.350.374730.105
mp-1222150Al1_1B10_{10}Mg4_4191-0.140.597687.488
mp-1189984C8_8Mo4_4Y4_462-0.240.614094.562
mp-3380C8_8La4_4Rh4_476-0.330.272630.0
mp-4262BeAlB216-0.050.165780.0
mp-5971YBPt2_2180-1.00.181910.0
mp-9596La(BIr)4_486-0.580.472010.503
mp-1105186Cu3_3B5_5Pt9_9189-0.210.41640.099
mp-1106165Nb5_5Si3_3B193-0.690.462820.638
mp-1106398V5_5Ge3_3B193-0.460.383030.11
mp-1188194Ta3_3B2_2Ru5_5127-0.530.331460.006

Table S15. EPC properties of dynamically stable compounds

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)
mp-29723LaB₂Ru₃191-0.410.671482.477
mp-22759CoBW62-0.430.373420.071
mp-28786Zn(BIr)₂139-0.320.083190.0
mp-9999Ni(BMo)₂71-0.480.33520.001
mp-1076987TaNiB63-0.620.32530.001
mp-3348LiBIr70-0.510.163250.0
mp-2760Nb₆C₅12-0.550.513931.632
mp-32679Nb₁₀C₇12-0.480.453010.537
mp-1226378Cr₂C164-0.040.383760.118
mp-2318Nb₂C164-0.450.312850.002
mp-974437Re₂C194-0.030.323480.007
mp-10037AlCo₃C221-0.220.12120.0
mp-9987Nb₂PC194-0.750.443450.493
mp-21003Y₂ReC₂62-0.420.292820.0
mp-28767Sc₅Re₂C₇65-0.480.213840.0
mp-567462Sc₃RhC₄12-0.50.683936.91
mp-7130ScRu₃C221-0.280.42270.12
mp-996161Nb₃AlC₂194-0.520.363840.048
mp-996162Nb₂AlC194-0.510.463370.768
mp-1078811Nb₂GeC194-0.510.453070.531
mp-1080835V₂GaC194-0.520.273750.0
mp-1189574YWC₂62-0.270.543892.392
mp-4992ScCrC₂194-0.340.624154.966
mp-8044V₂PC194-0.690.44090.227
mp-10046V₂AsC194-0.530.473540.878
mp-1212439Hf₅Al₃C193-0.470.229880.0
mp-1217764Ta₂CN123-0.822.712719.495
mp-37179Ta₂CN141-0.841.9416219.757
mp-4384Nb₂CS₂166-1.121.0322211.827
mp-559976Ta₂CS₂164-1.190.552511.71
mp-995201Ti₅Si₃C193-0.820.552531.784
mp-1025441Ta₂AlC194-0.520.442720.374
mp-1079546Nb₂GaC194-0.520.442630.424
mp-1220371NbAlVC164-0.470.443620.491
mp-3732Ti₂CS194-1.420.24730.0
mp-4563Ti₃TlC221-0.440.292290.0
mp-1216139Y₄C₄I₃Br8-0.910.841745.825
mp-1220491Nb₆V₂(CS₂)₃12-1.090.542781.768
mp-1220693Nb₂CuCS₂156-0.90.852077.255
mp-1215225ZrTaCN160-1.150.783018.323
mp-1025205Y₂Re₂Si₂C12-0.620.731723.887
mp-1215184ZrTiCN160-1.330.434370.5
mp-1224279HfTiCN160-1.410.44200.267
mp-1009894Zr₁C₁216-0.190.014440.405
mp-1068661ZrBRh₃221-0.750.011970.176
mp-1145B₂Ti₁191-1.060.116380.0

Table S16. EPC properties of dynamically stable compounds

IDCompoundSGFE (eV/atom)λ\lambdaωlog⁡\omega_{\log} (K)TcT_{\mathrm{c}} (K)
mp-1080664B4_4Cr4_4141-0.530.274490.0
mp-1994B2_2Hf191-1.020.165180.0
mp-2331B4_4Mo2_2166-0.430.44370.239
mp-20689B4_4Nb6_6127-0.630.263650.0
mp-13415B4_4Ta6_6127-0.670.243060.0
mp-910C Nb225-0.461.1130918.924
mp-1094093C2_2Nb2_2194-0.351.1430019.284
mp-999388C2_2Nb2_2194-0.330.8445515.269
mp-999377C2_2Nb2_2194-0.090.713206.7
mp-1086C Ta225-0.580.682364.239
mp-1207750Y5_5(SiB4_4)2_2127-0.680.334010.014
mp-1217023TiB2_2W38-0.780.34190.001
mp-1079333B2_2CN51-0.530.4610452.381
mp-13854B3_3Ru2_2194-0.330.144450.0
mp-14019NiB63-0.240.243760.0
mp-1018050CrC187-0.00.236100.0
mp-1215480Zr3_3NbC4_4166-0.730.773469.324
mp-1216691TiVC2_2166-0.620.534462.47
mp-1215174ZrTiC2_2123-0.710.254610.0
mp-1215218ZrMoC2_2166-0.351.7912013.656
mp-4893Hf2_2SnC194-0.780.312050.003
mp-5659Ti3_3SiC2_2194-0.820.373450.081
mp-1092281Ti2_2TlC194-0.570.142950.0
mp-3747Ti3_3AlC2_2194-0.760.263420.0
mp-3886Zr2_2AlC194-0.640.561631.181
mp-1188856V5_5Si2_2B140-0.390.373020.067
mp-1216445V9_9Cr3_3B8_810-0.70.254590.0
mp-1216643V10_{10}Si6_6B162-0.630.333230.014
mp-1220688Nb3_3Re3_3B4_438-0.350.532651.47
mp-1226327Cr3_3B4_4Mo3_338-0.290.782898.091
mp-1009817C Ta187-0.161.5423522.8
mp-1542YB2_2191-0.560.424730.453
mp-1216692TiNbB4_4191-0.890.294810.001
mp-4613Zr2_2SnC194-0.790.322590.004
mp-1232384ZrBC194-0.321.1232019.998

Table S17. EPC properties of dynamically stable compounds.

References

  1. [20]J. Chacón-Torres and T. Pichler, Defect modulated raman response of KC8 single crystals, Phys. Status Solidi (b) 248, 2744 (2011).DOI
  2. [21]U. Burkhardt, V. Gurin, F. Haarmann, H. Borrmann, W. Schnelle, A. Yaresko, and Y. Grin, On the electronic and structural properties of aluminum diboride Al₀.₉B₂, J Solid State Chem. 177, 389 (2004).
  3. [22]M. Bäcker, A. Simon, R. K. Kremer, H.-J. Mattauch, R. Dron skowski, and J. Rouxel, Superconductivity in intercalated and substituted y2br2c2, Angew. Chem. Int. Ed. Engl. 35, 752 (1996).DOI
  4. [23]M. Kayhan, E. Hildebrandt, M. Frotscher, A. Sesnyshyn, K. Hofmann, L. Alff, and B. Albert, Neutron diffraction and observation of superconductivity for tungsten borides, WB and W2B4, Solid State Sci. 14, 1656 (2012).
  5. [24]C. Pei, J. Zhang, C. Gong, Q. Wang, L. Gao, Y. Zhao, S. Tian, W. Cao, C. Li, Z.-Y. Lu, et al., Distinct superconducting behaviors of pressurized WB2 and ReB2 with different local layers, Sci. China: Phys. Mech. Astron. 65, 1 (2022).DOI
  6. [25]I. Bat’Ko, M. Bat’Kova, K. Flachbart, V. Filippov, Y. B. Paderno, N. Y. Shicevalova, and T. Wagner, Electrical resistivity and superconductivity of LaB6 and LuB12, J. Alloys Compd. 217, L1 (1995).DOI
  7. [26]V. Hadjiyev, E. Baburaj, and J. Guan, Phonons in ultra-high melting temperature Ta2C, EPL (Europhys. Lett.) 111, 68001 (2015).
  8. [27]Not reported.
  9. [28]S. Datta, A. Vasdev, S. Halder, J. Singh, Y. Singh, and G. Sheet, Spectroscopic signature of two superconducting gaps and their unusual field dependence in RuB2, J. Phys.: Condens. Matter. 32, 315701 (2020).
  10. [29]K. Langguth, Production of superconducting vanadium carbide by ion implantation, Tech. Rep. (Kernforschungszentrum Karlsruhe (Germany, 1977).
  11. [30]M.-S. Park, J. G. Kim, S.-H. Park, Y. Lee, S. Lee, and E. Choi, Physical properties of zincni3: comparison with superconducting mcgni3, Supercond. Sci. Technol. 17, 274 (2003).DOI
  12. [31]No transition observed above 2.0 K.
  13. [32]D. Johnston, Superconductivity in a new ternary structure class of boride compounds, Solid State Commun. 24, 699 (1977).DOI
  14. [33]No transition observed above 1.5 K.
  15. [34]H. Ku, G. Meisner, F. Acker, and D. Johnston, Superconducting and magnetic properties of new ternary borides with the CeCo3B2-type structure, Solid State Commun. 35, 91 (1980).DOI
  16. [35]No transition observed above 1.2 K.
  17. [36]National institute of materials science, materials information station, http://supercon.nims.go.jp/index_en.html, SuperCon (2018).
  18. [37]M.-S. Park, J. Guim, S.-H. Park, Y. Lee, S. Lee, and E. Choi, Physical properties of zincni3: comparison with superconducting MgCNi3, Supercond. Sci. Technol. 17, 274 (2003).DOI
  19. [38]N. Hoffmann, T. F. Cerqueira, J. Schmidt, and M. A. Marques, Superconductivity in antiperovskites, Npj Comput. Mater. 8, 150 (2022).
  20. [39]M. J. van Setten, M. Giantomassi, E. Bousquet, M. J. Verstraete, D. R. Hamann, X. Gonze, and G.-M. Rignanese, The pseudodojo: Training and grading a 85 element optimized norm-conserving pseudopotential table, Comput. Phys. Commun. 226, 39 (2018).DOI
  21. [40]N. Hoffmann, T. F. Cerqueira, J. Schmidt, and M. A. Marques, Superconductivity in antiperovskites, Materials Cloud Archive 2022.49 (2022).
  22. [41]M. Kayhan, E. Hildebrandt, M. Frotscher, A. Senyshyn, K. Hofmann, L. Alff, and B. Albert, Neutron diffraction and observation of superconductivity for tungsten borides, WB and W2B4, Solid State Sci. 14, 1656 (2012).
  23. [42]A. Bortolozo, Z. Fisk, O. Sant’Anna, C. Dos Santos, and A. Machado, Superconductivity in Nb2InC, Physica C Supercond 469, 256 (2009).DOI
  24. [43]B. Matthias, T. Geballe, K. Andres, E. Corenzwit, G. Hull, and J. Maita, Superconductivity and antiferromagnetism in boron-rich lattices, Science 159, 530 (1968).pubmed.ncbi.nlm.nih.gov/17792467
  25. [44]A. Czopnik, N. Shitsevalova, Y. Pluzhnikov, A. Krichikov, Y. Paderno, and Y. Onuki, Low-temperature thermal properties of yttrium and lutetium dodecaborides, J. Condens. Matter Phys. 17, 5971 (2005).DOI
  26. [45]G. P. Kahe, C. R. Tomassetti, I. I. Mazin, A. N. Kolmogorov, and E. R. Margine, Ab initio study of li-mg-b superconductors, Phys. Rev. Mater. 6, 084801 (2022).
  27. [46]E. Karaca, P. J. P. Byrne, P. J. Hasnip, and M. Probert, Prediction of phonon-mediated superconductivity in new Ti-based M2AX phases, Scientific reports 12, 13198 (2022).DOI
  28. [47]A. Bortolozo, O. Sant’Anna, C. Dos Santos, and A. Machado, Superconductivity in the hexagonal-layered nanolaminates ti2inc compound, Solid State Communications 144, 419 (2007).DOI

Paper details

vlnc:W_pfahzwkq
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