betaFeedback

WorksPhysics and Astronomy

Instability-driven segregation of V group transition metals on HEB2_2 grain boundaries influencing enhanced mechanical properties

Abstract

Five metals share a common sublattice in the (Hf,Ta,Ti,Nb,Zr)B2_2 (HEB2_2). Most high-entropy materials are produced as polycrystalline materials in the form of sintered ceramics. The question of which metal species are located at grain boundaries is still unknown. Here, we calculated the grain-boundary energy γGBγ_{GB} of four types of boundaries in five binary diborides and HEB2_2 using density functional theory (DFT). We analyzed the bonding at these boundaries using COHP and sampled cation segregation at six boundaries of HEB2_2, ordered and general, using Monte Carlo / molecular dynamics with a MACE potential that was fine-tuned for this system. According to our results γGBγ_{GB} splits by group: the group V diborides NbB2_2 and TaB2_2 form boundaries with about half the energy of the group IV at three of the four types of GBs, ranging from 0.8 to 2.7 compared to 2.5–4.8~J/m2^2. COHP shows a significant difference at the Fermi level, namely the Me–B bonds of NbB2_2 and TaB2_2 are antibonding at Fermi energy (EFE_F) in the bulk but bonding at the GB, while their integrated strength hardly changes. In HEB2_2 the segregation follows a similar pattern: Nb and Ta are enriched at five of the six grain boundaries, including general boundaries, up to 23–35~at.% at 2000~K, although the size of the atoms would favor Zr and Hf in the expanded sites. A segregated Σ193Σ193 grain boundary can withstand 26 against 20~GPa for a random boundary in tension at 300~K. The grain boundaries of HEB2_2 are composed of group V metals, as suggested by the boundary energies of binaries compounds and their bonding at the Fermi level. Grain boundaries of the (Hf,Ta,Ti,Nb,Zr)B2_2 preferentially include Nb and Ta. This lowers the grain boundary energy and strengthens the boundaries. This consists of the lower boundary energy and changes Fermi level bonding in NbB2_2 and TaB2_2.

Introduction

Transition-metal diborides MeB2\mathrm{MeB}_2 (Me=Ti,Zr,Hf,Nb,Ta\mathrm{Me} = \mathrm{Ti}, \mathrm{Zr}, \mathrm{Hf}, \mathrm{Nb}, \mathrm{Ta}) melt at 3280–3650 K [1, 2, 3], have a hardness of about 20–30 GPa for dense material [2, 4] and resist oxidation up to about 1100 ∘^\circC, while a layer of liquid B2O3\mathrm{B}_2\mathrm{O}_3 covers the surface [2]. This makes them candidates for the leading edges of hypersonic vehicles and for rocket nozzles, and TiB2\mathrm{TiB}_2 is already used for armor and cutting tools [2]. Structure of transition-metal diborides belongs to AlB2\mathrm{AlB}_2-type (P6/mmmP6/mmm) containing flat hexagonal boron nets alternate with metal layers, and the strong B–B and Me–B bonds carry the hardness and the stiffness [3]. The bonding follows the valence electron count (VEC, here per formula unit with three electrons per boron atom [5]). For the group IV metals (Ti,Zr,Hf\mathrm{Ti}, \mathrm{Zr}, \mathrm{Hf}) VEC = 10 and the Fermi level falls into the pseudogap between bonding and antibonding states. The extra electron of the group V metals Nb\mathrm{Nb} and Ta\mathrm{Ta} (VEC = 11) goes into antibonding states [6, 2, 7]. It is the same band-filling argument that puts the hardness maximum of the transition-metal carbonitrides at an intermediate electron count [8].

Mixing the five metals in equal proportions on the metal sublattice gives the high-entropy diboride (Hf,Ta,Ti,Nb,Zr)B2(\mathrm{Hf},\mathrm{Ta},\mathrm{Ti},\mathrm{Nb},\mathrm{Zr})\mathrm{B}_2 (HEB2_2), which is single-phase after spark plasma sintering at 2000 ∘^\circC [9]. Its hardness is close to the rule of mixtures of the binaries [9, 10], and it oxidizes slower than all of them but HfB2\mathrm{HfB}_2 [9]. With VEC = 10.4 it sits just above the optimum, because across a series of high-entropy diborides the calculated hardness is highest at VEC = 10 [5].

But a sintered ceramic is not a single crystal. Diborides densify only at temperatures where grain-boundary and volume diffusion become fast [2], and their strength and hardness depend on the grain size. The strength of ZrB2\mathrm{ZrB}_2 scales roughly with the inverse square root of the grain size [2], the hardness of a high-entropy diboride follows the Hall–Petch law [11], and adding Nb\mathrm{Nb} to (Hf,Zr,Ti,Ta)B2(\mathrm{Hf},\mathrm{Zr},\mathrm{Ti},\mathrm{Ta})\mathrm{B}_2, where it forms a second, Nb-rich boride, refines the grains and raises the hardness [12]. In HEB2_2 the grains grow five times slower than in ZrB2\mathrm{ZrB}_2, which was attributed to sluggish diffusion and to a boundary energy lowered by segregation, neither of them measured [13]. So what a high-entropy boundary is made of, and what it costs, enters both the microstructure and the mechanical response of the ceramic.

DFT gives the grain-boundary energy γGB\gamma_{\mathrm{GB}} of a few tilt boundaries in TiB2\mathrm{TiB}_2 [14] and ZrB2\mathrm{ZrB}_2 [15], and the structure of several [1] tilt boundaries in both [16]. For HfB2\mathrm{HfB}_2, NbB2\mathrm{NbB}_2 and TaB2\mathrm{TaB}_2 there are no available information. Segregation was computed for dilute solutes in these hosts, and it mostly follows size, where large solutes go to expanded boundary sites and small ones to compressed sites [15, 14], which in ZrB2\mathrm{ZrB}_2 sends Nb\mathrm{Nb} and Ta\mathrm{Ta} to the compressed ones [15]. At the expanded Σ6\Sigma6 site of TiB2\mathrm{TiB}_2, though, Nb\mathrm{Nb} and Ta\mathrm{Ta} segregate more strongly than the larger Zr\mathrm{Zr} and Hf\mathrm{Hf} [14].

Machine-learning potentials obtained the same picture at finite temperature, where W\mathrm{W} segregated to the boundaries of ZrB2\mathrm{ZrB}_2 and strengthened nearly all of them [17], and in (Ti,Zr,Hf)B2(\mathrm{Ti},\mathrm{Zr},\mathrm{Hf})\mathrm{B}_2 the smallest cation, Ti\mathrm{Ti}, went to the boundaries [18]. In related five-metal diborides, experiments found W\mathrm{W} and Mo\mathrm{Mo} replacing some of the host metals [19], and Er\mathrm{Er} replacing Nb\mathrm{Nb} [20] at the boundaries. Next to Er\mathrm{Er}, Zr\mathrm{Zr} and Hf\mathrm{Hf} were depleted, and Ta\mathrm{Ta} was not enriched [20]. In HEB₂ itself, the five metals looked homogeneous across a boundary in STEM-EDS maps[19]. For the five-metal diboride there was no γGB\gamma_{\mathrm{GB}} and no calculated segregation.

The electron count suggests a possible answer. At the boundary the boron nets reconstruct and the Me–B and Me–Me bonds are reformed with different lengths and angles. If the extra electron of group V is in the antibonding states of these bonds in the bulk, the reconstruction that moves these states away from the Fermi level is less costly for Nb and Ta than for Ti, Zr and Hf. This should cause a splitting of γGB\gamma_{\mathrm{GB}}. In HEB₂, where both groups share one sublattice, metals with cheaper boundaries should segregate. As Nb, with the lowest boundary energy, does in the NbMoTaW alloy[21] (where Ta, also from group V, is depleted), this would put Nb and Ta on the boundaries even at expanded sites. This is because size alone would send Zr and Hf there. Being a hypothesis about electron count, this can be tested on five binaries and then checked in a mixture. Since a COHP analysis has already supported a bonding origin of solute segregation to a boundary in Fe[22], the same analysis can be done at the boundary itself.

DFT with COHP provides γGB\gamma_{\mathrm{GB}} and bonding at a few boundaries, in cells containing a few hundred atoms, with one decoration of the metal sublattice at a time. Segregation requires thousands of atoms, thermal disorder and many decorations, which DFT cannot accommodate. The deep potential available for nine diborides[18] has no grain boundaries in its training set, and its authors found it less accurate for boundary energies than for other properties. Therefore, we built our own MACE model[23, 24] fine-tuned on a dataset assembled using active learning[25] and complemented by hand with amorphous and surface configurations that sample disordered and under-coordinated configurations of the boundary core. Boundaries were excluded from training, and the potential reproduced DFT γGB\gamma_{\mathrm{GB}} sufficiently well to rank boundaries and metals.

Here we first compute, in DFT, γGB\gamma_{\mathrm{GB}} of four types of grain boundaries (Σ6, Σ7 tilt, Σ7 twist, Σ8) in the five binary diborides and in HEB₂ modeled by special quasirandom structures (SQS)[26]. Then we compare the bonding at the boundary with the bulk by using COHP, and then sample cation segregation at six boundaries of HEB₂, ordered and general, by hybrid Monte Carlo / molecular dynamics. The γGB\gamma_{\mathrm{GB}} value of the group V diborides is approximately half that of the group IV at three out of the four boundaries. The COHP reveals the difference at the Fermi level. The Me–B bonds in NbB₂ and TaB₂ are antibonding at Fermi energy (EFE_{\mathrm{F}}) in the bulk and bonding at the boundary. In HEB₂ Nb and Ta segregate to five out of the six boundaries, including general boundaries. Only at the Σ7 tilt, which offers compressed sites next to expanded ones, does size prevail, and Ti and Zr segregate by radius. Finally, we pull the single crystal and the Σ193 boundary to failure. The segregated boundary is stronger than a random boundary by 31% (25.9 against 19.7 GPa at 300 K).

Methods

DFT calculations

DFT calculations are performed in VASP[27] using the Perdew–Burke–Ernzerhof (PBE) functional[28] and Ti_sv, Zr_sv, Nb_sv, Hf_pv, Ta_pv and B PAW datasets[29, 30]. The kinetic energy cutoff is set to 500 eV with kk-point mesh resolution of 0.16 Å−1^{-1} and Gaussian smearing of width σ=0.05\sigma= 0.05 eV. The convergence criterion for energy for the electronic self-consistent field (SCF) loop is set to 10−610^{-6} eV. PBE was chosen against RPBE[31] and PBEsol[32] by comparison with r2SCAN[33] on NbB2_2, where it gave the closest results on the (0001) surface energy (Supplementary Sec. ??). NbB2_2 was chosen for convergence tests because Nb exhibits a pronounced sensitivity of the band energy to Fermi-level shifts[34], which makes total energies and forces in NbB2_2 particularly sensitive to Brillouin-zone sampling and smearing. So we consider that settings converged for NbB2_2 are therefore conservative for the other diborides. Set parameters for energy cutoff and kk-point mesh resolution give are converged to about 1 meV/atom in absolute energy and better in energy differences (Supplementary Sec. ??). Grain-boundary cells are relaxed with respect to atoms, cell shape and volume are relaxed down to forces of 0.01 eV/Å at kk-point mesh resolution of 0.30 Å−1^{-1}, which fits into the memory of a single V100S GPU, and then recomputed as single points at 0.16 Å−1^{-1}. A relaxation at 0.16 Å−1^{-1} changes γGB\gamma_{GB} by at most 0.004 J/m2^2 (Supplementary Sec. ??). Slabs are relaxed with respect to atoms only, in the cell of the relaxed boundary.

Grain boundary energy γGB\gamma_{GB} (excess energy of GB compared to bulk per surface area), surface energy γs\gamma_s (excess energy of free surface compared to bulk per surface ares) and work of separation WsepW_{sep} (energy it takes to split the GB into two different surfaces) of a configuration with area SS parallel to the boundary or surface are calculated as:

γGB=EGB−Ebulk2S,γs=Es−Ebulk2S,Wsep=2γs−γGB\gamma_{GB} = \frac{E_{GB} - E_{bulk}}{2S}, \qquad\gamma_s = \frac{E_s - E_{bulk}}{2S}, \qquad W_{sep} = 2\gamma_s - \gamma_{GB}

, where EE is the total energy of the considered system (grain boundary, surface, bulk). For HEB2_2 the boundary cells are equimolar (§ 2.2), and EbulkE_{bulk} is their number of atoms times the energy per atom of an equimolar bulk SQS cell (90 atoms, Ti6_6Zr6_6Nb6_6Hf6_6Ta6_6B60_{60}, mean over three decorations).

Construction of grain boundaries and slabs

Four boundaries are considered in DFT: the Σ7\Sigma7 [1] 21.8° symmetric tilt boundary on the (1230) plane, the Σ7\Sigma7 [1] 21.8° twist boundary on the basal plane, a coherent (011ˉ\bar{1}1) twin boundary (78° rotation about [21ˉ\bar{1}10], conventionally labeled Σ8\Sigma8), and an asymmetric (2ˉ\bar{2}11ˉ\bar{1}7) || (12ˉ\bar{2}11ˉ\bar{1}) boundary (47.6° rotation about [101ˉ\bar{1}0], labeled Σ6\Sigma6). For Σ6\Sigma6 and Σ8\Sigma8 the rotation axis lies in the basal plane, where exact coincidence needs a rational c2/a2c^2/a^2, so their Σ\Sigma labels are nominal (construction and coincidence mismatch in Supplementary Sec. ??). The cells hold 105 (Σ6\Sigma6), 78 (Σ7\Sigma7 tilt), 48 (Σ8\Sigma8) and 210 (Σ7\Sigma7 twist) atoms, two boundaries each, and are shown in the top row of Figure 2 for TiB2_2. For HEB2_2 all four grain boundaries are computed in equimolar SQS cells: Σ6\Sigma6 in the 105-atom cell, relaxed in DFT, and the twist, the Σ7\Sigma7 tilt and Σ8\Sigma8 in cells of 210, 390 and 240 atoms, which are relaxed with the MACE potential of § 2.3, since a DFT run of an HEB2_2 cell takes much longer than one of a binary, and then computed as DFT single points. The slabs are cleaves of the boundary cells, so for each γGB\gamma_{\mathrm{GB}} the corresponding γs\gamma_{\mathrm{s}} is used.

Figure 2. The four grain boundaries of the DFT dataset and the accuracy of MACE on them. Top: relaxed cells (TiB2_2 is shown) colored by the excess energy per atom with respect to bulk, boundary planes are shown by dashed lines. The boundary plane is given above each panel. Bottom: γGB\gamma_{\mathrm{GB}} (circles), γsurf\gamma_{\mathrm{surf}} of the cleavage plane (squares) and WsepW_{\mathrm{sep}} (triangles) predicted by MACE against DFT, one panel per boundary, in J/m2^2. Colors mark the composition: group IV binaries in red, group V in blue, the SQS HEB2_2 cell in green (for Σ7\Sigma7 tilt and Σ8\Sigma8 the smaller SQS cells of 78 and 48 atoms, the same cell in DFT and in the potential, Supplementary Sec. ??). Dashed line is y=xy=x. The same plot for the NEP of the active-learning loop is Supplementary Fig. ??.

For MC/MD the four grain boundaries are rebuilt in cells of 1584–2304 atoms, together with one more ordered boundary, a (0552) twin (35.56° rotation about [21ˉ1ˉ0][2\bar{1}\bar{1}0], labeled Σ193\Sigma193, nominal in the same sense as Σ6\Sigma6 and Σ8\Sigma8) in a cell of 2340 atoms, and a three-grain polycrystal of 2919 atoms with random grain orientations, generated with Atomsk[35] (Supplementary Sec. ??).

Preparation of the machine-learning potential

The training set is built in three blocks, shown in Figure 1. Each block is labeled with the DFT protocol described in § 2.1. The seed set comprises the five binary diborides, strained and heated special quasirandom structure (SQS) cells of HEB2_2, and the corresponding elastically deformed configurations. NbB2_2 is overrepresented in the seed set (highlighted here and in Figure 1) because it served as the test system for the methodological experiments, including the convergence tests. It was chosen for this role as the most demanding case among the binary diborides: Nb is one of the elements on which machine-learning interatomic potentials consistently show elevated force errors, a difficulty attributed to the strong response of its electronic band energy to Fermi-level shifts [34]. The configurations that were not included in the initial set were collected using active learning (AL) with a NEP4 potential[36] in GPUMD[37] with implemented D-optimality framework[25] which uses the extrapolation grade γ\gamma of a local environment computed by MaxVol over the NEP descriptors. The same construction is implemented in the MLIP code for active learning of Moment Tensor Potentials [38, 39, 40]. Configurations of NEP MD with 1.5<γ<151.5 < \gamma< 15 are labeled and added until none is left, see middle part in the Figure 1. A configuration for which the extrapolation grade is below 1.5 is considered as interpolation, if the grade is above 15 the configuration is considered as unphysical. 90 iterations of AL is provided, see Supplementary Sec. ??. The loop samples only NEP MD visits, so 324 amorphous HEB2_2 cells and 324 randomly oriented HEB2_2 slabs are added manually for the disordered and under-coordinated environments of the boundary core, see right part in the Figure 1. The final dataset contains 3142 training configurations and 110 validation configurations (Supplementary Table ??) No grain boundaries are included in the dataset.

Figure 1. Assembly of the training set. Left: seed set of bulk NbB2_2 (702 frames), the five binar diborides MeB2_2, Me ∈{Ta,Ti,Zr,Hf,Nb}\in\{\mathrm{Ta}, \mathrm{Ti}, \mathrm{Zr}, \mathrm{Hf}, \mathrm{Nb}\} (280 frames) and SQS HEB2_2 (248 frames). Middle: active-learning loop, in which a NEP potential is (re)trained, MD with this NEP generates candidate configurations, the candidates are screened by the extrapolation grade 1.5<γ<151.5 < \gamma< 15 (D-optimality), the selected ones are labeled in DFT and added to the set; the loop is repeated until nothing passes the selection step. Right: amorphous HEB2_2 (324 frames) and randomly oriented HEB2_2 slabs (324 frames) added by hand. The production potential (MACE) is fine-tuned on the union of all three blocks.

The production potential is a fine-tune of the MACE-MH-1 foundation model[23, 24] on the entire dataset. The NEP of the loop was not used further, as its energy offset drifted over the AL iterations. On grain boundary energies it is 2–3 times less accurate than MACE (Supplementary Sec. ??). The fine-tuning settings and comparison with three other foundation models can be found in Supplementary Secs. ?? and ??. The accuracy results are presented in § 3.2.

Molecular Dynamics simulations

Molecular dynamics simulations are carried out in LAMMPS[41] with the MACE potential. Without domain decomposition one V100S GPU holds about 3600 atoms, which sets the cell sizes. The cation sublattice of each cell is decorated by annealing the Warren–Cowley parameters[42] of the first metal shells to zero. Segregation is sampled by hybrid Monte Carlo/molecular dynamics MC/MD at 2000, 2500 and 3000 K using NPT MD at zero pressure, 120 ps per run. Metropolis swaps are used to exchange the metal species every 100 timesteps, with boron never being swapped. Boundary cation sites are found from the boron coordination number of each metal atom, CNBCN_B for short, as number of B atoms within 3.5 Å; 12 in the bulk, or for Σ6 and Σ8 from the shape of its boron cage. Run settings, site definitions and diffusion coefficients are provided in Supplementary Secs. ??, ?? and ??.

Segregation energies are calculated in two ways. The first method is the swap spectrum of a relaxed 0 K cell with random decoration. In this method every boundary site ss is exchanged with each of the four other species taken from a bulk site bb, and then a local relaxation is performed,

ΔE(s,A→i)=E(i at s,A at b)−E(A at s,i at b),\Delta E(s,A \to i) = E(i\text{ at }s,A\text{ at }b) - E(A\text{ at }s,i\text{ at }b),

where AA is the species that occupied ss. Per site the five energies are shifted to zero mean, Ei(s)=ΔE(s,A→i)−⟨ΔE(s,A→j)⟩jE_i(s) = \Delta E(s,A \to i) - \langle\Delta E(s,A \to j)\rangle_j with ΔE(s,A→A)=0\Delta E(s,A \to A) = 0, so that they do not depend on which cation sits at the site, they are reduced to one number per species by the Boltzmann average:

Eiseg=−kTln⁡⟨e−Ei(s)/kT⟩s∈GB+kTln⁡⟨e−Ei(s)/kT⟩s∈int,E_i^{\mathrm{seg}} = -kT\ln\langle e^{-E_i(s)/kT}\rangle_{s\in GB} + kT\ln\langle e^{-E_i(s)/kT}\rangle_{s\in int},

at T=2000T = 2000 K.

The second way is the McLean isotherm[43] inverted from the MC/MD compositions. A cation can only reach a boundary site by exchanging with another cation. Therefore the isotherm for inversion is the exchange form, which only gives differences in segregation free energies. With the gauge ∑iΔGi=0\sum_i \Delta G_i = 0:

ΔGi=−kT(ln⁡ei−⟨ln⁡ej⟩j),ei=xiGB/xiint,\Delta G_i = -kT\left(\ln e_i - \langle\ln e_j\rangle_j\right), \qquad e_i = x_i^{GB}/x_i^{int},

where xiGBx_i^{GB} and xiintx_i^{int} are the fractions of species ii at the boundary and in the interior. In both routes a negative energy means that the boundary takes the species. The two routes are not the same quantity: the spectrum is the energy of one exchange in a random boundary at 0 K, the isotherm is the free energy of an equilibrated boundary at TT, with the interactions between segregants and the vibrational part is included. The site sets, the outlier cut of the spectrum and the gauge at the Σ7 tilt are given in Supplementary Sec. ??.

γGB\gamma_{GB} after segregation is computed on MC/MD snapshots quenched to 0 K and is compared with γrand\gamma_{rand} of the relaxed initial cell, which carries a random decoration. The expected drop in the swap spectrum, Δγ(T)\Delta\gamma(T), is the free energy that the boundary sites gain when their occupations are in equilibrium with the bulk rather than random, with site energies Ei(s)E_i(s) (Supplementary Sec. ??).

Uniaxial tension is simulated at a strain rate of 5×108 s−15 \times10^{8}\ \mathrm{s}^{-1} on a 432-atom HEB2\mathrm{HEB}_{2} single crystal and on the Σ193\Sigma193 cell. The latter is performed in three segregated cells that are quenched from a 2000 K run and in the same cells with cations permuted randomly, so that only the arrangement of the cations differed within such a pair (Supplementary Sec. ??). Atomic configurations are rendered with OVITO[44].

Bond characterization using the COHP analysis

Bonding in bulk diborides and at grain boundaries is compared using the projected COHP method[45, 46] in LOBSTER 5.1.1[47, 48] and, for the 168-atom twist cells (which give the same γGB\gamma_{\mathrm{GB}} as the 210-atom ones), in gpucohp[49], which is our GPU implementation of the same projection scheme, checked against LOBSTER in Supplementary Sec. ??. Every relaxed structure is recalculated as a single point with the settings described in § 2.1, and with symmetry turned off. The number of virtual energy bands is set to be roughly twice the number of occupied bands. The PAW wavefunctions are then projected onto the pbeVaspFit2015 Slater-type basis with the semicore states of the metals included. The absolute charge spilling is 0.62–1.71%. Bond strength is compared through the −ICOHP integrated up to the Fermi level:

ICOHPij=∫−∞EFCOHPij(E) dE\mathrm{ICOHP}_{ij} = \int_{-\infty}^{E_{\mathrm{F}}} \mathrm{COHP}_{ij}(E)\,\mathrm{d}E

where a pair of atoms ii, jj enters the analysis as a first-shell bond if the distance between them is shorter than 1.25 times the average bond length of the same type (B–B, Me–B or Me–Me) in the bulk diboride. A boundary atom is an atom that is within 1.2 Å of the boundary plane, as determined by the excess of its per-atom energy. The boundary atoms are selected once per boundary in the TiB2\mathrm{TiB}_{2} cell and transferred to the other compositions through one-to-one site matching. We use −ICOHP only as a relative measure to compare Me–B and Me–Me bonds between the five metals and between the bulk and grain boundaries, not as an absolute bond energy. Another descriptor is −COHPij(EF)\mathrm{COHP}_{ij}(E_{\mathrm{F}}), the value of the curve at the Fermi level, which indicates whether the states of that bond at EFE_{\mathrm{F}} are bonding or antibonding. Its dependence on the energy grid, the integration method and for the HEB2\mathrm{HEB}_{2} twist, on the MACE relaxation of the cell is checked in Supplementary Sec. ??.

Results and Discussion

Grain-boundary and surface energetics from DFT

DFT values of γGB\gamma_{\mathrm{GB}} for the four types of grain boundaries, five binary diborides and the HEB2\mathrm{HEB}_{2} SQS cells are collected in Table 1. The surface energies and the works of separation of the same cells, which serve below only as a test of the potential, are given in Supplementary Table ?? and Fig. ??.

BoundaryGroup IVGroup V
TiB2TiB_2ZrB2ZrB_2HfB2HfB_2NbB2NbB_2TaB2TaB_2HEB2HEB_2
Σ6\Sigma6, asymmetric4.784.254.512.732.183.65
Σ7\Sigma7 tilt, symmetric2.652.582.742.362.262.56
Σ8\Sigma8, coherent twin2.872.532.641.370.832.28
Σ7\Sigma7 twist, basal2.802.592.661.270.841.95

Table 1. DFT grain-boundary energies γGB\gamma_{GB} of four grain boundaries in the five binary diborides and in one SQS decoration of HEB2HEB_2, in J/m2^2. For HEB2HEB_2, Σ7\Sigma7 tilt and Σ8\Sigma8 are the equimolar cells of 390 and 240 atoms. These and the HEB2HEB_2 twist are DFT single points on MACE-relaxed cells.

Table 1 splits by group. For the Σ6\Sigma6, Σ8\Sigma8 and twist boundaries of the group IV diborides give γGB\gamma_{GB} of 2.5–4.8 J/m2^2.

The group V diborides gives γGB\gamma_{GB} of 0.8–2.7 J/m2^2, i.e. roughly two times lower, with TaB2TaB_2 always the lowest. The only grain boundary where the groups do not separate is the symmetric Σ7\Sigma7 tilt with the γGB\gamma_{GB} ranging from 2.26 to 2.74 J/m2^2 for all five metals, a spread of 0.5 J/m2^2. Within a group the order of the metals holds at three out of four grain boundaries (Ti>Hf>ZrTi > Hf > Zr at Σ6\Sigma6, Σ8\Sigma8 and the twist, Hf>Ti>ZrHf > Ti > Zr at the tilt; Nb>TaNb > Ta everywhere), and at the three grain boundaries that split the differences inside a group are 0.2–0.6 J/m2^2 against 1.2–2.6 J/m2^2 between the groups.

The Σ7\Sigma7 tilt is computed before on the same plane, and our values agree with the earlier DFT results, namely 2.65 against 2.64 J/m2^2 for TiB2TiB_2[14] and 2.58 against 2.52 J/m2^2 for ZrB2ZrB_2[15] (reference values converted from meV/Å2^2).

Previously studied Σ6\Sigma6[14, 15] is a symmetric tilt about [011ˉ\bar{1}0], which differs from our asymmetric one, and their Σ8\Sigma8 is the same twin but off-stoichiometric, and its γGB\gamma_{GB} depends on the chemical potential of boron (0.40–1.44 J/m2^2 in TiB2TiB_2, 1.35–2.48 J/m2^2 in ZrB2ZrB_2), so we do not compare it with our stoichiometric twin. The group split goes with the extra valence electron of NbNb and TaTa. At the boundary the Me–B bonds have to reconstruct, and the COHP analysis of § 3.4 looks for the trace of that electron in the bonding.

For the six-element cells, with one SQS decoration per boundary, HEB2HEB_2 gives γGB=3.65\gamma_{GB} = 3.65 J/m2^2 for Σ6\Sigma6 against a average across five binary diborides is 3.69 J/m2^2, 2.56 against 2.52 J/m2^2 for the Σ7\Sigma7 tilt, 1.95 against 2.03 J/m2^2 for the twist and 2.28 against 2.05 J/m2^2 for Σ8\Sigma8. The second, smaller SQS cells of the tilt and of Σ8\Sigma8 (78 and 48 atoms, not equimolar, Figure 2) show how much of this belongs to the decoration. They give γGB\gamma_{GB} of 2.36 and 2.60 J/m2^2, 0.2–0.3 J/m2^2 away from the equimolar cells. Within this scatter, the random HEB2HEB_2 boundary cost the mean of the binary diborides at all four considered grain boundaries. Chemical disorder by itself does not move γGB\gamma_{GB} by more than this.

The random decorations of the much larger MACE cells of § 3.3 gives γGB\gamma_{GB} of 3.76, 2.44, 2.02 and 2.24 J/m2^2, which is in the same order (Figure 4(b)), and are within 0.12 J/m2^2 of the DFT values, which is below the error of the potential on γGB\gamma_{GB} (§ 3.2). The same comparison applies to the closest pair of boundaries in the DFT set. The two grain boundaries of the 105-atom Σ6\Sigma6 cell are only 8.4–9.0 Å apart, compared to 31 Å in the MACE cell. The difference between these two values is 0.11 J/m2^2, so their interaction within the DFT cell is within the accuracy of the potential.

Figure 4. Migration of atoms and boundary energy at the six grain boundaries. (a) Self-diffusion coefficients of boron (yellow) and metal (purple) atoms at the boundary (dashed lines, squares) and in the grain boundary (solid lines, circles). Points below the resolution of 5×10−9 cm2/s5 \times10^{-9}\,\mathrm{cm}^{2}/\mathrm{s} are placed on a common floor at the bottom of the axis. Grey band shows the melting of the polycrystal. (b) γGB\gamma_{\mathrm{GB}} of MC/MD snapshots quenched to 0 K, by run temperature, and of the relaxed initial cell of the runs, with a random decoration and the boundary geometry as built (white bar, right of the axis break). The black line is the γGB\gamma_{\mathrm{GB}} predicted from the swap spectrum, γrand+Δγ(T)\gamma_{\mathrm{rand}} + \Delta\gamma(T), with a bootstrap band over sites. Molten runs are not shown.

Accuracy of the potential on grain boundaries and surfaces

No grain boundary is present in the training set. Therefore, the boundaries described in the Table 1 and their cleavage surfaces are testing the potential of a configuration that the model has not encountered (Figure 2, Table 2c). The surfaces are trained as randomly oriented HEB2\mathrm{HEB}_{2} slabs, so the test is weaker for them. Over the 24 boundary cells (four boundaries in six compositions), the RMSE of γGB\gamma_{\mathrm{GB}} is 0.17 J/m2^{2}, about 4% of the 0.8–4.8 J/m2^{2} range. WsepW_{\mathrm{sep}} is reproduced to 0.17 J/m2^{2} and the (101ˉ0)(10\bar{1}0) surfaces to 0.16 J/m2^{2}. The order of the groups and the metals within each group from Table 1 is preserved. On the six-element cells (green in Figure 2), MACE stays within 0.02–0.23 J/m2^{2} of DFT for all twelve quantities. The same level as for binary diborides. One systematic error is the compression of the γGB\gamma_{\mathrm{GB}} range: the low-energy Nb and Ta boundaries come out 0.2–0.45 J/m2^{2} above DFT at Σ6\Sigma6, Σ8\Sigma8 and the twist (within 0.1 J/m2^{2} at the tilt). So, the MLIP averages are biased against low boundary energies. The potential is good enough to determine the boundaries and metals, but it is not able to distinguish differences below 0.2 J/m2^{2} between individual grain boundaries. The errors in the bulk configurations of the training data (2.2 meV/atom and 23 meV/Å) and in the elastic moduli, and the NEP of the active-learning loop, 2–3 times worse on γGB\gamma_{\mathrm{GB}}, discussed in Supplementary Section ??.

Table 2: RMSE of MACE against DFT. Sections a and b represent validation data, with different configuration types present during training. Section c is the test set, where grain boundaries are not included in the training, and the surfaces are represented only by randomly oriented HEB2_2 slabs. The NEP of the active-learning loop on the same sets is given in Supplementary Table ??. EE (meV/at) is energy per atom RMSE, FF (meV/Å) is the force per atom RMSE, σ\sigma (GPa) is stress of the configuration calculated from virial. Elastic moduli CijC_{ij} in (b) are calculated from finite deformation methods along specific axes. Bulk modulus B0B_0 is calculated from finite deformation along three main axes.

a. Validation
EE (meV/at)FF (meV/Å)σ\sigma (GPa)
1. Relaxed decorations (n=35n=35)2.269.90.143
2. Equation of state, ±2\pm2–6% (n=60n=60)2.4214.70.480
3. Uniaxial strain, ±2\pm2–8% (n=70n=70)2.268.70.192
4. Shear strain, 2–6% (n=48n=48)1.9710.10.170
5. Size check, 54-atom cells (n=5n=5)2.6712.60.215
6. Thermal MD, 300–2300 K (n=72n=72)2.2942.60.196
7. Coupled strain + MD (n=16n=16)1.7725.50.184
8. Amorphous (n=36n=36)11.29191.00.750
9. Generated surfaces (n=36n=36)8.28174.90.164
All bulk subsets (1–7) (n=306n=306)2.2423.10.268
b. Elastic moduli
B0B_0 (GPa)C11C_{11} (GPa)C33C_{33} (GPa)
1. Bulk and axial (n=10n=10)1.16.614.2
C44C_{44} (GPa)C66C_{66} (GPa)
2. Shear (n=8n=8)9.411.2
c. Test
γGB\gamma_{\mathrm{GB}} (J/m2^2)γsurf,clv.\gamma_{\mathrm{surf, clv.}} (J/m2^2)WsepW_{\mathrm{sep}} (J/m2^2)
1. Σ\Sigma6, asymmetric (n=6n=6)0.2090.0460.162
2. Σ\Sigma7 tilt, symmetric (n=6n=6)0.0650.0810.119
3. Σ\Sigma8, coherent twin (n=6n=6)0.2170.1050.067
4. Σ\Sigma7 twist, basal (n=6n=6)0.1580.1530.256
All boundaries (1–4) (n=24n=24)0.1730.1040.166
γsurf(10‾10)\gamma_{\mathrm{surf}}(\overline{10}10) (J/m2^2)
5. Free surfaces (n=5n=5)0.156

Table 2. RMSE of MACE against DFT. Sections a and b represent validation data, with different configuration types present during training. Section c is the test set, where grain boundaries are not included in the training, and the surfaces are represented only by randomly oriented HEB2_2 slabs. The NEP of the active-learning loop on the same sets is given in Supplementary Table ??. EE (meV/at) is energy per atom RMSE, FF (meV/Å) is the force per atom RMSE, σ\sigma (GPa) is stress of the configuration calculated from virial. Elastic moduli CijC_{ij} in (b) are calculated from finite deformation methods along specific axes. Bulk modulus B0B_0 is calculated from finite deformation along three main axes.

HEB2_2 grain boundary segregation inferred with MACE

Figure 3 collects the MC/MD compositions of six boundaries at 2000–3000 K, the segregation energy at 2000 K and the map of site volume at each boundary. The boundaries can be divided into two types based on this map, with Σ\Sigma8 in the middle, and the type determines which metals will segregate.

Figure 3. Cation segregation at the six grain boundaries of HEB2 from MC/MD simulations with MACE. (a) Site volume ΔV\Delta V of the boundary cation relative to the interior cation of the same species, with the color scale in Å3^3 and saturated at ±3\pm3 Å3^3. The data is from a slice of a relaxed 0 K cell. The boundary plane is shown by green dashed line, metal atoms of the different grains are orange, purple and green, boron is white. (b) Metal fraction at the boundary sites against temperature. The dotted line is the nominal 20 at.%, error bars are block standard errors. Gray band shows the region where simulated cell of polycrystalline HEB2 is molten at considered conditions of 3000 K. (c) Segregation energies at 2000 K from the swap spectrum (open circles, error bars from a bootstrap over sites) and from the McLean isotherm (filled squares). A negative value indicates that the boundary includes the species. For Σ\Sigma7 tilt the two site classes are shown separately, the compressed CNB=10C_{NB}=10 sites (10 boron neighbours instead of 12 in the bulk) (blue, down triangles) and the expanded CNB=14C_{NB}=14 sites (red, up triangles), colored as in (a). McLean values for Hf and Zr on the CNB=10C_{NB}=10 sites are absent, since these metals never occupy them. Open symbols represents the spectrum, and filled symbols represent isotherm. The columns are divided into chemistry-driven (Σ\Sigma6, Σ\Sigma7 twist, polycrystal, Σ\Sigma193), mixed (Σ\Sigma8) and geometry-driven (Σ\Sigma7 tilt). The y-axis scales in (c) are different between panels, and tilt values are gauged over Ti, Nb and Ta only (see Supplementary Sec. ??). Swap spectra are provided in Supplementary Fig. ??.

Where every site at the boundary is larger than in the bulk, segregation is controlled by chemistry. This is the case for Σ\Sigma6, Σ\Sigma7 twist, Σ\Sigma193 and the general boundaries of the polycrystalline sample: at all of them Nb and Ta are enriched and Hf is depleted, even though the expanded sites would prefer Zr and Hf. Which of Ti and Zr is also depleted depends on the boundary. Ti is depleted at Σ\Sigma6 and Σ\Sigma193 (by 12.5 and 12.9 at.% at 2000 K), while Zr at the twist (by 7.2 at.%). The effect is strongest at the Σ\Sigma7 twist boundary, 35 at.% of Nb and 35 at.% of Ta at 2000 K against 7 at.% each of Zr and Hf. With temperature it weakens at the twist (Nb and Ta at 25 at.% at 3000 K) and at Σ\Sigma193 (from 26 and 30 to 23 and 25 at.%, respectively for Nb and Ta), while at Σ\Sigma6 it stays at 25–29 at.% up to 3000 K. The Σ\Sigma193 runs started from a random decoration, so this enrichment is reached by the MC. The polycrystal has a similar behavior to the ordered grain boundaries with 23 and 24 at.% of Nb and Ta at 2000 K. Its boundaries are less ordered compared to others, and it melts at 3000 K; with 2919 atoms in three grains, the grains are approximately 2 nm in size and about 450 out of 973 cations (46%) are located at the boundary sites, so the interior acts as a small reservoir. A larger cell would be beyond the single-GPU limit in § 2.4, since the mean site volume is 18% higher than the bulk value, and Nb and Ta occupy these sites equally, so the enrichment does not require an ordered boundary for its appearance.

The enrichment is shown in the site energies (Figure 3(c); the spectrum is taken over the energy-defined boundary sites in Supplementary Section ?? and the isotherm over the structural ones). At all four grain boundaries of this kind, both routes give negative energies for Nb and Ta, ranging from −0.03-0.03 to −0.17-0.17 eV from the spectrum and −0.05-0.05 to −0.15-0.15 eV from the isotherm. Positive energies are found for Ti (except for the isotherm at the twist, −0.004-0.004 eV), which agree within 0.02 eV everywhere except the twist. Hf and Zr showed different results: the spectrum put them within 0.06 eV of zero (except Zr in the polycrystal, −0.12-0.12 eV), while the isotherm ranged up to +0.15+0.15 eV. The routes agree in sign for metals that segregated and differed in size by a factor of 2–3, in both directions. For Nb and Ta the isotherm is the larger of the two at Σ6\Sigma6 and the twist, and the smaller at Σ193\Sigma193 and in the polycrystal. The two are not the same in quantity (§ 2.4). Te spectrum is one exchange in a random boundary at 0 K and the isotherm is an equilibrated boundary at 2000 K. We take only the sign from them together.

Where the boundary between expanded and compressed sites is located, geometry takes over. At the Σ7\Sigma7 tilt boundary, there are two sites with an equal number of boron atoms with coordination numbers of 14 and 10, and atomic volumes that are 18% larger and 13% smaller than those in the bulk crystal. The runs fill these sites according to the radius of the segregant. At 2000 K Zr and Ti accounts for 40 and 39 at.% of the boundary, respectively, i.e. each occupies most of its own class of sites. Zr occupies the large site, while Ti occupies the smaller one. Nb and Ta are depleted to 5–8 at.% at this grain boundary. Site energies indicate the same trend, and are an order of magnitude stronger than at any other boundary. On the CNB=14CN_{B}=14 sites Zr is at −0.97-0.97 eV and Hf at −0.70-0.70 eV (−0.85-0.85 and −0.59-0.59 eV from the isotherm) Hf would also occupy the large site, but it is lost to Zr. On the CNB=10CN_{B}=10 sites Ti is at −0.26-0.26 eV (−0.22-0.22 eV), and Hf and Zr are 1.4 and 1.7 eV uphill. Σ8\Sigma8, with a weaker splitting of the site volumes, is a mixed case. At 2000 K Zr is enriched to 27 at.% and Ti is depleted to 9 at.%, as by size with isotherm energies of −0.07-0.07 and +0.14+0.14 eV. Nb and Ta are enriched to 23 and 25 at.%, as at the chemical kind. The entire spectrum lies within ±0.03\pm0.03 eV. The effect weakens with temperature at a different rate for each element. Nb and Ta stop segregating towards Σ8\Sigma8 at T=2500T=2500 K, while Zr remains at 27–30% up to 3000 K. At Σ7\Sigma7, all five metals return to within 7% of their nominal value at 3000 K.

The tilt site energies can be compared with DFT without new calculations, as the sites of the pentagonal and heptagonal rings of this boundary were calculated for dilute solutes previously. Re-referencing to Ti, the expanded CNB=14CN_{B}=14 site gives −1.02-1.02, −0.75-0.75 and −0.16-0.16 eV for Zr, Hf, and Nb, respectively, compared to −1.04-1.04, −0.85-0.85 and −0.13 eV at the heptagonal site of TiB2_2[14]. Re-referencing to Zr, the compressed CNB=10CN_{\mathrm{B}} = 10 site gives −1.59 and −1.63 eV for Nb and Ta, respectively, compared to −1.49 and −1.36 eV at the pentagonal site of ZrB2_2[15]. The hosts differ (dilute solutes in a binary vs HEB2_2), and the comparison covers the geometric sites where the energies vary by 0.1–1 eV. The chemical energies, −0.03 to −0.17 eV, are comparable to the error of the potential on γGB\gamma_{\mathrm{GB}} converted to one boundary cation (−0.06 to 0.2 eV), so we can use the potential to determine the sign of Nb and Ta enrichment but not their magnitude. The potential bias on the binaries (§ 3.2) is working against Nb and Ta.

The loss of segregation with temperature is accompanied by boron diffusion across the sample volume towards the grain boundaries (see Figure 4(a)). At 2000 K, boron at the boundary diffuses at a rate of D=8×10−8D = 8 \times10^{-8}–10−6 cm2/s10^{-6}\,\mathrm{cm}^{2}/\mathrm{s} at every boundary, except for the Σ7\Sigma7 tilt. This is at least an order of magnitude faster than the rate of diffusion in the grain interior (circles on the Figure 4(a)). The metals at the boundaries stay below 10−7 cm2/s10^{-7}\,\mathrm{cm}^{2}/\mathrm{s}, and in the interior, mostly below the resolution. The exception to this is the twist, which has a diffusion coefficient of up to 10−6 cm2/s10^{-6}\,\mathrm{cm}^{2}/\mathrm{s} at 2500 K. Its two grains slide against each other along the basal plane as rigid blocks. The local-sublattice reference in Supplementary Section ?? reduces this motion partially, but at 3000 K the residual sliding also sets the diffusion coefficient of the metal in the twist region. The cation sublattice, on which the compositions in Figure 3(b) were measured, is static, while boron is the element that rearranges at the boundary. At the Σ7\Sigma7 tilt boundary, boron diffusion was below the resolution until 2500 K, when it jumped to 2.9×10−6 cm2/s2.9 \times10^{-6}\,\mathrm{cm}^2/\mathrm{s} at 3000 K. This is the same temperature where the size sorting collapsed. Two factors contribute to this collapse and are not discussed here. The simple Boltzmann weakening of the fixed segregation energy with TT, and the growth of the boundary itself. As the number of boundary sites increases from 24 at 2000 K to 36 at 3000 K (and the twist changes from 24 to 52), the new sites diffuse the old ones, diluting them. Boron diffusion alone cannot reverse segregation: at Σ8\Sigma8, boron diffuses at the boundary from 2000 K on, and zirconium remains enriched to 27-30% up to 3000 K. In the polycrystal, at 3000 K, boron and metal both diffuse at 10−510^{-5}–10−4 cm2/s10^{-4}\,\mathrm{cm}^2/\mathrm{s} at the boundary and in the grains, which means it is molten.

Figure 4(b) shows the effect of segregation on the boundary energy at different temperatures. At 2000 K, every boundary has a lower energy than the random case: by 0.17 J/m2^2 at the twist, 0.47 J/m2^2 at Σ8\Sigma8, and 0.86-1.28 J/m2^2 at Σ7\Sigma7 tilt, Σ193\Sigma193, Σ6\Sigma6, and in the polycrystal. With increasing temperature, γGB\gamma_{\mathrm{GB}} increases again. At 3000 K, it returns to the random value at Σ6\Sigma6 (3.77 vs 3.76 J/m2^2) and is above it for the twist (2.50 vs 2.02 J/m2^2), but it remains 0.17-0.51 J/m2^2 lower than the random value for the tilt, Σ8\Sigma8 and Σ193\Sigma193 (2.27 vs 2.44 J/m2^2, 2.02 vs 2.24 J/m2^2, and 3.04 vs 3.55 J/m2^2 respectively). The twist still has 25% Nb and Ta at 3000 K, so its excess energy over the random value is not due to chemistry. The quench retains some of the thermal disorder in the boundary, which is not separated from the chemistry in the other samples either. The prediction from the swap spectrum accurately reproduces the twist and Σ7\Sigma7 tilt at 2000 and 2500 K within a range of 0.07-0.16 J/m2^2, which is comparable to the spread of the snapshots. Additionally, Σ8\Sigma8 is also accurately predicted at 3000 K.

At these three boundaries, the drop of γGB\gamma_{\mathrm{GB}} is determined by the site preferences at the fixed 0 K boundary. At 3000 K, the tilt and twist lie above the prediction based on thermal disorder, which is kept by the quench. At Σ6\Sigma6 and Σ193\Sigma193, the measured drop at 2000 K is larger than predicted by 0.5 and 0.4 J/m2^2, respectively, and in the polycrystal, it is 0.9 J/m2^2 larger. The prediction accounts only for the chemical preferences of the sites at a fixed boundary. We attribute the rest to the reconstruction of these boundaries, mainly their boron network, during the process, which cannot be captured by the spectrum, since boron never swaps and relaxes only locally around an exchange. MC also swaps cations in grain interiors, and any short-range order they develop is counted as boundary energy, because the bulk reference is random. The boundaries that the prediction produces indicate that this area is small: its interior is similar in nature to that of the other cells. In this region, the prediction, which does not include an interior term, leaves a maximum of 0.16 J/m2^2 unexplained. For the polycrystal, this can be verified. Its 2500 K geometry with the cation positions randomized (not shown in Figure 4(b)) has a GB energy of 3.42 J/m2^2, so when the geometry is annealed alone, it reduces the as-built GB energy by 0.85 J/m2^2. From this reference, the predicted value (3.08 J/m2^2 at 2000 K) matches the measured value (2.99 J/m2^{2}) within 0.1 J/m2^{2}. Therefore, roughly two thirds of the reduction is due to reconstruction, and one third is due to chemistry. The reference geometry is at 2500 K, so the proportion is valid to the extent that the GB boundary at 2000 and 2500 K is the same. For ordered boundaries, the same randomization was not applied.

The two types of boundaries, with Σ8\Sigma8 in between, separate different metals. The type of boundary depends on its site volume map, which can be ordered or general. For a polycrystal, the chemical type is most important, as general boundaries belong to this category. The segregation of Nb and Ta agrees in sign with the density functional theory (DFT) energetics of ordered boundaries. The γGB\gamma_{GB} value of the group V diboride is lower than that of group IV, so a boundary enriched with Nb and Ta is cheaper. However, this agreement is only qualitative. When converted to energy per boundary cation, the difference between the binary duborides is several times greater than the segregation energies in Figure 3c. It is discussed in § 3.1. The role of the extra valence electrons of Nb and Ta in the bonding is discussed in § 3.4.

The geometric kind is what was previously found in diborides, in ZrB2_{2} with Nb and Ta taking the compressed boundary sites[15], and in (Ti,Zr,Hf)B2_{2}, the smallest cation, Ti, segregates[18]. Both of these are in line with the size. Chemical species such as Nb and Ta occupy expanded sites, but their size does not predict this. A binary boundary that is cheaper does not always determine segregation in a lattice. In the high-entropy carbide, with the same five metals, a Monte Carlo simulation with a universal MACE model (not fine-tuned) at 300 and 2000 K puts Ti and Zr on the Σ5\Sigma5 boundary[50]. Although there too, the group V binary species have the cheaper grain boundaries. The values for TiC, ZrC, and HfC are between 1.88 and 2.96 J/m2^{2} in DFT[50], while the values for NbC and TaC are 2.38 and 2.69 J/m2^{2}. In the diboride, the cheaper boundaries for Nb and Ta coincide with the change in bonding at EFE_{F}, as discussed in § 3.4. We did not check whether the carbide lacked this.

The measurement of host metal at the HEB2_{2} boundary does not show enrichment. STEM-EDS maps across the boundary of HEB2_{2}, created by reactive spark plasma sintering with a final 10-minute hold at 2000 °C (2273 K), show all five metals to be homogeneous[19]. The maps shown are qualitative and do not provide a profile of the composition at a given point. It is not known whether the sample has reached equilibrium, as in the case of MC samples. A profile of the composition across the boundaries of an annealed sample would provide a way to test the prediction.

COHP analysis of bonding at the grain boundaries

To analyze chemical bonding in grain boundaries in HEB2_{2} two types of descriptors are used, shown in Figure 5(a) for the example of B–B bond in NbB2_{2}. These descriptors are –COHP(EFE_{F}), whose sign tells whether the states at the Fermi level are bonding or antibonding, and –ICOHP, the integral up to EFE_{F}, i.e. the total strength of the bond. For the bulk four types of bonds can be considered, see Figure 5(b), namely Me–Me in-plane (denoted as 1 in Figure 5(b)), Me–B (2), B–B (3) and Me–Me inter-plane (4). At the grain boundary Σ8\Sigma8 also four types of bonds are considered, as shown in right panel of Figure 5(b), namely MeGB\mathrm{Me}_{GB}–Mebulk\mathrm{Me}_{bulk} (1), MeGB\mathrm{Me}_{GB}–BGB\mathrm{B}_{GB} (2), BGB\mathrm{B}_{GB}–BGB\mathrm{B}_{GB} (3) and BGB\mathrm{B}_{GB}–Bbulk\mathrm{B}_{bulk} (4). At the base of the twist, the last layer is empty, as the next layer along the cc-axis lies outside the first layer, and the bonds between the MeGB\mathrm{Me}_{GB} and Mebulk\mathrm{Me}_{bulk} atoms are along the cc-direction.

COHP curves of the bonds in the bulk and at the grain boundary

Figure 5. COHP curves of the bonds in the bulk and at the grain boundary. (a) Example: the B–B bond in bulk NbB₂; the shaded area up to EFE_F is −ICOHP, the inset shows −COHP(EFE_F), a positive value means bonding states at the Fermi level. (b) The four bond types of the bulk crystal (left) and the four bond classes at the Σ8 (011ˉ\bar{1}1) boundary (right, boundary atoms outlined with dashes, boundary slab in grey) for TiB₂; the color of a bond is its −ICOHP. (c) −COHP against energy for each bond type, upper row bulk, lower row Σ8; the curve of a composition is the mean over all bonds of that class, insets magnify the window E−EF=±0.8E - E_F = \pm0.8 eV. Atom colors of (b) and composition colors of (a, c) as in the two keys.

The COHP curves of all compositions are given in Figure 5(c) for Σ8\Sigma8 and for all four boundaries in Supplementary Fig. ??; the descriptors, averaged over Σ6\Sigma6, Σ7\Sigma7 tilt, Σ7\Sigma7 twist and Σ8\Sigma8, are collected in Figure 6, and per boundary in Supplementary Table ??.

Figure 6. Bond descriptors on going from the bulk to the grain boundary. (a) The Σ6, Σ7 tilt, Σ7 twist and Σ8 cells and the bulk crystal. Boundary atoms are marked with dashed rings, boundary region in grey. (b) −COHP(EF)-\mathrm{COHP}(E_F) and (c) −ICOHP-\mathrm{ICOHP} of the B–B, Me–B and Me–Me bonds in the bulk and at the boundary for the five binary diborides and HEB2_2; the boundary point is the mean over all bonds of that class at the four boundaries (the twist, with the largest boundary area, contributes most of them and has no BGBB_{GB}–BbulkB_{bulk} bonds), the violins in (c) show the distribution over individual bonds. The two halves of a marker are colored by the two atoms of the bond (metal purple, boron yellow), so B–B, Me–B and Me–Me read from the marker. The three types of bonding are labeled explicitly for the TiB2_2 panel of (c). Double dashed lines join the bulk bond to the boundary–boundary bonds, dashed-solid lines are referred to the boundary–bulk bonds. Per-boundary values are presented in Supplementary Table ??.

By ICOHP the bonds at the boundary hardly differ from the bulk ones. Me–B bonds at the boundary are stronger than in the bulk by 0.10–0.27 eV (from 1.59–1.87 eV in the bulk to 1.70–2.07 eV at the boundary), Me–Me bonds by 0.01–0.13 eV, for all six compositions. BGBB_{GB}–BbulkB_{bulk} bonds are as strong as the bulk ones (5.8–6.6 against 5.8–6.7 eV), and the mean of the BGBB_{GB}–BGBB_{GB} bonds moves by −0.17 to +0.09 eV, with a tail of a few bonds with −ICOHP-\mathrm{ICOHP} down to 2.9 eV at Σ6 and the Σ7 tilt (see violins in Figure 6(c)). All these shifts are smaller than the spread of −ICOHP-\mathrm{ICOHP} over the bonds of one class (standard deviation 0.4–0.9 eV).

The difference between groups IV and V can be observed using −COHP-\mathrm{COHP} at the Fermi level. In the bulk TiB2_2, ZrB2_2 and HfB2_2 the states at EFE_F are bonding for both Me–B (−COHP(EF)=0.009-\mathrm{COHP}(E_F) = 0.009–0.016 eV−10.016\ \mathrm{eV}^{-1}) and for Me–Me (0.0200.020–0.024 eV−10.024\ \mathrm{eV}^{-1}). In contrast, in NbB2_2 and TaB2_2 the Me–B bonds are antibonding at EFE_F with values of −0.027-0.027 and −0.013 eV−1-0.013\ \mathrm{eV}^{-1}, respectively. The Me–Me bond is antibonding for Nb–Nb (−0.010 eV−1-0.010\ \mathrm{eV}^{-1}) and close to zero for Ta–Ta (+0.005+0.005 eV−1^{-1}). For the B–B bond, the situation is reversed. Only NbB2_2 has bonding states at EFE_F with −COHP(EF)=+0.038-\mathrm{COHP}(E_F) = +0.038 eV−1^{-1}, as shown in Figure 5(a). The other five compositions have weakly antibonding states (−0.005-0.005 to −0.019-0.019 eV−1^{-1}). Overall, it can be seen that the Fermi level of group V is located in antibonding Me–B states in the bulk. This means that the extra electron in Nb and Ta occupies into states that weaken the Me–B bond.

At the grain boundary the signs change. For group IV, the Me–B bonds at the boundary become 4–8 times stronger, where EFE_F is ranging from 0.064 to 0.075 eV−1^{-1} and the Me–Me bonds become 2–3 times stronger. For group V, the antibonding states disappear: the Me–B goes from −0.027-0.027 to +0.011+0.011 for Nb and from −0.013-0.013 to +0.022+0.022 eV−1^{-1} for Ta, and the Me–Me goes from −0.010-0.010 to zero (+0.001+0.001) for Nb and from +0.005+0.005 to +0.021+0.021 eV−1^{-1} for Ta. The BGB_{GB}–BGB_{GB} bonds of group V at the grain boundary are also stronger than B–B in the bulk with EFE_F of +0.062+0.062 against +0.038+0.038 for Nb, +0.039+0.039 against −0.013-0.013 eV−1^{-1} for Ta, while the bonds fo group IV are between −0.012-0.012 and +0.001+0.001 eV−1^{-1}. The size of this difference depends on the grain boundary, and roughly follows the change in γGB\gamma_{GB}. The shift of the group V Me–B states is largest at the twist (+0.050+0.050 eV−1^{-1}, mean of Nb and Ta), followed by Σ6\Sigma6 and Σ8\Sigma8 (+0.039+0.039 and +0.035+0.035), and smallest at Σ7\Sigma7 tilt (+0.010+0.010 eV−1^{-1} with Nb–B remaining antibonding at −0.018-0.018 eV−1^{-1}). The ratio of γGBgroupV\gamma_{GB}^{groupV} to γGBgroupIV\gamma_{GB}^{groupIV} from Table 1 runs 0.39, 0.54, 0.41 and 0.87 in the same order. This trend holds for the two ends (twist and tilt), but for Σ6\Sigma6 and Σ8\Sigma8 they are reversed, so this is only a general trend with four grain boundaries. The tilt has the smallest swing and no group splitting, and is also the boundary where segregation in § 3.3 is determined by geometry. For HEB2_2 the descriptors averaged across all metals behave like a group IV (Me–B range from 0.008 to 0.043 eV−1^{-1}), and we did not differentiate them by metal species weithin HEB2_2.

In short, ICOHP shows that the bonds at the grain boundary are as strong as they are in the bulk for all six compositions. The groups differ at the Fermi level, where in the bulk for group V diborides EFE_F corresponds to the antibonding Me–B (and Me–Me for Nb) states. At the grain boundary these states either leave the Fermi level or become bonding states. For group IV there are no antibonding states to remove. The descriptor is read based on its sign, rather than its magnitude. Group IV gains more bonding at EFE_F at the boundary compared to group V, and NbB2_2 is more antibonding in the bulk than TaB2_2, but its boundaries are more expensive (γGB(NbB2)>γGB(TaB2)\gamma_{GB}(\mathrm{NbB}_2) > \gamma_{GB}(\mathrm{TaB}_2)). So, the argument relies on the antibonding states in group V in the bulk that are lost at the boundary. Between boundaries, the magnitude of the shift of group V follows how much cheaper its boundaries are in DFT (§ 3.1). The same metals, Nb and Ta, segregate chemically in HEB2_2 (§ 3.3). This is a correlation between Fermi-level descriptor and the boundary energy. COHP(EF)\mathrm{COHP}(E_F) is not an energy and we do not derive γGB\gamma_{GB} from it.

Strength of HEB2_2 and of a segregated boundary

We first performed tensile deformation on the HEB2_2 single crystal (Figure 7(a,Figure 7b)). Along the [1] direction the ultimate tensile strength (UTS) is 46.4 GPa at 300 K and sample fails at a strain of 20%. Along the [01\bar{1}0] direction the UTS is lower, 32.9 GPa at a strain of 10%, followed by a plateau of a few percent before failure. As temperature increases both values decrease: σmax⁡\sigma_{\max} along the [1] decreases to 33.7 GPa and along the [01\bar{1}0] it decreases to 27.6 GPa at 1900 K. The Young’s modulus of HEB2_2 is the average of the five binary diborides values (Figure 7(c,Figure 7d)): 365 against 370 GPa along the [1] and 525 against 527 GPa along the [01\bar{1}0] at 300 K. Along the [1] it stays within 3% of this average up to 1900 K, along the [011ˉ0][01\bar{1}0] it stays within 3% up to 1500 K and 7% above it at 1900 K (445 against 417 GPa). The stiffness along the cc-axis is also where the potential accuracy is lowest (C33C_{33} is within 3%, Table 2b), so the values along the [0001][0001] direction carry this uncertainty.

Figure 7. Uniaxial tension of single-crystal and Σ193\Sigma193 boundary of HEB2_2 with MACE. For single crystal, (a,b) Averaged stress–strain curves along [1] and [01\bar{1}0] directions at 300–1900 K (see color scale). The circular insets represent slices of the cell at the strains they point to, before and after failure. (c,d) Young’s modulus EE against temperature for HEB2_2 (teal) and the five binary diborides (colors as in the key), from stepped finite deformation. The cell sketches show the direction of tensile deformation. (e,f) Maximum stress σmax⁡\sigma_{\max} of HEB2_2 against temperature. Averaged stress–strain curves of three segregated cells (blue; cation placements taken from 2000 K runs) and three random cells (orange; the same cells with the cations permuted at random) calculated at (g) 300 and (h) 2000 K. Open circles mark σmax⁡\sigma_{\max} for each run. (i, j) Atomic structures of segregated and random cells before loading at 300 K. Below each structure the Nb+Ta fraction of all cations in the cell is plotted in 2.5 Å slices, with a nominal 40 at.% dotted line. Dashed lines mark the two boundary planes.

The tensile deformation along the normal of the segregated and random Σ193\Sigma193 grain boundary of HEB2_2 is shown in the Figure 7(g,h,i,j). As can be seen the Σ193\Sigma193 boundary fails at 20–26 GPa (group means) and 6–7% strain at 300 K, against 33 and 46 GPa at the strain of 10 and 20% for the single crystal along two directions, and in every run the crack opens at a boundary (Supplementary Fig. ??). Before loading (Figure 7(i,Figure 7j)) the two boundary planes of the segregated cell hold 62 and 65 at.% of Nb and Ta against about 37 at.% in the grains, while in the random cell the same planes hold 32 and 46 at.%, within its fluctuations from slice to slice. Segregation makes this boundary stronger: at 300 K the three segregated cells hold 25.9 ± 0.8 GPa and the three random ones 19.7 ± 2.0 GPa, at a strain of 7.0 against 5.7%, and the segregated cells are also 4% stiffer. At 2000 K the difference shrinks to 16.3 ± 0.7 against 14.4 ± 1.3 GPa; every segregated cell is still above every random one, but the difference is comparable to the spread of the random cells. Since within each pair of cells the geometry is the same and only the cations are rearranged, the gain at 300 K comes from the chemistry of the boundary, with the caveat that this common geometry was annealed together with the segregated chemistry. It goes in the direction expected from the boundary energy: at fixed surface energy the work of separation is Wsep=2γs−γGBW_{\mathrm{sep}} = 2\gamma_{\mathrm{s}} - \gamma_{\mathrm{GB}}, and segregation lowers γGB\gamma_{\mathrm{GB}} of Σ193\Sigma193 by 1.1 J/m2^2 against its random reference of Figure 4(b) (γGB\gamma_{\mathrm{GB}} of the tension cells themselves was not computed). The surface term was not computed either, and it works the other way: in DFT the Nb- and Ta-rich surfaces are the cheaper ones (Supplementary Table ??), so a Nb- and Ta-rich fracture surface lowers WsepW_{\mathrm{sep}}. The three segregated cells are snapshots of one MC/MD run, 20 ps apart, so this is one boundary and one run per chemistry, and the result shows that segregation of Nb and Ta can strengthen a boundary. For Nb and Ta, earlier DFT calculations give either sign, depending on the host: on the sites they prefer they raise the ideal tensile strength of ZrB2_2 boundaries by 12.5–34%[15] and lower it at all three TiB2_2 boundaries studied, by up to 24%[14]. W, a different segregant, strengthens 13 of 14 boundaries of ZrB2_2, by up to 45% in MD at 2000 K[17]. Our gain of 31% at 300 K is of the size found for ZrB2_2, for a segregant that makes up most of the boundary plane rather than a single site.

Discussion

According to our calculations γGB\gamma_{\mathrm{GB}} splits by group at the Σ6\Sigma6, Σ8\Sigma8 and basal twist boundaries with grain boundary energies of 2.5–4.8 J/m2^2 for group IV, 0.8–2.7 J/m2^2 for group V, with TaB2_2 having the lowest energy everywhere and NbB2_2 with highest energy. The symmetric Σ7\Sigma7 tilt boundary shows no split, and for TiB2_2 and ZrB2_2 it agrees with earlier DFT within 0.06 J/m2^2. A random decoration of HEB2_2 gives average value of γGB\gamma_{\mathrm{GB}} among the binary diborides, within 0.1 J/m2^2 at three grain boundaries and 0.23 J/m2^2 above the average at Σ8, and the large considered cell with MACE agree with the average value within 0.12 J/m2^2.

By ICOHP analysis, the Me–B and Me–Me bonds at the boundary are, on average, as strong as or slightly stronger than in the bulk, by 0.01–0.27 eV per bond. The difference between the two groups is at the Fermi level. In the bulk of NbB2_2 and TaB2_2, the Me–B bonds are antibonding at EFE_F (and so are Nb–Nb bonds, while Ta–Ta is close to zero, +0.005 eV−1^{-1}). However, directly at the boundary, they become bonding, on average over four boundaries. For group IV, Nb–Nb and Ta–Ta bonds are bonding in both cases. The shift is largest at the twist boundary and smallest at the Σ7 tilt boundary, which has no group split. The results for group V are consistent with these changes at EFE_F, and there is a correlation between the four boundaries.

In HEB2_2, the segregation follows a similar pattern. At the Σ6, Σ7 twist, Σ193 and the general boundaries of a three-grain polycrystal, both Nb and Ta are enriched to 23–35% at 2000 K, while Hf is depleted. This is in contrast to what would be predicted based on size alone. Both the swap spectrum and McLean isotherms give negative segregation energies for Nb and Ta at these sites, ranging from −0.03 to −0.17 eV. At this scale, the potential determines the sign of the segregation rather than its magnitude. Only the Σ7 tilt, with compressed and expanded sites next to each other, sorts the metals according to their radius: Ti occupies the small site, Zr the large site, and their site energies are within 0.3 eV of those predicted by DFT for dilute solutes in the same positions. This sorting is lost at 3000 K, along with the onset of boron diffusion at the boundaries. Σ8 is in between, where Zr is enriched alongside Nb and Ta at 2000 K and Ti is depleted. At 2000 K, the segregated boundaries have a separation of 0.17–1.28 J/m2^2 below the random decoration in γGB\gamma_{\mathrm{GB}}. The site preferences alone account for 0.25–0.74 J/m2^2, reproducing the drop at twist, Σ8 and Σ7 tilt; at Σ6 and Σ193, the rest is attributed to boundary reconstruction during the run. In polycrystals, two-thirds of the drop comes from the annealing of boundary geometry, and one-third from chemistry. At 300 K tension, the Σ193 boundary fails at 20–26 GPa, compared to 33–46 GPa for single crystals, whose Young’s modulus is the average of five binary values. The segregated boundary has a strength of 25.9 ± 0.8 GPa, compared to 19.7 ± 2.0 GPa for a random boundary.

For the ceramic, this has consequences that we did not expect. A boundary that is made cheaper by segregation provides less driving force for grain growth, and even the random HEB2_2 boundaries, on average, are below those of ZrB2_2 at all four boundaries (by 0.02–0.64 J/m2^2), which explains the slower grain growth of HEB2_2 compared to ZrB2_2[13]. Since the hardness of high-entropy diborides depends on the grain size[11, 12], both a finer grain size and a stronger boundary will affect it. To determine whether they do, we need to measure grain growth and hardness on annealed samples.

Oxidation is outside the scope of this work, but it is important to note that it has an impact on the segregation process in two opposing ways. In HEB2_2, the metals oxidize in the following order: Hf, Zr > Ti > Ta, Nb. At temperatures between 1500 and 1800 °C, Nb and Ta remain in borides, while Hf and Zr have already formed oxides, resulting in a boundary that is enriched in Nb and Ta but depleted in Hf. This boundary may take longer to oxidize compared to a random boundary.

Once Nb and Ta have been oxidized, they form eutectic compounds with TiO2_2 and ZrO2_2 that melt at temperatures between 1436 and 1621 °C. In the high-entropy carbide, these liquids leave behind voids that allow oxygen to enter through the boundaries. The outcome of these processes depends on the conditions of the oxidation run and the potential extension to the oxide phase.

The potential compresses the γGB\gamma_{\mathrm{GB}} range by 0.2–0.45 J/m2^2 at its low end, so it works against Nb and Ta, and it is unlikely that it has produced their enrichment. DFT cells contain one decoration per boundary, and SQS cells do not place Nb on a boundary site. Therefore, the bonding of Nb and Ta inside HEB2_2 is inferred from binaries. The electron count argument is specific to diborides: in high-entropy carbides of the same metals, group V binaries also have cheaper boundaries, but Ti and Zr still segregate. One experiment on host metals at a HEB2_2 boundary, a qualitative STEM-EDS map of a spark plasma-sintered sample, showed no enrichment. A composition profile across the boundaries of an annealed sample would confirm or refute this prediction.

Conclusions

In this work, we combined density functional theory, crystal orbital Hamilton population analysis, and machine-learning-accelerated Monte Carlo/molecular dynamics to elucidate the grain-boundary behavior of the high-entropy diboride (Hf,Ta,Ti,Nb,Zr)B2_2 and its binary constituents. Our calculations reveal a clear group-dependent split in grain-boundary energies: group V diborides (NbB2_2, TaB2_2) exhibit roughly half the boundary energy of group IV diborides (TiB2_2, ZrB2_2, HfB2_2) at Σ6\Sigma6, Σ8\Sigma8, and basal twist boundaries, while the symmetric Σ7\Sigma7 tilt remains insensitive to composition. This distinction originates at the Fermi level, where the Me–B bonds in bulk NbB2_2 and TaB2_2 are antibonding but become bonding at the boundary, a change absent in group IV diborides. In the high-entropy diboride, this electronic effect drives the segregation of Nb and Ta to five of the six studied boundaries, reaching 23–35 at.% at 2000 K, contrary to size-based predictions. Only the Σ7\Sigma7 tilt, which presents compressed and expanded sites, sorts metals by atomic radius. Segregation lowers the grain-boundary energy by 0.17–1.28 J/m2^2 and strengthens the Σ193\Sigma193 boundary, which sustains 25.9 ± 0.8 GPa in tension at 300 K compared to 19.7 ± 2.0 GPa for a random boundary.

These findings have implications for the processing and properties of high-entropy diboride ceramics. The reduced boundary energy due to segregation suggests a lower driving force for grain growth, consistent with the slower grain growth observed in HEB2_2 relative to ZrB2_2. Since hardness in these materials follows a Hall–Petch relationship, a finer grain size combined with stronger boundaries could enhance mechanical performance. However, the effect of segregation on oxidation resistance remains ambiguous: while Nb- and Ta-enriched boundaries may delay oxidation, they can also form low-melting eutectics with TiO2\mathrm{TiO_2} and ZrO2\mathrm{ZrO_2}, potentially compromising integrity at elevated temperatures.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work the authors used Claude Opus (Anthropic) to polish the text of the manuscript and help with code for calculations analysis. All results were rechecked by the authors, who take full responsibility for the content of the publication.

Data availability

The MACE potential, its training and validation sets and the atomic structures of this work are available at Zenodo (DOI: 10.5281/zenodo.23453025).

References

  1. [1]F. Thorpe, E. Sobalvarro Converse, J. Rivera, W. Rosenberg, S. J. McCormack, Melting point of group IV and V transition metal diborides, Journal of the American Ceramic Society 109 (1) (2026) e70299. doi:10.1111/jace.70299.DOI
  2. [2]W. G. Fahrenholtz, G. E. Hilmas, I. G. Talmy, J. A. Zaykoski, Refractory diborides of zirconium and hafnium, Journal of the American Ceramic Society 90 (5) (2007) 1347–1364. doi:10.1111/j.1551-2916.2007.01583.x.DOI
  3. [3]M. Magnuson, L. Hultman, H. Högberg, Review of transition-metal diboride thin films, Vacuum 196 (2022) 110567. doi:10.1016/j.vacuum.2021.110567.DOI
  4. [4]L. Cappai, M. Casu, A. M. Locci, G. Cao, S. Garroni, V. Cannillo, D. Bellucci, R. Orrù, From individual to equimolar binary transition metal diborides: The case of (Nb₀.5Mo₀.5)B2, M = Ti, Hf, Journal of the European Ceramic Society 45 (8) (2025) 117277. doi:10.1016/j.jeurceramsoc.2025.117277.DOI
  5. [5]M. D. Hossain, N. S. McIlwaine, N. O. Marquez-Rios, A. C. Feltrin, V. Chawla, R. A. Mayanovic, W. G. Fahrenholtz, D. Penumadu, E. Zurek, D. W. Brenner, D. E. Wolfe, S. Divilov, H. Eckert, S. Curtarolo, J.-P. Maria, Superhard refractory high-entropy diborides, Physical Review Materials 10 (3) (2026) 033604. doi:10.1103/hqlp-dcvl.DOI
  6. [6]P. Vajeeston, P. Ravindran, C. Ravi, R. Asokamani, Electronic structure, bonding, and ground-state properties of AlB2-type transition-metal diborides, Physical Review B 63 (4) (2001) 045115. doi:10.1103/PhysRevB.63.045115.DOI
  7. [7]H. Zhai, W. Gong, X. Shen, X. Ye, X. Song, P. Liu, D. Yan, Y. Shi, R. Yu, X. Yu, Q. Li, Y. Ma, Significant enhancement of hardness and stability in the ScₓTa₁₋ₓB₂ transition-metal diborides system, Fundamental Research 6 (2) (2026) 1078–1085. doi:10.1016/j.fmre.2024.12.025.DOI
  8. [8]S.-H. Jhi, J. Ihm, S. G. Louie, M. L. Cohen, Electronic mechanism of hardness enhancement in transition-metal carbonitrides, Nature 399 (6732) (1999) 132–134. doi:10.1038/20148.DOI
  9. [9]J. Gild, Y. Zhang, T. Harrington, S. Jiang, T. Hu, M. C. Quinn, W. M. Mellor, N. Zhou, K. Vecchio, J. Luo, High-entropy metal diborides: A new class of high-entropy materials and a new type of ultrahigh temperature ceramics, Scientific Reports 6 (2016) 37946. doi:10.1038/srep37946.
  10. [10]L. Feng, W. G. Fahrenholtz, G. E. Hilmas, Processing of dense high-entropy boride ceramics, Journal of the European Ceramic Society 40 (12) (2020) 3815–3823. doi:10.1016/j.jeurceramsoc.2020.03.065.DOI
  11. [11]Y. Yang, J. Bi, L. Qiao, G. Liang, S. Liang, S. Liu, S. Wang, H. Gong, Z. Qian, J. Shi, W. Li, Effect of grain size on mechanical properties and tribological behavior of size-tunable high entropy diboride ceramics obtained by two-step SPS sintering, Journal of Advanced Ceramics 13 (11) (2024) 1835–1843. doi:10.26599/JAC.2024.9220980.DOI
  12. [12]L. Feng, W. G. Fahrenholtz, G. E. Hilmas, F. Monteverde, Effect of Nb content on the phase composition, densification, microstructure, and mechanical properties of high-entropy boride ceramics, Journal of the European Ceramic Society 41 (1) (2021) 92–100. doi:10.1016/j.jeurceramsoc.2020.08.058.DOI
  13. [13]Y.-Y. Cheng, L. Zhou, J.-X. Liu, Y.-F. Tan, G.-J. Zhang, Grain growth inhibition by sluggish diffusion and Zener pinning in high-entropy diboride ceramics, Journal of the American Ceramic Society 106 (8) (2023) 4997–5004. doi:10.1111/jace.19128.DOI
  14. [14]F.-Z. Dai, H. Xiang, Y. Zhou, Strategy to design high performance TiB₂-based materials: Strengthen grain boundaries by solid solute segregation, Journal of the American Ceramic Society 103 (5) (2020) 3311–3320. doi:10.1111/jace.17018.DOI
  15. [15]F.-Z. Dai, Y. Zhou, W. Sun, Segregation of solute atoms (Y, Nb, Ta, Mo and W) in ZrB₂ grain boundaries and their effects on grain boundary strengths: A first-principles investigation, Acta Materialia 127 (2017) 312–318. doi:10.1016/j.actamat.2017.01.048.DOI
  16. [16]S. Guo, H. Sun, Superhardness induced by grain boundary vertical sliding in (001)-textured ZrB₂ and TiB₂ nano films, Acta Materialia 218 (2021) 117212. doi:10.1016/j.actamat.2021.117212.DOI
  17. [17]F.-Z. Dai, B. Wen, H. Xiang, Y. Zhou, Grain boundary strengthening in ZrB₂ by segregation of W: Atomistic simulations with deep learning potential, Journal of the European Ceramic Society 40 (15) (2020) 5029–5036. doi:10.1016/j.jeurceramsoc.2020.06.007.DOI
  18. [18]F.-Z. Dai, B. Wen, Y. Hu, X.-F. Gu, A deep neural network potential model for transition metal diborides, Journal of Materials Informatics 4 (3) (2024) 10. doi:10.20517/jmi.2024.14.DOI
  19. [19]C. Wang, M. Qin, T. Lei, Y. He, K. Kisslinger, T. J. Rupert, J. Luo, H. L. Xin, Synergic grain boundary segregation and precipitation in W- and W-Mo-containing high-entropy borides, Journal of the European Ceramic Society 41 (10) (2021) 5380–5387. doi:10.1016/j.jeurceramsoc.2021.04.004.DOI
  20. [20]M. Qin, S. Shivakumar, T. Lei, J. Gild, E. C. Hessong, H. Wang, K. S. Vecchio, T. J. Rupert, J. Luo, Processing-dependent stabilization of a dissimilar rare-earth boride in high-entropy (Ti₀.₂Zr₀.₂Hf₀.₂Ta₀.₂Er₀.₂)B₂ with enhanced hardness and grain boundary segregation, Journal of the European Ceramic Society 42 (12) (2022) 5164–5171. doi:10.1016/j.jeurceramsoc.2022.05.034.DOI
  21. [21]X.-G. Li, C. Chen, H. Zheng, Y. Zuo, S. P. Ong, Complex strengthening mechanisms in the NbMoTaW multi-principal element alloy, npj Computational Materials 6 (1) (2020) 70. doi:10.1038/s41524-020-0339-0.DOI
  22. [22]K. Ito, H. Sawada, S. Tanaka, S. Ogata, M. Kohyama, Electronic origin of grain boundary segregation of Al, Si, P, and S in bcc-Fe: combined analysis of ab initio local energy and crystal orbital Hamilton population, Modelling and Simulation in Materials Science and Engineering 29 (1) (2021) 015001. doi:10.1088/1361-651X/abc04c.DOI
  23. [23]I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner, G. Csányi, MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields, in: Advances in Neural Information Processing Systems 35 (NeurIPS 2022), 2022, pp. 11423–11436. doi:10.52202/068431-0830.DOI
  24. [24]I. Batatia, C. Lin, J. Hart, E. Kasaor, A. M. Elena, S. W. Norwood, T. Wolf, G. Csányi, Cross Learning between Electronic Structure Theories for Unifying Molecular, Surface, and Inorganic Crystal Foundation Force Fields, arXiv preprint (2025). arXiv:2510.25380, doi:10.48550/arXiv.2510.25380.
  25. [25]J. Wang, H. Zhang, S. Pan, Q. Jia, C. Ding, M. M. Lukanov, F. Jalalov, A. G. Kvashnin, Z. Fan, J. Sun, NEPmaker: Active learning of neuroevolution machine learning potential for large cells, Chinese Physics B doi:10.1088/1674-1056/ae85f8.DOI
  26. [26]A. Zunger, S.-H. Wei, L. G. Ferreira, J. E. Bernard, Special quasirandom structures, Physical Review Letters 65 (3) (1990) 353–356. doi:10.1103/PhysRevLett.65.353.DOI
  27. [27]G. Kresse, J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set, Physical Review B 54 (16) (1996) 11169–11186. doi:10.1103/PhysRevB.54.11169.DOI
  28. [28]J. P. Perdew, K. Burke, M. Ernzerhof, Generalized Gradient Approximation Made Simple, Physical Review Letters 77 (18) (1996) 3865–3868. doi:10.1103/PhysRevLett.77.3865.DOI
  29. [29]P. E. Blöchl, Projector augmented-wave method, Physical Review B 50 (24) (1994) 17953–17979. doi:10.1103/PhysRevB.50.17953.DOI
  30. [30]G. Kresse, D. Joubert, From ultrasoft pseudopotentials to the projector augmented-wave method, Physical Review B 59 (3) (1999) 1758–1775. doi:10.1103/PhysRevB.59.1758.DOI
  31. [31]B. Hammer, L. B. Hansen, J. K. Nørskov, Improved adsorption energetics within density-functional theory using revised Perdew-Burke-Ernzerhof functionals, Physical Review B 59 (11) (1999) 7413–7421. doi:10.1103/PhysRevB.59.7413.DOI
  32. [32]J. P. Perdew, A. Ruzsinszky, G. I. Csonka, O. A. Vydrov, G. E. Scuseria, L. A. Constantin, X. Zhou, K. Burke, Restoring the Density-Gradient Expansion for Exchange in Solids and Surfaces, Physical Review Letters 100 (13) (2008) 136406. doi:10.1103/PhysRevLett.100.136406.DOI
  33. [33]J. W. Furness, A. D. Kaplan, J. Ning, J. P. Perdew, J. Sun, Accurate and Numerically Efficient r2SCAN Meta-Generalized Gradient Approximation, The Journal of Physical Chemistry Letters 11 (19) (2020) 8208–8215. doi:10.1021/acs.jpclett.0c02405.DOI
  34. [34]H. Du, D. Lv, B. Wei, Y. Ren, F. Yu, X. He, B. Zhu, Y. Huang, Y. Li, J. Liu, S. Shi, H. Wang, Z. Lu, Origins of universal machine learning force-field errors in multicomponent materials (2026). arXiv:2610.09837. URL https://arxiv.org/abs/2610.09837
  35. [35]P. Hirel, Atomsk: A tool for manipulating and converting atomic data files, Computer Physics Communications 197 (2015) 212–219. doi:10.1016/j.cpc.2015.07.012.DOI
  36. [36]K. Song, R. Zhao, J. Liu, Y. Wang, E. Lindgren, Y. Wang, S. Chen, K. Xu, T. Liang, P. Ying, N. Xu, Z. Zhao, J. Shi, J. Wang, S. Lyu, Z. Zeng, S. Liang, H. Dong, L. Sun, Y. Chen, Z. Zhang, W. Guo, P. Qian, J. Sun, P. Erhart, T. Ala-Nissila, Y. Su, Z. Fan, General-purpose machine-learned potential for 16 elemental metals and their alloys, Nature Communications 15 (1) (2024) 10208. doi:10.1038/s41467-024-54554-x.DOI
  37. [37]Z. Fan, Y. Wang, P. Ying, K. Song, J. Wang, Y. Wang, Z. Zeng, K. Xu, E. Lindgren, J. M. Rahm, A. J. Gabourie, J. Liu, H. Dong, J. Wu, Y. Chen, Z. Zhong, J. Sun, P. Erhart, Y. Su, T. Ala-Nissila, GPUMD: A package for constructing accurate machine-learned potentials and performing highly efficient atomistic simulations, The Journal of Chemical Physics 157 (11) (2022) 114801. doi:10.1063/5.0106617.DOI
  38. [38]A. V. Shapeev, Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Potentials, Multiscale Modeling & Simulation 14 (3) (2016) 1153–1173. doi:10.1137/15M1054183.DOI
  39. [39]I. S. Novikov, K. Gubaev, E. V. Podryabinkin, A. V. Shapeev, The MLIP package: moment tensor potentials with MPI and active learning, Machine Learning: Science and Technology 2 (2) (2021) 025002. doi:10.1088/2632-2153/abc9fe.DOI
  40. [40]E. Podryabinkin, K. Garifullin, A. Shapeev, I. Novikov, MLIP-3: Active learning on atomic environments with moment tensor potentials, The Journal of Chemical Physics 159 (8) (2023) 084112. doi:10.1063/5.0155887.DOI
  41. [41]A. P. Thompson, H. M. Aktulga, R. Berger, D. S. Bolintineanu, W. M. Brown, P. S. Crozier, P. J. in ’t Veld, A. Kohlmeyer, S. G. Moore, T. D. Nguyen, R. Shan, M. J. Stevens, J. Tranchida, C. Trott, S. J. Plimpton, LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales, Computer Physics Communications 271 (2022) 108171. doi:10.1016/j.cpc.2021.108171.DOI
  42. [42]J. M. Cowley, An Approximate Theory of Order in Alloys, Physical Review 77 (5) (1950) 669–675. doi:10.1103/PhysRev.77.669.DOI
  43. [43]D. McLean, Grain Boundaries in Metals, Clarendon Press, Oxford, 1957.DOI
  44. [44]A. Stukowski, Visualization and analysis of atomistic simulation data with OVITO–the Open Visualization Tool, Modelling and Simulation in Materials Science and Engineering 18 (1) (2010) 015012. doi:10.1088/0965-0393/18/1/015012.DOI
  45. [45]R. Dronskowski, P. E. Blöchl, Crystal orbital Hamilton populations (COHP): energy-resolved visualization of chemical bonding in solids based on density-functional calculations, The Journal of Physical Chemistry 97 (33) (1993) 8617–8624. doi:10.1021/j100135a014.DOI
  46. [46]V. L. Deringer, A. L. Tchougréeff, R. Dronskowski, Crystal Orbital Hamilton Population (COHP) Analysis As Projected from Plane-Wave Basis Sets, The Journal of Physical Chemistry A 115 (21) (2011) 5461–5466. doi:10.1021/jp202489s.DOI
  47. [47]S. Maintz, V. L. Deringer, A. L. Tchougréeff, R. Dronskowski, LOBSTER: A tool to extract chemical bonding from plane-wave based DFT, Journal of Computational Chemistry 37 (11) (2016) 1030–1035. doi:10.1002/jcc.24300.DOI
  48. [48]R. Nelson, C. Eltural, J. George, V. L. Deringer, G. Hautier, R. Dronskowski, LOBSTER: Local orbital projections, atomic charges, and chemical-bonding analysis from projector-augmented-wave-based density-functional theory, Journal of Computational Chemistry 41 (21) (2020) 1931–1940. doi:10.1002/jcc.26353.DOI
  49. [49]D. S. Alekseev, I. V. Chepkasov, A. G. Kvashnin, gpuchp: GPU projection of plane-wave wavefunctions onto local orbitals for COHP analysis (2026). doi:10.5281/zenodo.22995816. URL https://doi.org/10.5281/zenodo.22995816
  50. [50]M. M. Mou, T. M. A. Haque, S. E. Daigle, J. Roberts, W. G. Fahrenholtz, J.-P. Maria, D. E. Wolfe, E. Zurek, S. Curtarolo, D. W. Brenner, Finite-temperature grain-boundary segregation in high-entropy carbides, Journal of the American Ceramic Society 109 (7) (2026) e71032. doi:10.1111/jace.71032.DOI

Paper details

vlnc:W_n7xdyrvm
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

Claude Opus

View statement
Looking for related papers…
Contents