betaFeedback

WorksPhysics and Astronomy

Performance of universal machine learning potentials in global optimization of inorganic crystal structures

Abstract

Rapid development of universal machine learning potentials (uMLPs) and expansion of training data sets are reshaping the state of the art in atomistic simulation, highlighting the need for concurrent systematic benchmarking of their capabilities. Global optimization is among the most demanding uMLP applications because unconstrained exploration includes probing motifs not present in reference sets. We examined twelve pretrained uMLPs in unconstrained evolutionary searches to assess whether these models can consistently predict complex nonmagnetic crystal structure ground states under ambient pressure across diverse inorganic systems. Our findings demonstrate that the considered M3GNet, MACE-MATPES, MACE-mh-1, SevenNet, EquiformerV2, EquiformerV3, MatterSim, GRACE, eSEN, Orb-v3, PET-MAD, and PET-OAM models span a wide performance range, from near ab initio to essentially non-predictive, in their ability to resolve competing phases within low-energy basins. Additional tests on hcp-Zn, MB4_4 (M = Cr, Mn, and Fe), and LiBy_{y} (y≈0.9y\approx 0.9) ground states reveal that several uMLPs capture fine energy differences arising from subtle electronic structure features.

Introduction

Machine learning interatomic potentials (MLPs) have earned a place in mainstream computational materials research as efficient alternatives to electronic structure calculations. The first models capable of describing arbitrary atomic environments, the Behler–Parrinello neural networks (NNs) and the Gaussian approximation potentials (GAPs), demonstrated that density functional theory (DFT)-level accuracy can be retained in simulations where system sizes and times scales are extended by orders of magnitude [1, 2]. In the following years, MLPs were applied to a broad range of tasks, including molecular dynamics (MD) simulations of phase stability, thermal transport, and mechanical response, geometry optimizations of clusters, surfaces, and extended defects, and transition state searches of diffusion and reaction pathways [3, 4, 5].

The impact of MLPs on crystal structure prediction was less immediate. Many early studies focused on elemental materials, e.g., B, C, Si, and P, and either reproduced known structures or identified metastable polymorphs [6, 7, 8, 9]. The first examples of new thermodynamically stable compounds found with MLP-accelerated global searches appeared only in the late 2010s [10, 11, 12, 13, 14]. The main challenge was the construction of practical potentials that could resolve small energy differences between competing structures while remaining reliable in previously unexplored regions of configuration space. This was addressed, in particular, by extending models from elements to multicomponent systems and by introducing automated workflows for iterative reference data generation and model parametrization [15, 16, 17, 18]. The resulting MLPs offered genuine acceleration, leading to the identification of new complex M-Sn [11, 12] and SiH4_4 [19] ground states overlooked in prior searches at the DFT level. Despite these successes, the need to build robust system-specific models has limited the widespread use of MLPs in routine crystal structure prediction.

Recently introduced universal MLPs (uMLPs) represent a conceptually different approach. Trained on large and broadly representative materials databases, these models enable users to either bypass system-specific parametrization altogether or carry out further fine-tuning for target applications. uMLPs have already been used for large-scale screening of stability, transport properties, and finite-temperature behavior and have been benchmarked in several studies [20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30]. In particular, Yu et al. [23] assessed four uMLPs on equation-of-state curves, relaxed geometries, and formation energies for extensive sets of crystalline materials. Riebesell et al. [25] developed the Matbench Discovery framework to evaluate property predictions, such as relaxed formation energies and distances to the convex hull, from unrelaxed structures. Chiang et al. [26] presented the MLIP Arena system, which compares the performance of nine uMLPs across equation-of-state curves and MD simulations, focusing on physically grounded performance independent of DFT references. Tahmasbi et al. [27] extended these evaluations to global structure searches by using a minima hopping algorithm to probe the potential energy surfaces (PES) of unary materials. These studies establish that current uMLPs are sufficiently accurate for local optimization of pre-defined structure prototypes and near-equilibrium properties as well as applicable to global optimization of elemental systems, but their performance in unconstrained structure searches for stable chemically diverse inorganic compounds has yet to be explored systematically.

In this work, we assess the performance of the latest generation of uMLPs as well as three legacy models. els in global structure searches in multicomponent systems. The workflow complements existing benchmarks by probing the surrogate PES beyond the configurations represented in the training datasets, as we investigated materials with both well-established [31, 32, 33, 34, 35, 36, 37, 39, 40] [38] and recently proposed [11, 12, 41] ambient-pressure ground states. The twelve selected models, M3GNet [20], MACE-MATPES [42, 43], MACE-mh-1 [44], SevenNet [45, 46], EquiformerV2 [47, 48, 49, 50], EquiformerV3 [51], MatterSim [52], GRACE [53, 54], eSEN [48, 55], Orb-v3 [56, 57], PET-MAD [58], and PET-OAM [59] span various architectures and vastly different training sets, but nearly all of them were fitted to the same DFT approximation (see Table I). To determine their suitability for ground state searches across materials with different chemistries and bonding types, we used the uMLPs as out-of-the-box surrogate models combined with an evolutionary algorithm [16]. Within this framework, which previously enabled the prediction of new thermodynamically stable compounds [11, 12], the small pools of identified low-energy candidate structures are re-examined with the reference DFT approximation. We also consider three challenging cases, hcp-Zn with an anomalous c/ac/a ratio, distorted MB4\mathrm{MB}_{4} derivatives (M=Cr\mathrm{M}=\mathrm{Cr}, Mn, and Fe), and off-stoichiometric LiBy\mathrm{LiB}_{y} phases, in which electronic features are known to play a key role in defining the ground state structure. By simulating all selected systems with three DFT flavors, the Perdew–Burke–Ernzerhof (PBE) [60], PBE for solids (PBEsol) [61], and regularized-restored strongly constrained and appropriately normed (r2SCAN) [62] exchange-correlation functionals, we contextualize the uMLP errors and demonstrate that some models, most notably EquiformerV3, PET-OAM, and eSEN, reproduce the reference method with accuracy better than the systematic spread among these DFT approximations.

uMLP versionID adoptedYear reporteduMLP architectureFeaturizing basis functionsDataset sourcesDFT functionalStructures (×106\times10^{6})Parameters (×106\times10^{6})
M3GNetMG2022GNNBessel RBF YℓmY_{\ell}^{m}MPPBE0.21.1
MACE-MATPESMC∗^{*}2023Eqv. MPNNBessel RBF YℓmY_{\ell}^{m}MATPES-PBEPBE19.8
MACE-mh-1MC2025E(3) GNNBessel RBF YℓmY_{\ell}^{m}OMat24PBE/PBE+U11019.8
SevenNet-MFSN2024E(3) NequIPBessel RBF YℓmY_{\ell}^{m}OMat24 MPtrj sAlexPBE11325.7
EquiformerV2EQ∗^{*}2024E(3)/SE(3)Gaussian RBF YℓmY_{\ell}^{m}OMat24 MPtrj sAlexPBE10298.4
EquiformerV3EQ2026SE(3)Gaussian RBF YℓmY_{\ell}^{m}OMat24 MPtrj sAlexPBE11330
MatterSimMS2024M3GNet GraphSBF YℓmY_{\ell}^{m} GKMatterSimPBE174.6
GRACE-2LGR2025ACE GraphChebyshev RBF YℓmY_{\ell}^{m}OMat24PBE25.2
eSENEN2025E(3)/SE(3)Gaussian RBF YℓmY_{\ell}^{m}OMat24PBE1030.2
Orb-v3OR2025GNSBessel RBF YℓmY_{\ell}^{m}OMat24PBE5525.5
PET-MADPT∗^{*}2025GNNMachine learnedMADPBEsol0.13.5
PET-OAMPT2026GNNMachine learnedOMat24 MPtrj sAlexMAD PBE113730

Table 1. Overview of uMLPs evaluated in this work, with additional information given in Section II B. For each model we report the publication year, model architecture, featurizing basis functions, training data source, reference DFT method, training set size, and the parameter count. RBF, SBF, and GK stand for radial basis functions, spherical Bessel functions, and Gaussian kernels, respectively. Blank entries represent data not reported or readily available.

Methodology

DFT Calculations

All DFT calculations were performed with the Vienna ab initio simulation package (VASP) [63, 64, 65, 66] and projector-augmented wave potentials [67] with the maximum number of semi-core electrons for each element. We used a 500 eV plane-wave cutoff and dense Monkhorst–Pack [68] kk-meshes with Δk≤2π×0.025\Delta k \le2\pi\times0.025 Å−1^{-1} to ensure good numerical convergence. All final energies were evaluated with the tetrahedron integration method with Blöchl corrections [69] for fully optimized unit cells. As the default DFT method for nearly all uMLPs (see Table I), we employed the PBE exchange-correlation functional [60] within the generalized gradient approximation (GGA) [70]. For examining systematic DFT errors, we also carried out calculations with the PBEsol [61] functional, the r2^{2}SCAN functional [62] that is currently adopted in the Materials Project [71] for its improved description of materials formation energies, and four different parametrizations [72, 73, 74, 75] that treat van der Waals (vdW) interactions important in layered materials.

uMLPs

Pretrained models were taken directly from publicly available repositories and were not fine-tuned in this study in order to evaluate their baseline performance. M3GNet (denoted as MG henceforth) is a pioneering model trained on three-body interactions across 89 elements [20]. MACE-MATPES (MC∗^{*}) utilizes the atomic cluster expansion as a descriptor and performs efficient calculations due to its streamlined message-passing architecture [42, 43]. MACE-mh-1 (MC) is based on MACE-OMAT-0, which shares an architecture with MACE-MATPES but differs in training data, and adds higher order messages, nonlinear blocks, and a multi-head replay post-training strategy [44]. SevenNet (SN) maintains equivariance, incorporates the NequIP framework [76], and uses a combination of low- and high-fidelity data in its training [45, 46]. EquiformerV2 (EQ{sup}) is an equivariant model that incorporates eSCN convolutions [77] for higher degree representations of atomic environments [47, 48, 49, 50]. EquiformerV3 (EQ) further develops this architecture by introducing SwiGLU-S{sup}2 activations for strict equivariance on coarser grids and a gradual radius cutoff that helps produce smoothly varying PES [51]. MatterSim (MS, v1.0.0–5M) uses active learning, uncertainty-aware data sampling, and a training dataset containing off-equilibrium structures under a variety of temperatures and pressures [52]. GRACE (GR) uses equivariant message passing and extends the atomic cluster expansion to recursively evaluated star- and tree-like basis functions, constructing a complete basis for describing atomic interactions [53, 54]. eSEN (EN) is a message-passing NN with an Equiformer-like architecture [47] designed to strictly maintain equivariance and a smoothly varying PES [48, 55]. The conservative, unlimited neighbor Orb-v3 (OR) variant is a roto-equivariant graph network simulator (GNS) that builds off of the previous generation of direct-force Orb models [56, 57]. PET-MAD (PT{sup}, v1.0.2) is a rotationally unconstrained, transformer-based model trained on the Massive Atomic Diversity (MAD) dataset [58]. The recently introduced PET-OAM (PT) updates the PET framework by passing node features up GNN layers, introducing a new direct force head, and increasing the number of node features, allowing for dramatically more parameters (see Table I) at minimal cost [59]. Among these uMLPs, EQ{sup}* is the only explicitly non-conservative model that calculates forces directly rather than as an energy gradient.

In the present tests, local structure relaxations were performed with MatGL [78] in MG simulations or with the Atomic Simulation Environment (ASE) for all other uMLPs [79], and the accuracy of each uMLP’s predictions was evaluated against the DFT functional used in its training, i.e., PBEsol for PET-MAD and PBE for the rest (see Table I). Importantly, none of the tested properties involved absolute DFT energies, which made results independent of the atomic reference values in pseudopotentials, particularly relevant in the case of the PT{sup}* baseline method [58]. For clarity, the main-text comparisons focus primarily on the nine latest uMLPs, whereas the detailed performance of the legacy variants, PT{sup}, MC{sup}, and EQ{sup}*, is reported in the Supplemental Material.

Information on uMLP relative efficiencies in GPU-based high-throughput and MD workflows can be found in previous studies, e.g., Refs. [28, 58]. All present uMLP global searches, requiring sequential generation and local optimization of structures, were carried out on CPU nodes. Our single-core CPU benchmark for fixed-length relaxations of crystal structures with 10–16 atoms per unit cell, shown in Fig. S1, is qualitatively consistent with the reported GPU trends: runtimes depend strongly on model architecture and chemical system, spanning a factor of over 30 between MG and EQ{sup}* in our tests.

Global search benchmarking protocol

Efficient acceleration of ab initio global structure optimization depends on surrogate models meeting two key requirements [11, 16, 27, 80]. First, a uMLP must be robust enough that the PES is not dominated by spuriously over-stabilized unphysical structures and that the true DFT ground state can be reached from a uMLP minimum structure via a barrierless trajectory. Second, it must be sufficiently accurate that the uMLP-based energy ranking retains, within a tractably small pool, at least one candidate that converges to the true ground state upon DFT re-optimization. Systematic comparison of the uMLPs’ performance is not straightforward due to the inherently stochastic nature of global optimization methods. The suitability of uMLPs for finding ground states with no structural input was benchmarked with the following protocol.

The starting point involved carrying out typical zero-temperature fixed-composition evolutionary searches appropriate for locating small- to medium-sized ground states of crystalline solids [11, 12]. Populations of 100 structures with a given number of atoms were seeded randomly and evolved over 100 generations with standard operations implemented in the module for ab initio structure evolution (MAISE) [16]: mutation of individual members (20%), crossover of two parent structures (60%), and injection of additional random seeds (20%). Such searches are usually sufficient but not foolproof for locating ground states in this size range. When the reference global minimum was missed, we carried out up to four additional attempts and report in Table S1 the generation at which that structure entered the population. The 10410^{4} uMLP-level local optimizations in each run proceeded until the maximum atomic force fell below 0.05 eV/Å, which enabled reliable sampling of different basins by detecting and eliminating duplicate motifs based on the similarity in energy and structural fingerprint [16, 81]. To remove any underconverged duplicates, all visited minima up to 20 meV/atom above the best candidate and with the similarity fingerprint scalar product below 0.92, controlled by the SCUT flag in MAISE, were optimized with the uMLP with a finer 0.001 eV/Å tolerance and filtered again with the 0.92 threshold. For compounds where few minima were found after filtering, we increased the energy window by 10 meV/atom and repeated the filtering step until more than 10 minima were found or the energy window reached 100 meV/atom (e.g., for LiB3). Inspection of the pools provided information on the quality of the PES fit for individual uMLPs, judged by the presence of minima significantly below the global one and an overabundance of near-degenerate lattice decorations or low-symmetry satellite minima around structures improperly modeled as dynamically unstable. Accordingly, we used a more restrictive 0.80 (Li3_3Sn) or 0.85 (Na2_2CN2_2) SCUT flag for our selection criterion.

The variance in the number of candidate structures per uMLP pool complicated consistent quantitative assessment of the models’ ability to select competitive candidates in relevant regions of the configuration space. To account for the non-deterministic sampling of low-energy basins and ensure that each uMLP test pool included the ground state for each compound, we merged selected minima from one evolutionary run for each uMLP (as detailed in Table S2) so that cases requiring additional tries did not contribute disproportionately to the aggregate. When multiple attempts were available, we retained the run reaching the lowest-energy uMLP minimum. Subsequent uMLP-level relaxation (0.001 eV/Å tolerance) and filtering (0.92 SCUT) of the merged sets generated comparable pools for each classical model. In addition to the SCUT filter criteria, we also only collected minima up to 100 meV/atom above the lowest energy candidate to avoid the appearance of high energy artifacts that could skew our evaluation. This energy window filter led to an average of 17% of structures in uMLP relaxed pools being filtered out, leaving only relevant low energy candidates for further analysis. Finally, candidates in each uMLP pool were re-optimized with the reference DFT method, generally after conservative 0.01-tolerance MAISE symmetrization to remove numerical noise.

The fingerprint scalar product and the DFT energy difference between the uMLP- and DFT-relaxed configurations, defining structure and energy proximity metrics, respectively, indicated how close the competing minima were on the surrogate and reference PESs. The ranking accuracy was gauged using an RMSE defined as

Ranking RMSE⁡=[∑n=1N(EnuMLP−EnDFT−ΔE‾)2N−1]12.\operatorname{Ranking\ RMSE} = \left[\frac{\sum_{n=1}^{N}\left(E_{n}^{\mathrm{uMLP}}-E_{n}^{\mathrm{DFT}}-\overline{\Delta E}\right)^{2}}{N-1}\right]^{\frac{1}{2}}.

Here, NN is the number of structures in each pool, ΔE‾=E‾uMLP−E‾DFT\overline{\Delta E}=\overline{E}^{\mathrm{uMLP}}-\overline{E}^{\mathrm{DFT}} is the difference between pool averages, and the N−1N-1 factor accounts for the Bessel correction. This metric differs from the typically reported overall model error in that it eliminates the average shifts irrelevant to relative stability, evaluates energies of configurations fully relaxed within their respective methods rather than identical snapshots, and probes accuracy only in the low-energy basins important in global structure searches.

Metrics for ground state screening efficiency are important but not straightforward to define, in part because establishing whether multiple uMLP minima converge to the same DFT ground state is sensitive to the tolerance limits of the reference calculations. We therefore constructed the ground state hit set for each compound as all shortlisted minima of a given uMLP that, after DFT re-optimization, were assigned to this endpoint using the structural fingerprint cutoff. The uMLP candidates were ordered by their uMLP energies, and ranki_i was defined as the earliest position in this ordered list among members of the hit set for compound ii. The mean reciprocal rank was then calculated as

MRR=1Nc∑i=1Nc1ranki.\mathrm{MRR}=\frac{1}{N_c}\sum_{i=1}^{N_c}\frac{1}{\mathrm{rank}_i}.

where NcN_c is the number of considered compounds. The uncertainty in the average MRR was estimated from 95% Student tt confidence intervals over the considered compounds.

The size of the hit set defined the ground state multiplicity (GSM). Large GSM values indicate that the shortlist contains many ground state satellites, and numerous DFT re-optimizations are spent rediscovering the known minimum instead of screening distinct metastable candidates. Reliable ordering of metastable candidates would be valuable for screening workflows [82, 83], especially when low lying phases beyond the 0 K ground state are relevant [11]. We considered KK ranking metrics, including the normalized discounted reciprocal rank (nDRR@KK), but did not include them for two reasons. First, most successful uMLPs already recovered the ground state at rank one, so finite KK scores would add little beyond MRR for those models. Second, automated hit list construction beyond the known ground state would suffer from accumulated ambiguities, as fingerprint-based assignments of shallow DFT minima are susceptible to numerical noise and cannot be readily checked in regions of genuinely fragmented reference PES landscapes.

Operationally, ranking RMSE, MRR, and GSM provided complementary information on the relative stability within the candidate shortlists, rank efficiency in finding the reference ground state for each compound, and DFT screening overhead from redundant global minimum hits.

Evolutionary optimization

Evolutionary searches were performed for 12 inorganic nonmagnetic compounds with bonding environments ranging from extended metal-stabilized covalent-ionic frameworks to interpenetrating metal-metalloid sublattices and structural complexities ranging from 5 to 16 atoms per primitive unit cell (see Table S1). The inclusion of Li3_3Sn, Pd5_5Sn3_3, and MgB3_3C3_3 with recently proposed prototypes [11, 12, 41] served to probe whether the uMLPs can recognize unfamiliar, non-sampled motifs as ground states.

Table II summarizes the qualitative outcome of the searches (see Supplementary Listing 1 for MAISE settings). For each model and compound, we report the uMLP energy of the corresponding lowest-energy candidate relative to the ground state of the reference DFT method and the position of the DFT ground-state structure within the uMLP-ranked pool as a superscript. Negative entries identify competing minima favored by the uMLP over the reference phase, whereas small positive superscripts mean that this phase would be recovered from a shallow post-search DFT shortlist. In the particularly challenging MgB3_3C3_3 case, five runs were insufficient for some uMLPs despite the TETRIS-based generation [84] of initial structures containing B-C building blocks (see Supplementary Listing 2) used to accelerate the convergence. We would like to emphasize that this dataset is not statistically sufficient to disentangle whether missed ground states reflect surrogate deficiencies or seed/trajectory dependence in the evolutionary exploration. However, the investigation did identify two clear model- and compound-level outliers, MG and AgClO4_4, as discussed below.

uMLP IDLiB3_3 tP16Be4_4B tP10TiO2_2 mS24MgIr2_2 hP12Pd5_5Sn3_3 mS32Li3_3Sn hR48Na2_2CN2_2 tI10 (mS10)Na2_2IrO3_3 mS24Si3_3CaPt tI10MgB3_3C3_3 oI28AgClO4_4 tI12
EN0.00.00.00.00.00.00.00.00.00.0−187.40
EQ∗^*0.00.00.00.00.00.0−1.000.00.015.7−185.30
EQ0.00.00.00.00.00.00.00.00.028.4−180.00
OR0.00.00.00.00.00.00.00.00.00.0−189.90
SN0.00.00.00.00.00.00.00.00.00.0−205.70
GR0.00.0−3.420.00.0−2.64−0.430.00.00.0−210.90
MC∗^*0.00.0−37.0270.00.0−7.220.00.00.0−45.10−1e40
MC0.00.00.00.00.00.0−4.320.00.00.0−205.60
MS0.00.00.00.00.00.00.00.00.00.0−179.00
PT∗^*0.00.0−5.220.00.00.00.00.00.00.0−157.00
PT0.00.00.00.00.00.00.00.00.025.3−160.30
MG−47.60−14.50−9.40−29.70−3.820−14.90−56.400.0−14.80−180.30−220.80

Table 2. uMLP energy difference of best candidates from evolutionary searches relative to reference ground states given in meV/atom. The formula and Pearson symbol for each reference are provided in the first two rows. In the case of Na2_2CN2_2, the mS10 experimental structure favored by PBEsol is used as the reference for PT∗^* while the tI10 PBE artifact is used as the reference for all other models. The relative rank of reference structures found in each pool is given as a positive superscript (degenerate structures with energy differences < 0.1 meV/atom are assigned the same rank). Searches where the reference global minimum was not found have a positive energy difference or a zero superscript.

Evolutionary searches with nearly all uMLPs reached the global minima for the majority of considered compounds, mostly independent of the primitive unit cell size or perceived complexity. For example, the relatively large LiB3_3 ground state (tP16) is comprised of compact B6_6 units that recur frequently across metal boride chemistries [85], which is apparently well reproduced as a broad basin of attraction on surrogate PESs. Likewise, the low-symmetry Na2_2IrO3_3 (mS24) features a typical layered arrangement of edge-sharing IrO6_6 octahedra, which tends to constrain the accessible low-energy motifs. The recently proposed Li3_3Sn (hR48) and Pd5_5Sn3_3 (mS32) ground states with new structure types turned out not to pose particular difficulty, as their local environments can be viewed as ordered decorations of familiar metallic and metal-metalloid coordination motifs [11, 12], and the uMLPs learned the underlying stability trend well enough to identify these phases as global minima. Interestingly, about half of the uMLPs favored a distorted mS32-Pd5_5Sn3_3 variant, indicating that reducing the reported C2/c symmetry to Cc lowers the energy by 0.1–0.3 meV/atom. PBE and PBEsol yield similar energy gains along a Γ-point eigenvector in the conventional unit cell, associated with a single imaginary mode that breaks inversion symmetry through buckling of the Pd-Sn framework. However, the extremely small energy change and minor atomic displacements of about 0.1 Å do not necessarily imply that the distortion would be detectable, as it could be suppressed by anharmonic effects. Hence, despite the similarity factor of 0.74 being below our typical threshold, we treated the two phases as identical in our MRR and GSM metric calculations.

The few misses, either failing to access the global-minimum basin or misranking closely competing minima, were concentrated in systems that are challenging even at the DFT level. TiO2_2 exhibits a rich polymorphism, with the relative stability of low-energy phases being highly sensitive to the treatment of dispersive and correlation effects [86]. This complexity of the PES helps frame why two uMLPs reversed the stability order of anatase (tI12) and the PBE-level ground state bronze (mS24) and why another did not allow the searches to log the ground state as a viable candidate. The uMLP-based exploration successfully recovered other known phases, namely, rutile (tP6), columbite (oP12), and baddeleyite (mP12), which indicates that the models correctly distinguished the PBE ground state from a realistic line-up of candidates.

Global optimization of AgClO4_4 with all 12 uMLPs produced numerous spuriously deep minima containing O2_2 molecules, such as mS24 given in the Supplementary Material (Fig. S2), starkly contrasting the accepted tI12 ground state built from ClO4_4 tetrahedra. This pathology indicates that the considered uMLPs are not sufficiently familiar with molecular oxygen to disfavor its appearance in solid-state phases, which should be straightforward to fix by retraining the models on datasets that include this motif. Since the failure occurs across diverse architectures, AgClO4_4 proves to be an out-of-distribution stress test of the released models rather than a conventional low-energy ranking error. In an uncertainty-aware or active-learning workflow, such chemically implausible configurations would be natural triggers for early DFT evaluation and targeted retraining before they dominate evolutionary sampling. Similar measures may be needed when exploring related systems to prevent the unphysical phase separation into notably stable low-dimensional units, such as nitrogen or hydrogen molecules.

In addition to the primary intended benchmarking objective, our global searches with several uMLPs unexpectedly identified lower-energy candidates at the reference level: two apparent semi-local functional artifacts and a robustly better polymorph. For Na2_2CN2_2, a substantially less dense packing of CN2_2 units in a tI10 candidate (17.8 Å3^3/atom) than in the known mS10 structure (13.4 Å3^3/atom) improved the PBE-level stability by 1.9 meV/atom (see Figure 1). However, all other tested functionals soundly disfavor tI10 relative to mS10 by 16–56 meV/atom, which makes the outlier likely PBE-specific. For CYI, the experimentally reported mS12 phase [37] [38] is reproduced by PBEsol and the vdW-corrected functionals, while PBE and r2^2SCAN favor a related mS24 variant with tilted C2_2 dimers (see Figure 1). This near degeneracy is expected for a layered compound in which shifted Y2_2I2_2 sheets and C2_2 dimers are sensitive to dispersive interactions. In fact, this is reflected in the different stacking orders found among the collected uMLP minima and a consistent 1–2 Å overestimation of the interlayer spacing, an artifact inherited from the underlying semi-local DFT approximations that misestimate the measured values by 0.2–0.5 Å. Due to the abundance of distorted, nearly degenerate structures that made the minima ranking unreliable, we excluded this compound from Table II and used it only in the calculation of ranking RMSE and proximity metrics. For MgB3_3C3_3, proposed to be a high-TcT_{\rm c} superconductor in hP7 [87] or hP14 [41] honeycomb layer structures potentially accessible through deintercalation of the known MgB2_2C2_2 compound, prior ab initio searches located an mS28 phase with an alternative layered morphology considered more stable by a set of DFT functionals [41]. The current uMLP-accelerated global optimization generated an oI28 polymorph, shown in Figure 1, with a BC framework connected in 3D via sp3^3 links uniformly favored by all tested DFT approximations.

Stability of the mS24-CYI, tI10-Na2CN2, and oI28-MgB3C3 phases

Figure 1. Stability of the mS24-CYI, tI10-Na2_2CN2_2, and oI28-MgB3_3C3_3 phases identified in this work relative to the reported mS12-CYI, mS10-Na2_2CN2_2, mS28-MgB3_3C3_3, and hP14-MgB3_3C3_3 evaluated with common DFT functionals.

The quantitative correspondence between surrogate and DFT low-energy landscapes was assessed using merged pools and the three metrics described in Section II C. Table S1 shows a relatively large dispersion in the number of candidate phases in each uMLP’s original set, e.g., between 14 for MC* and 45 for EN in Li3_3Sn, which illustrates why aggregating relevant minima was necessary for evaluating the uMLP performance on a more comparable basis. MG was treated separately because its native pools were unreasonably large (see Table S1).

To enable a controlled comparison of alternative DFT flavors and the MG model, we chose the representative merged EN set. For MG, this choice is favorable because it excludes numerous artifacts generated in its own searches, but the metrics below show that its ranking fidelity remains poor. The resulting test sets for 11 compounds (not including AgClO4_4) were mostly comparable in size, varying between 500 and 870 structures, with the only notable exception (EQ*) exceeding 2,000 candidates. The inflated EQ* pools contained a large number of distorted structure variants, which complicated duplicate removal and suggests practical inefficiency in evolutionary searches that rely on compact candidate lists. The similar trends obtained from simple and size-weighted averages, with weights Nminimacompound/NminimatotalN_{\mathrm{minima}}^{\mathrm{compound}}/N_{\mathrm{minima}}^{\mathrm{total}}, indicate that the conclusions were not controlled by the largest sets.

Fig. 2 summarizes the pool size-weighted metrics measuring the proximity of uMLP minima to the corresponding DFT counterparts and ranking fidelity within a low-energy shortlist. The changes in the radial distribution function (RDF) fingerprint and the DFT energy upon re-optimization span 0.43–0.76 and 1–8 meV/atom, respectively, with EQ, PT, and EN showing particularly small deviations. The fact that the average DFT relaxation produces only a modest and fairly systematic energy drop is important for downstream selection strategies that aim to improve effective ranking. For instance, our study on Au clusters [80] showed that a hybrid workflow, in which structures are relaxed with a NN potential while final energies are evaluated with single-shot DFT, can markedly improve convergence to the ground state,

@bibliography 񟿿 and that the added DFT cost during the evolutionary run can be offset by the reduction in pool sizes.

Some of the ranking RMSEs, ranging between 3 and 22 meV/atom, are notably lower than the overall errors typically reported for uMLPs, e.g., around 20 meV/atom in MatBench tests [25]. This observation is consistent with the different functional purpose of this metric that involves only competing structures near the ground states and explicitly removes the average uMLP-DFT energy offset in each pool. Fig. S4 breaks down the ranking RMSE by compound and shows that the values are reasonably consistent across distinct chemical classes. The violin plots presented in Fig. S5 underscore a general absence of outliers that could dominate and skew the computed metric. Fig. S6 shows that the ranking RMSEs vary little between native and merged pools, differing within 1 meV/atom for most models. The results show that relative energy accuracy correlates with the uMLP fitted-parameter count listed in Table I. Three models outperforming this trend are the medium-tier EQ and EN models with accuracies close to the large-sized EQ* and PT models while the small-tier PT* approaches the level of the medium-sized uMLPs.

The computed MRR values of 0.82–0.91 for all but three uMLPs indicate remarkable consistency among most models in ranking the DFT ground state near the top of the candidate pool. The contribution from the artifact-dominated AgClO₄ case was set to zero. The confidence-interval half-widths of 0.18–0.25 do not allow for a precise ordering within this band. MC* and GR, with lower values of 0.70 ± 0.25 and 0.71 ± 0.25 respectively, suggest less consistent ranking fidelity, although their intervals still overlap with the main group. MG remains the clear exception despite being evaluated on the favorable EN-derived candidate set: its much lower value of 0.32 ± 0.22 indicates that the model frequently places the ground state deep in the pool. The GSM values show a similarly compact main group, with most models falling between 1.0 and 1.4 with intervals ranging from 0.2 to 0.8 while EQ* is a clear redundancy outlier at a high 13.4 ± 9.4 value. The full MRR and GSM values are reported in the Summary.

Finally, it is instructive to compare the uMLPs against our system-specific Behler-Parrinello NN potentials fitted in previous studies for two M-Sn binaries [11, 12], with the reported total RMSE and the current ranking RMSEs of 10.2 meV/atom and 9.2 meV/atom for Li-Sn, and 9.6 meV/atom and 14.9 meV/atom for Pd-Sn, respectively. Based on the stated 45% rate of matching the DFT global minima in our study of Li-Sn alloys [11], our fitted model would have a fairly modest estimated MRR score between 0.45 and 0.73. Nearly all uMLPs fare better, highlighting the benefit of training larger architectures with more flexible descriptors on more extensive datasets across the periodic table to capture transferable stability trends across both structures and chemistries.

Ground state perturbations

Ground state of Zn

Among hcp metals, Zn is known to exhibit a c/ac/a ratio significantly deviating from the ideal close-packed 8/3≈1.633\sqrt{8/3} \approx1.633 value, an anomalous geometry attributed to the underlying electronic structure [88, 89, 90]. Häussermann and Simak [88] argued that the axial anisotropy is driven by a band-energy gain due to an ss–pp hybridization that opens up a density of states (DOS) pseudogap and places it near EFE_{\mathrm{F}}. The demonstrated synchronized alignment of multiple hybridization gaps from opposite directions suggests that the structural distortion is a manifestation of an electronic topological transition triggered by the varying lattice parameters. To illustrate that this stabilization is tied directly to symmetry-controlled hybridization features at EFE_{\mathrm{F}}, we compare band structures of hcp- and fcc-Zn in Figure 3. Representation of both close-packed crystal structures with hexagonal unit cells at the experimental c/a=1.826c/a = 1.826 ratio extrapolated to zero Kelvin [89] helps appreciate that it is the stacking sequence that determines the band splitting at the L and M k-points, which ultimately leads to the appearance of a pseudogap near EFE_{\mathrm{F}}, present in hcp-Zn (ABABAB) but absent in fcc-Zn (ABCABC).

Band structure and density of states of hcp-Zn and fcc-Zn

Figure 3. Band structure and density of states calculated with PBE for hcp-Zn (black) and fcc-Zn (red) represented with 6-atom hexagonal unit cells at the experimental c/a=1.826c/a = 1.826.

We examined the ability of the uMLPs to reproduce the structural anomaly in hcp-Zn by first relaxing the volume with each model and then performing single-point calculations for c/ac/a ratios around 1.826 at the fixed optimal volume. This constraint allowed us to sample the shallow trough on the PES over a wide range of structural parameters with a variation of ∼20\sim20 meV/atom, the energy scale of interest (see Fig. S7). Figure 4 shows that the three DFT approximations correctly predict the departure from the ideal packing and place the c/ac/a minimum at 1.894 (PBE), 1.826 (PBEsol), and 1.840 (r2SCAN), but most uMLPs have difficulty matching their reference methods. SN, EQ, and PT reproduce the large c/ac/a ratio but show notable deviations from the PBE profile at lower c/ac/a values. EN, GR, MS, and MC are aware of the energy basin’s peculiar shape but do not properly favor the structural distortion, yielding nearly flat curves with a ∼2\sim2 meV/atom variation between the ideal and the experimental ratios. OR fails to recognize the stabilizing effect of the distortion and defines a pronounced minimum around the ideal c/ac/a. MG agrees with PBE qualitatively, predicting a large c/ac/a optimal ratio, but the feature appears fortuitous because of the presence of a second comparable minimum at unphysically small c/ac/a around 1.42. Fig. S8 illustrates a clearly better performance of the new PT, MC, and EQ parameterizations, as PT* and MC* have deep minima around the ideal c/ac/a ratio while EQ* produces ∼5\sim5 meV/atom jumps along the trajectory (the observed artifact is consistent with the findings of Fu et al. [55] who noted that grid-based spherical harmonic projections in EQ* and eSCN can introduce discretization artifacts that degrade both smoothness and energy conservation). Our own attempts to accurately reproduce the dependence with a Behler-Parrinello NN potential constructed with our standard protocol [16] failed even after the distorted configurations were heavily weighted in the reference data set and the regularization factor was relaxed, with best versions (not shown) displaying EN-like behavior.

DFT and uMLP energy profiles of hcp-Zn

Figure 4. DFT and uMLP energy profiles of hcp-Zn as a function of c/ac/a for fixed-volume unit cells. Each curve is set to zero at the experimental c/a=1.826c/a = 1.826 ratio.

Competing MB4_4 phases for 3d metals

Ground states of 3d metal tetraborides feature a complex morphology of interlinked B4_4 units (Figure 5). The shape of the 3D B framework and the arrangement of the metal atoms are sensitive to the electron count and remained misidentified in CrB4_4 and MnB4_4 long after their discovery. In 2010, an ab initio evolutionary search for FeB4_4 found a new oP10 structure [81, 91], a distorted derivative of the oI10 CrB4_4 prototype, to be thermodynamically stable under moderate pressures; the phase was subsequently synthesized at pressures above 8 GPa and quenched to ambient conditions [92]. An ab initio reexamination of CrB4_4 revealed that the compound also prefers the oP10 polymorph over the originally assigned oI10; the prediction was later confirmed by single-crystal XRD measurements [93, 94]. Similarly, several DFT studies reanalyzed MnB4_4 and proposed a lower-symmetry mP20 ground state [85, 95, 96]; the refined solution was verified experimentally as well [95, 96]. The definitive determination of ground states proved essential for this materials family, as FeB4_4 and MnB4_4 have been shown to superconduct at 2.7 K (1 bar) [92] and 14.2 K (150 GPa) [97], respectively. While all three compounds are non-magnetic, with ferromagnetic order vanishing in the Fe-B binary around the dilute 1:4 metal concentration [85], spin fluctuations appear to play an important role in the pairing interaction, suppressing the conventional TcT_{\mathrm{c}} in FeB4_4 and possibly contributing to the high TcT_{\mathrm{c}} in MnB4_4. Hence, the electronically frustrated MB4_4 compounds present a trying test case for the capability of uMLPs to resolve competing crystal structures.

Stability of distorted oP10 and mP20 derivatives relative to oI10 for MB4 compounds

Figure 5. Stability of distorted oP10 and mP20 derivatives relative to oI10 calculated with DFT and uMLPs for MB4_4 with (a) M = Cr, (b) M = Mn, and (c) M = Fe. The oI10 and oP10 structures are displayed with supercells to illustrate the connection with the lowest-symmetry mP20.

As a baseline, we examined how commonly used DFT flavors, PBE, r2r^{2}SCAN, and PBEsol, describe the relative stability of the three configurations. The three approximations correctly predict that CrB4_4 and FeB4_4 benefit from the oI10→\tooP10 transformation that skews the B4_4 rectangles and that MnB4_4 further stabilizes through the oP10→\tomP20 symmetry breaking that dimerizes Mn atoms along the metal linear chains. Figure 5 and Table S3 show that the energy gains agree to within 3–5 meV/atom: 5.4, 8.6, and 7.6 meV/atom for CrB4_4, 25.7, 25.1, and 27.9 meV/atom for FeB4_4, and 28.5, 32.4, and 33.0 meV/atom for MnB4_4, respectively. In CrB4_4, the higher stabilization found for r2r^{2}SCAN and PBEsol reflects the larger level of the B4_4 distortion.

Local optimizations of the three prototypes with each uMLP were performed starting from conventional PBE-relaxed unit cells, and Figure 5 displays the energy of the oP10 and mP20 derivatives relative to the undistorted oI10 structure in each compound. For FeB4_4, all uMLPs capture the dynamical instability of the oI10 prototype with respect to the Γ\Gamma-point B phonon mode which generates oP10 with an unexpected increase of the DOS at the Fermi level [81]. However, MG shows no measurable energy gain, GR underestimates the stabilization by a factor of five, MS overestimates it twofold, and MC, SN, and MS favor further monoclinic distortion. The full relaxation of the oP10 prototype reduces the right angle in the B4_4 units to about 60∘60^\circ and increases the coordination of half of the B atoms by bringing a neighbor within a near-covalent 1.9 A˚\text{\AA} bond length in all but the MG parametrization (see Table S3). For CrB4_4, EQ, EN, MC, SN, PT, and MS are able to reproduce the favorability of the oP10 prototype. For MnB4_4, GR, OR, and MG find oI10 to be locally stable, SN, MS and MC improve it to the oP10 derivative, and EN, EQ, and PT correctly identify mP20 as the ground state. The three models are in quantitative agreement with PBE, matching the stabilization gain to within 15 meV/atom and the dimerized Mn-Mn nearest distances, which contract from 2.93 A˚\text{\AA} in oI10 down to 2.65 A˚\text{\AA} in mP20, to within 1%. As shown in Fig. S9, all three legacy variants perform well describing CrB4_4, but PT∗^{*} and MC∗^{*} fail to reproduce the mP20 stabilization for MnB4_4 and PT∗^{*} and EQ∗^{*} incorrectly favor the mP20 distortion for FeB4_4. Overall, the latest EQ and PT versions improve the description across the MB4_4 series, whereas MC displays more pronounced deviations from the reference method.

These tests illustrate that some of the examined uMLPs could successfully serve as surrogate models in unconstrained searches to locate the low-symmetry ground states in this materials family. The distorted prototypes would be correctly identified to have the lowest energy at the DFT level with a simple re-optimization of the candidate pool rather than an expensive phonon dispersion analysis. EQ, EN, and PT perform best by reproducing the distorted derivatives for all three MB4_4 compounds; SN and MC correctly identify the ground state only for CrB4_4; GR and OR are able to recognize the favorability of the symmetry breaking only for FeB4_4; MS points to oI10 instability across all three MB4_4 compounds without clearly preferring any of the correct ground states; and MG predicts a very shallow basin around the oI10 parent structure in this materials family.

Off-stoichiometry LiBy_y phases

The structure and the composition of LiBy<1_{y<1}, discovered in the late 1970s [98], remained a puzzle for decades until an in-depth XRD analysis showed that the compound is comprised of linear B chains and a Li sublattice with incommensurate periods [99]. Subsequent modeling of LiBy_y with structures of varying sublattice ratios revealed that their formation energies follow a parabolic dependence on the composition [100], which defines a finite range of stability around the experimentally observed y=0.9y = 0.9 (x=(1+y)−1=0.526x = (1+y)^{-1} = 0.526). While deviations from ideal stoichiometries are common at elevated temperatures due to the configurational entropy contribution to the formation free energy, the existence of a finite stability range at zero temperature is unique. According to the ab initio analysis [100], the Li to B charge transfer does not follow a simple rigid band shift scenario because Li preferentially downshifts px,yp_{x,y} B orbitals, and the resulting DOS pseudogap moves close to the Fermi level at the observed xx values. The established parabolic dependence has helped explain the material’s response to H2_2 exposure [101], track its composition under compression [100, 102, 103], and prime its proposed function as a precursor for synthesis of a layered LiB superconductor [104].

The material is a particularly valuable testbed because the off-stoichiometric ground state phases, not yet added to major materials databases, have not been included in the uMLP training sets. For example, the Materials Project has only one suboptimal representative of the material, α\alpha-LiB. Figure 6 and Table S4 illustrate how we used the Li2n_{2n}Bm_m series to probe whether uMLPs are capable of capturing the stability range, with xmin⁡x_{\min} and xmax⁡x_{\max} defined by tangents to the parabolic fits 12c(x−x0)2+E0\frac{1}{2}c(x-x_0)^2+E_0 from LiB3_3 and Li points on the formation energy plot, respectively.

Stability ranges for Li_xB_{1-x} phases with linear B chains

Figure 6. Stability ranges for Lix_xB1−x_{1-x} phases with linear B chains for DFT and uMLPs. The inset shows the parabolic fit to the PBE formation energies, with xmin⁡x_{\min} and xmax⁡x_{\max} values determined as parabola tangents connecting LiB3_3 and Li, respectively.

Our DFT results reflect the known strong dependence of formation energies on the exchange-correlation functional: for α\alpha-LiB and LiB3_3, values vary up to nearly 100 meV/atom. For this reason, our uMLP benchmarking focuses primarily on properties that are either intrinsic to the LiBy_y family, such as the energy difference between α\alpha and β\beta LiB phases with alternative relative placement of the B and Li sublattices, or at least invariant under affine transformations (E′=E+A+BxE'=E+A+Bx), such as the fitted parabola curvature and the ranges of stability (see Supplementary Note I). The three DFT approximations indeed produce closely matching values for ΔEα-LiB=9.2−10.3\Delta E_{\alpha\text{-LiB}}=9.2-10.3 meV/atom and c=72−75c=72-75 eV/atom, but the [xmin⁡,xmax⁡][x_{\min},x_{\max}] estimates show a larger dispersion, with r2^2SCAN and PBEsol defining non-overlapping ranges different in size by a factor of two (Figure 6).

The uMLPs correctly favor β\beta over α\alpha but, with the exception of GR and PT, the energy difference is noticeably underestimated, especially for MS, SN, and PT*. The fitted curvatures are also consistently softer, at 60–85% of the PBE value. The results show that the majority of these classical models do not sufficiently penalize the sublattice shifts away from the natural Li-B registry or the compositional deviations from the optimal stoichiometry. At the same time, the PBE stability ranges are reproduced by the uMLPs with accuracy that is better (EN, MC*, OR, and EQ*) or comparable (SN, MC, PT, EQ, GR, and MS) to that of the PBEsol or r2^2SCAN estimates, while PT* is in reasonable agreement with the PBEsol reference. The fact that all uMLPs place the minimum of the LiBy_y formation energy curve in the Li-rich region, in quantitative agreement with the ab initio methods, could be an indication that the stable stoichiometries are determined predominantly by classical factors, e.g., the ratio of the effective Li and B sizes [100], and that the redistribution of electronic states is consequential rather than causal.

For assessing the quality of the LiBy_y description further away from equilibrium, we also calculated the relative stability between β\beta-LiB and a layered hP8-LiB phase. The latter was predicted to stabilize under pressure [105], and was later synthesized via cold compression and annealed to ambient conditions [103]. The conversion apparently follows a low-barrier pathway linking B chains into honeycomb B layers, a process that requires large-scale simulations to model. uMLPs could therefore shed light on the kinetics of the solid-state transformation, provided that they accurately resolve the relative energy between the hP8-LiB and β\beta-LiB endpoints. In practice, the values vary substantially even across the DFT approximations (−3.5-3.5 meV/atom for PBE, +26.6+26.6 meV/atom for r2^2SCAN, and −25.6-25.6 meV/atom for PBEsol) and disperse over a wider range with uMLPs, with −110-110 meV/atom for MC* and +153+153 meV/atom for MG (see Table S4). Among the considered models, only EN, MC, GR, OR, and PT* reproduce their respective references closely enough to distinguish reliably between the two competing phases with markedly different morphologies. Overall, the updated PT, MC, and EQ variants do not yield a uniform improvement in the LiBy_y description over their legacy counterparts (see Fig. S10).

Summary

This study presents a systematic benchmark of nine state-of-the-art uMLPs and three legacy models through evolutionary crystal structure prediction. Unlike traditional testing frameworks that often rely on libraries of commonly observed high-symmetry structures, our protocol probed the models’ readiness for practical unconstrained explorations that have a track record in the accelerated discovery of new thermodynamically stable prototypes [11, 12]. By focusing on multicomponent materials, we evaluated the uMLPs’ ability to describe diverse bonding mechanisms and complex ground states. The inclusion of three compounds with recently proposed ground states, Li3_3Sn, Pd5_5Sn3_3, and MgB3_3C3_3, fur-

ther tested whether out-of-the-box uMLPs can handle configurations absent from their training sets.

The results reveal an impressive level of predictive maturity across modern uMLPs. With the exception of the earliest architecture (MG), the considered uMLPs successfully managed unconstrained searches, locating ground states for most compounds (see Table II and Figure 2). The remarkably low rate of unphysical artifacts indicates that current training sets provide a robust baseline for inorganic chemistry. AgClO4_4 was the sole compound where all models struggled, likely due to insufficient sampling of specific oxygen bonding environments.

Performance metrics of nine uMLPs on merged pools

Figure 2. Performance metrics of nine uMLPs on merged pools assessed relative to the reference DFT method, and averaged over 11 compounds, excluding AgClO₄. The comparison of EQ*, PT*, and MC* against their latest counterparts is presented in Fig. S3. The schematics at the top clarify the definitions of the proximity metrics and ranking RMSE introduced in the text. The representative merged EN pools were used to evaluate the metrics for the alternative DFT approximations and MG, with dashed horizontal lines marking the ranking RMSE for PBEsol and r²SCAN relative to PBE. The energy proximity average for MG exceeded 60 meV/atom (see Fig. S4 for further discussion).

Our combined analysis of over three million uMLP local relaxations and more than ten thousand DFT calculations summarized in Table III allows for a more contextual understanding of model behavior that a single accuracy metric cannot capture alone. For example, EQ* has the second-lowest ranking RMSE, yet it generated numerous redundant candidates with an average GSM of 13.4±9.413.4 \pm9.4, and, therefore, has limited practical value for evolutionary searches. At the other end of the accuracy spectrum, MG exhibits RMSE values typical of smaller uMLPs and comparable to the DFT systematic spread, but proves to be singularly ineffective in nearly all unconstrained runs.

TABLE III. Aggregate performance of the considered uMLPs with respect to their reference DFT method. The models are ordered by the ranking RMSE shown in Figure 2 and split into three size groups by the fitted-parameter count listed in Table I. Accuracy metrics averaged over the identified pools of low-energy candidates are shown in Figure 2. The MRR scores measure the fidelity of ranking the ground state, while the GSM reflects the degree of uMLP ground state redundancy, with both quantities defined in Section II C. Description quality of near-equilibrium configurations for selected compounds is assessed with a composite score on a 12-point scale. For hcp-Zn, the evaluation criteria are the location of the global minimum at the anomalous c/ac/a ratio, PES smoothness, and absence of pronounced secondary minima (Figure 4). For the MB4_4 family, the grades reflect the qualitative identification of the ground-state polymorph (mP20 for Mn and oP10 for Cr and Fe) and the quantitative agreement, within 10 meV/atom, of the distortion-driven stabilization (Figure 5 and Table S3). For LiBy_y, the scores indicate the favorability of Li-rich phases (y≈0.9y \approx0.9) and reproduction of the relative stabilities among relevant LiB phases (Figure 6 and Table S4).

uMLPMerged pool metricsGround state perturbations
NameIDSizeRMSEProx(E)Prox(S)MRRGSMhcp-ZnMB4_4LiBy_y
EquiformerV3EQMedium3.410.330.760.91 ± 0.181.1 ± 0.411128
EquiformerV2EQ*Large4.221.170.690.82 ± 0.2413.4 ± 9.46127
PET-OAMPTLarge4.310.510.760.91 ± 0.181.2 ± 0.311108
eSENENMedium4.490.410.750.91 ± 0.181.1 ± 0.281211
MACE-mh-1MCMedium9.172.910.660.84 ± 0.211.1 ± 0.29411
Orb-v3ORMedium9.642.450.660.91 ± 0.181.1 ± 0.22211
SevenNetSNMedium9.922.900.630.91 ± 0.181.4 ± 0.81266
GRACEGRMedium11.682.290.630.71 ± 0.251.1 ± 0.28212
PET-MADPT*Small15.946.500.540.91 ± 0.181.4 ± 0.44610
MACE-MATPESMC*Medium18.285.990.550.70 ± 0.251.1 ± 0.2469
MatterSimMSSmall21.667.960.560.86 ± 0.191.0 ± 0.0606
M3GNetMGSmall22.45†65.0†0.44†0.32† ± 0.221.1† ± 0.2006

Table III.

Rather unexpectedly, the extensive structure searches carried out purely for benchmarking purposes identified three phases lying below the previously reported global minima at the reference DFT level. The TlIO-

Na2_2CN2_2 candidate is favored only by PBE, whereas the mS24-CY1 candidate is stabilized by PBE and r2^2SCAN but not by PBEsol or vdW-corrected treatments, making both likely approximation-dependent artifacts. The oI28-MgB2_2C3_3 phase ranks below the proposed layered superconducting polymorph [41] across all considered exchange-correlation functionals, indicating that deintercalation of the MgB2_2C2_2 precursor may lead to products with morphologies unrelated to the starting honeycomb framework.

We also examined whether current uMLPs can map subtle near-equilibrium PES features arising from electronic structure peculiarities. As in our multi-metric evaluation of global optimization performance, we assigned composite qualitative and quantitative scores to assess reproduction of the relevant observables described in the caption of Table III. In this case, we focused on prospective simulations probing structural response and phase transformations on scales for which direct DFT validation may be prohibitively expensive.

For the well-studied hcp-Zn, most uMLPs struggled to reproduce the shallow basin around the anomalous c/ac/a ratio defined by the electronic band topology [88]. Only SN, EQ, and PT closely follow the reference energy profile and could be used in MD simulations to model, in particular, the observed c/ac/a response to temperature [89]. For the CrB4_4, FeB4_4, and MnB4_4 compounds, several uMLPs correctly reproduce the chemistry-dependent symmetry lowering among the competing oI10, oP10, and mP20 polymorphs. The performance is encouraging for future large-scale exploration of the (Cr, Mn, Fe)B4_4 phase space aimed at identifying synthesizable superconducting and superhard materials without exhaustive ab initio stability analyses for every distortion pattern [81, 85, 91]. For off-stoichiometric LiBy_y, all uMLPs correctly place the stability minimum in the Li-rich region near y≈0.9y \approx0.9, despite standard databases still listing LiB as the ground-state composition. Detailed modeling of the experimentally observed structural transformation from the ambient-pressure precursor to a layered phase with potential for high-TcT_c superconductivity [103, 104] demands accurate description of the competing Li-B phases. While several uMLPs agree with the reference method, the common semi-local DFT approximations themselves do not account for the vdW interactions important in this materials class.

Overall, the benchmark indicates that current pretrained uMLPs can already boost ab initio global optimization of nonmagnetic inorganic crystal structures under ambient pressure and accurately describe selected near-equilibrium PES features, with EQ, PT, and EN delivering the most consistent performance across the tested cases. The results are consistent with the general trend that larger models with more expressive descriptors and broader training coverage transfer more reliably across competing structures and chemistries. In this respect, system-specific potentials may no longer need to be the default starting point for crystal structure exploration. Any application to a specific materials problem should still be preceded by case-specific validation and, when needed, retraining against an appropriate reference set. This caution is especially important because no exchange-correlation approximation is universally accurate across the periodic table.

Conflicts of interest

There are no conflicts to declare.

Data Availability

The data supporting the findings of this study and descriptions of our workflows (Supplementary Note II) are available within the Supplemental Material and in the Zenodo repository at 10.5281/zenodo.20880271. The open-source MAISE code can be acquired at https://github.com/maise-guide/maise. Additional data are available from the corresponding author upon request.

Supplementary Material: uMLP single-core CPU efficiency tests, evolutionary search and native minima pool details, merged pool details, representative AgClO4_4 artifact, performance metrics for legacy models, ranking RMSEs per compound for merged pools, violin plots for pool-based energy differences between uMLPs and DFT, native and merged pool ranking RMSE comparison, PBE PES for hcp-Zn, ground state perturbation tests for legacy models, stability and structural information of MB4_4 phases (M = Cr, Mn, and Fe), stability of Lix_xB1−x_{1-x} phases, LiBy_y properties under affine transformations, example MAISE settings for evolutionary searches, BLOCK file for MgB3_3C3_3 TETRIS generation, evolutionary optimization and perturbation calculation workflows, information for used VASP pseudopotentials, structural information for select phases in CIF format.

Acknowledgments

The authors thank Aidan Thorn and Erin Delargy for their help with constructing neural network potentials for Zn and acknowledge support from the National Science Foundation (NSF) (Award No. DMR-2320073). This work used the Expanse system at the San Diego Supercomputer Center via allocation TG-DMR180071. Expanse is supported by the Extreme Science and Engineering Discovery Environment (XSEDE) program [106] through NSF Award No. ACI-1548562. The work of O.G. was supported by Grant #13001218-080 through SCALE: A Microelectronic Workforce Development Production Project.

References

  1. [1]J. Behler and M. Parrinello, Generalized neural-network representation of high-dimensional potential-energy surfaces, Phys. Rev. Lett. 98, 146401 (2007).pubmed.ncbi.nlm.nih.gov/17501293
  2. [2]A. P. Bartók, M. C. Payne, R. Kondor, and G. Csányi, Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons, Phys. Rev. Lett. 104, 136403 (2010).pubmed.ncbi.nlm.nih.gov/20481899
  3. [3]V. L. Deringer, M. A. Caro, and G. Csányi, Machine learning interatomic potentials as emerging tools for materials science, Adv. Mater. 31, 1902765 (2019).DOI
  4. [4]G. L. W. Hart, T. Mueller, C. Toher, and S. Curtarolo, Machine learning for alloys, Nat. Rev. Mater. 6, 730 (2021).DOI
  5. [5]Y. Mishin, Machine-learning interatomic potentials for materials science, Acta Materialia 214, 116980 (2021).DOI
  6. [6]S.-D. Huang, C. Shang, P.-L. Kang, and Z.-P. Liu, Atomic structure of boron resolved using machine learning and global sampling, Chem. Sci. 9, 8644 (2018).DOI
  7. [7]E. V. Podryabinkin, E. V. Tikhonov, A. V. Shapeev, and A. R. Oganov, Accelerating crystal structure prediction by machine-learning interatomic potentials with active learning, Phys. Rev. B 99, 064114 (2019).DOI
  8. [8]J. Behler, R. Martoňák, D. Donadio, and M. Parrinello, Metadynamics simulations of the high-pressure phases of silicon employing a high-dimensional neural network potential, Phys. Rev. Lett. 100, 185501 (2008).pubmed.ncbi.nlm.nih.gov/18518388
  9. [9]V. L. Deringer, D. M. Proserpio, G. Csányi, and C. J. Pickard, Data-driven learning and prediction of inorganic crystal structures, Faraday Discuss. 211, 45 (2018).DOI
  10. [10]W. Ibarra-Hernández, S. Hajinazar, G. Avendaño-Franco, A. Bautista-Hernández, A. N. Kolmogorov, and A. H. Romero, Structural search for stable Mg–Ca alloys accelerated with a neural network interatomic model, Phys. Chem. Chem. Phys. 20, 27545 (2018).
  11. [11]S. Kharabadze, A. Thorn, E. A. Koulakova, and A. N. Kolmogorov, Prediction of stable Li-Sn compounds: boosting ab initio searches with neural network potentials, npj Comput. Mater. 8, 136 (2022).DOI
  12. [12]A. Thorn, D. Gochitashvili, S. Kharabadze, and A. N. Kolmogorov, Machine learning search for stable binary Sn alloys with Na, Ca, Cu, Pd, and Ag, Phys. Chem. Chem. Phys. 25, 22415 (2023).DOI
  13. [13]R. Ouyang, Y. Xie, and D.-e. Jiang, Global minimization of gold clusters by combining neural network potentials and the basin-hopping method, Nanoscale 7, 14817–14821 (2015).DOI
  14. [14]K. Gubaev, E. V. Podryabinkin, G. L. W. Hart, and A. V. Shapeev, Accelerating high-throughput searches for new alloys with active learning of interatomic potentials, Comput. Mater. Sci. 156, 148 (2019).arxiv.org/abs/1806.10567
  15. [15]S. Hajinazar, J. Shao, and A. N. Kolmogorov, Stratified construction of neural network based interatomic models for multicomponent materials, Phys. Rev. B 95, 014114 (2017).DOI
  16. [16]S. Hajinazar, A. Thorn, E. D. Sandoval, S. Kharabadze, and A. N. Kolmogorov, MAISE: Construction of neural network interatomic models and evolutionary structure optimization, Comput. Phys. Commun. 259, 107679 (2021).
  17. [17]P. E. Dolgirev, I. A. Kruglov, and A. R. Oganov, Machine learning scheme for fast extraction of chemically interpretable interatomic potentials, AIP Adv. 6, 085318 (2016).
  18. [18]N. Artrith, T. Morawietz, and J. Behler, High-dimensional neural-network potentials for multicomponent systems: Applications to zinc oxide, Phys. Rev. B 83, 153101 (2011).DOI
  19. [19]C. J. Pickard, Ephemeral data derived potentials for random structure search, Phys. Rev. B 106, 014102 (2022).DOI
  20. [20]C. Chen and S. P. Ong, A universal graph deep learning interatomic potential for the periodic table, Nat. Comput. Sci. 2, 718 (2022).DOI
  21. [21]B. Póta, P. Ahlawat, G. Csányi, and M. Simoncelli, Thermal conductivity predictions with foundation atomistic models, arXiv preprint arXiv:2408.00755 10.48550/arXiv.2408.00755 (2024).arxiv.org/abs/2408.00755
  22. [22]K. Sharma, A. Loew, H. Wang, F. A. Nilsson, M. Jain, M. A. L. Marques, and K. S. Thygesen, Accelerating point defect photo-emission calculations with machine learning interatomic potentials, npj Comput. Mater. 11, 18201 (2025).arxiv.org/abs/2505.01403
  23. [23]H. Yu, M. Giantomassi, G. Materzanini, J. Wang, and G.-M. Rignanese, Systematic assessment of various universal machine-learning interatomic potentials, Mater. Genome Eng. Adv. 2, e58 (2024).arxiv.org/abs/2403.05729
  24. [24]A. Loew, D. Sun, H.-C. Wang, S. Botti, and M. A. L. Marques, Universal machine learning interatomic potentials are ready for phonons, npj Comput. Mater. 11, 10 (2025).arxiv.org/abs/2412.16551
  25. [25]J. Riebesell, R. E. A. Goodall, P. Benner, Y. Chiang, B. Deng, G. Ceder, M. Asta, A. A. Lee, A. Jain, and K. A. Persson, A framework to evaluate machine learning crystal stability predictions, Nat. Mach. Intell. 7, 836 (2025).DOI
  26. [26]Y. Chiang, T. Kreiman, C. Zhang, M. C. Kuner, E. Weaver, I. Amin, H. Park, Y. Lim, J. Kim, D. Chrzan, A. Walsh, S. M. Blau, M. Asta, and A. S. Krishnapriyan, MLIP Arena: Advancing Fairness and Transparency in Machine Learning Interatomic Potentials via an Open, Accessible Benchmark Platform, arXiv preprint arXiv:2509.20630 10.48550/arXiv.2509.20630 (2025).
  27. [27]H. Tahmasbi, A. Knüpfer, T. D. Kühne, and H. Mirhoseini, Benchmarking Universal Machine Learning Interatomic Potentials on Elemental Systems, arXiv preprint arXiv:2512.20230 10.48550/arXiv.2512.20230 (2025).arxiv.org/abs/2512.20230
  28. [28]A. Yang, C. Cai, M. Guo, D. Zhang, C. Zhang, W. Jiang, Y. Wang, A. Loew, C. Wu, W. E, L. Zhang, and H. Wang, Lambench: a benchmark for large atomistic models, npj Comput. Mater. 12, 62 (2026).
  29. [29]A. Loew, J. Schmidt, S. Botti, and M. A. L. Marques, Universal machine learning potentials under pressure, Journal of Physics: Materials 9, 015010 (2025).
  30. [30]P. Lyngby, C. Larsen, and K. W. Jacobsen, Bayesian optimization of atomic structures with prior probabilities from universal interatomic potentials, Phys. Rev. Mater. 8, 10.1103/physrevmaterials.8.123802 (2024).arxiv.org/abs/2408.15590
  31. [31]G. Mair, H.-G. von Schnering, M. Wörle, and R. Nesper, Dilithium Hexaboride, Li₂B₆, Z. Anorg. Allg. Chem. 625, 1207 (1999).doi.org/10.1002/(sici)1521-3749(199907)625:7<1207::aid-zaac1207>3.0.co;2-9
  32. [32]S. K. Choi, R. Coldea, A. N. Kolmogorov, T. Lancaster, I. I. Mazin, S. J. Blundell, P. G. Radaelli, Y. Singh, P. Gegenwart, K. R. Choi, S.-W. Cheong, P. J. Baker, C. Stock, and J. Taylor, Spin Waves and Revised Crystal Structure of Honeycomb Iridate Na₂IrO₃, Phys. Rev. Lett. 108, 127204 (2012).pubmed.ncbi.nlm.nih.gov/22540621
  33. [33]H. J. Becher and A. Schäfer, Darstellung und Struktur des Berylliumborids Be₄B, Z. Anorg. Allg. Chem. 318, 304 (1962).DOI
  34. [34]V. Hlukhy and R. Pöttgen, The Hexagonal Laves Phase MgIr₂, Zeitschrift für Naturforschung B 59, 943–946 (2004).
  35. [35]G. Eguchi, D. C. Peets, M. Kriener, Y. Maeno, E. Nishibori, Y. Kumazawa, K. Banno, S. Maki, and H. Sawa, Crystallographic and superconducting properties of the fully gapped noncentrosymmetric 5d-electron superconductors CaMSi₃ (M = Ir, Pt), Phys. Rev. B 83, 024512 (2011).DOI
  36. [36]C. J. Howard, T. M. Sabine, and F. Dickson, Structural and thermal parameters for rutile and anatase, Acta Crystallogr. Sect. B: Struct. Sci. 47, 462 (1991).DOI
  37. [37]R. W. Henn, W. Schnelle, R. K. Kremer, and A. Simon, Bulk Superconductivity at 10 K in the Layered Compounds Y₂C₂I₂ and Y₂C₂Br₂, Phys. Rev. Lett. 77, K. Ahn, R. K. Kremer, A. Simon, W. G. Marshall, and A. Muñoz, Pressure-structure relationships in the 10 K layered carbide halide superconductors Y2C2I2, Journal of Physics: Condensed Matter 28, 375703 (2016).
  38. [39]H. J. Berthold, J. M. Molepo, and R. Wartchow, Die Kristallstruktur des AgClO4, Zeitschrift für Kristallographie - Crystalline Materials 144, 116–125 (1976).
  39. [40]M. Becker, J. Nuss, and M. Jansen, Synthese und Charakterisierung von Natriumcyanamid, Zeitschrift für anorganische und allgemeine Chemie 626, 2505 (2000).
  40. [41]C. R. Tomassetti, D. Gochitashvili, C. Renskers, E. R. Margine, and A. N. Kolmogorov, First-principles design of ambient-pressure MgxB2C2 and NaxBC superconductors, Phys. Rev. Mater. 8, 114801 (2024).
  41. [42]I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner, and G. Csányi, MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields, in Adv. Neural Inf. Process. Syst. (NeurIPS), Vol. 35 (2022) pp. 11423–11436.
  42. [43]I. Batatia, S. Batzner, D. P. Kovács, A. Musaelian, G. N. C. Simm, R. Drautz, C. Ortner, B. Kozinsky, and G. Csányi, The design space of E(3)-equivariant atom-centred interatomic potentials, Nature Machine Intelligence 7, 56–67 (2025).DOI
  43. [44]I. Batatia, C. Lin, J. Hart, E. Kasoar, A. M. Elena, S. W. Norwood, T. Wolf, and G. Csányi, Cross learning between electronic structure theories for unifying molecular, surface, and inorganic crystal foundation force fields, arXiv preprint arXiv:2510.25380 10.48550/arXiv.2510.25380 (2025).arxiv.org/abs/2510.25380
  44. [45]Y. Park, J. Kim, S. Hwang, and S. Han, Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations, J. Chem. Theory Comput. 20, 4857 (2024).DOI
  45. [46]J. Kim, J. Kim, J. Kim, J. Lee, Y. Park, Y. Kang, and S. Han, Data-Efficient Multifidelity Training for High-Fidelity Machine Learning Interatomic Potentials, J. Am. Chem. Soc. 147, 1042 (2024).DOI
  46. [47]Y.-L. Liao, B. Wood, A. Das, and T. Smidt, EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations, in International Conference on Learning Representations (ICLR) (2024).arxiv.org/abs/2306.12059
  47. [48]L. Barroso-Luque, M. Shuaibi, X. Fu, B. M. Wood, M. Dzamba, M. Gao, A. Rizvi, C. L. Zitnick, and Z. W. Ulissi, Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models, arXiv preprint arXiv:2410.12771 10.48550/arXiv.2410.12771 (2024).
  48. [49]Y.-L. Liao and T. Smidt, Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs, in International Conference on Learning Representations (ICLR) (2023).arxiv.org/abs/2206.11990
  49. [50]L. Chanussot, A. Das, S. Goyal, T. Lavril, M. Shuaibi, M. Riviere, K. Tran, J. Heras-Domingo, C. Ho, W. Hu, A. Palizhati, A. Sriram, B. Wood, J. Yoon, D. Parikh, C. L. Zitnick, and Z. Ulissi, Open Catalyst 2020 (OC20) Dataset and Community Challenges, ACS Catal. 11, 6059 (2021).DOI
  50. [51]Y.-L. Liao, A. J. Hoffman, S. C. Shen, A. Duval, S. W. Norwood, and T. Smidt, EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers, arXiv preprint arXiv:2604.09130 10.48550/arXiv.2604.09130 (2026).arxiv.org/abs/2604.09130
  51. [52]H. Yang, C. Hu, Y. Zhou, X. Liu, Y. Shi, J. Li, G. Li, Z. Chen, S. Chen, C. Zeni, M. Horton, R. Pinsler, A. Fowler, D. Zügner, T. Xie, J. Smith, L. Sun, Q. Wang, L. Kong, C. Liu, H. Hao, and Z. Lu, MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures, arXiv preprint arXiv:2405.04967 10.48550/arXiv.2405.04967 (2024).arxiv.org/abs/2405.04967
  52. [53]A. Bochkarev, Y. Lysogorskiy, and R. Drautz, Graph atomic cluster expansion for semilocal interactions beyond equivariant message passing, Phys. Rev. X 14, 021036 (2024).arxiv.org/abs/2311.16326
  53. [54]Y. Lysogorskiy, A. Bochkarev, and R. Drautz, Graph atomic cluster expansion for foundational machine learning interatomic potentials, arXiv preprint arXiv:2508.17936 10.48550/arXiv.2508.17936 (2025).arxiv.org/abs/2508.17936
  54. [55]X. Fu, B. M. Wood, L. Barroso-Luque, D. S. Levine, M. Gao, M. Dzamba, and C. L. Zitnick, Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction, arXiv preprint arXiv:2502.12147 10.48550/arXiv.2502.12147 (2025).arxiv.org/abs/2502.12147
  55. [56]B. Rhodes, S. Vandenhaute, V. Simkus, J. Gin, J. Godwin, T. Duignan, and M. Neumann, Orb-v3: atomistic simulation at scale, arXiv preprint arXiv:2504.06231 10.48550/arXiv.2504.06231 (2025).
  56. [57]M. Neumann, J. Gin, B. Rhodes, S. Bennett, Z. Li, H. Choubisa, A. Hussey, and J. Godwin, Orb: A Fast, Scalable Neural Network Potential, arXiv preprint arXiv:2410.22570 10.48550/arXiv.2410.22570 (2024).arxiv.org/abs/2410.22570
  57. [58]A. Mazitov, F. Bigi, M. Kellner, P. Pegolo, D. Tisi, G. Fraux, S. Pozdnyakov, P. Loche, and M. Ceriotti, PET-MAD as a lightweight universal interatomic potential for advanced materials modeling, Nat. Commun. 16, 65662 (2025).
  58. [59]F. Bigi, P. Pegolo, A. Mazitov, and M. Ceriotti, Pushing the limits of unconstrained machine-learned interatomic potentials, arXiv preprint arXiv:2601.16195 10.48550/arXiv.2601.16195 (2026).arxiv.org/abs/2601.16195
  59. [60]J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized Gradient Approximation Made Simple, Phys. Rev. Lett. 77, 3865 (1996).pubmed.ncbi.nlm.nih.gov/10062328
  60. [61]J. P. Perdew, A. Ruzsinszky, G. I. Csonka, O. A. Vydrov, G. E. Scuseria, L. A. Constantin, X. Zhou, and K. Burke, Restoring the density-gradient expansion for exchange in solids and surfaces, Phys. Rev. Lett. 100, 136406 (2008).pubmed.ncbi.nlm.nih.gov/18517979
  61. [62]J. W. Furness, A. D. Kaplan, J. Ning, J. P. Perdew, and J. Sun, Accurate and numerically efficient r2SCAN meta-generalized gradient approximation, J. Phys. Chem. Lett. 11, 8208 (2020).DOI
  62. [63]G. Kresse and J. Hafner, Ab initio molecular dynamics for liquid metals, Phys. Rev. B 47, 558 (1993).
  63. [64]G. Kresse and J. Hafner, Ab initio molecular-dynamics simulation of the liquid-metal–amorphous-semiconductor transition in germanium, Phys. Rev. B 49, 14251 (1994).pubmed.ncbi.nlm.nih.gov/10010505
  64. [65]G. Kresse and J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set, Phys. Rev. B 54, 11169 (1996).pubmed.ncbi.nlm.nih.gov/9984901
  65. [66]G. Kresse and J. Furthmüller, Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set, Comput. Mater. Sci. 6, 15 (1996).DOI
  66. [67]P. E. Blöchl, Projector augmented-wave method, Phys. Rev. B 50, 17953 (1994).pubmed.ncbi.nlm.nih.gov/9976227
  67. [68]H. J. Monkhorst and J. D. Pack, Special points for Brillouin-zone integrations, Phys. Rev. B 13, 5188 (1976).DOI
  68. [69]P. E. Blöchl, O. Jepsen, and O. K. Andersen, Improved tetrahedron method for Brillouin-zone integrations, Phys. Rev. B 49, 16223 (1994).pubmed.ncbi.nlm.nih.gov/10010769
  69. [70]D. C. Langreth and M. J. Mehl, Beyond the local-density approximation in calculations of ground-state electronic properties, Phys. Rev. B 28, 1809 (1983).DOI
  70. [71]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, Commentary: The Materials Project: A materials genome approach to accelerating materials innovation, APL Mater. 1, 011002 (2013).DOI
  71. [72]J. Klimeš, D. R. Bowler, and A. Michaelides, Chemical accuracy for the van der Waals density functional, J. Phys.: Condens. Matter 22, 022201 (2009).
  72. [73]J. Klimeš, D. R. Bowler, and A. Michaelides, Van der Waals density functionals applied to solids, Phys. Rev. B 83, 195131 (2011).
  73. [74]J. Ning, M. Kothakonda, J. W. Furness, A. D. Kaplan, S. Ehlert, J. G. Brandenburg, J. P. Perdew, and J. Sun, Workhorse minimally empirical dispersion-corrected density functional with tests for weakly bound systems: r2SCAN + rVV10, Phys. Rev. B 106, 075422 (2022).DOI
  74. [75]D. Chakraborty, K. Berland, and T. Thonhauser, Next-Generation Nonlocal van der Waals Density Functional, J. Chem. Theory Comput. 16, 5893 (2020).
  75. [76]S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials, Nat. Commun. 13, 2453 (2022).DOI
  76. [77]S. Passaro and C. L. Zitnick, Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs, in Proc. Mach. Learn. Res. (PMLR), Vol. 202 (PMLR, 2023) pp. 27424–27438.arxiv.org/abs/2302.03655
  77. [78]T. W. Ko, B. Deng, M. Nassar, L. Barroso-Luque, R. Liu, J. Qi, A. C. Thakur, A. R. Mishra, E. Liu, G. Ceder, S. Miret, and S. P. Ong, Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry, npj Comput. Mater. 11, 253 (2025).arxiv.org/abs/2503.03837
  78. [79]A. H. Larsen, J. J. Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Dułak, J. Friis, M. Groves, B. Hammer, C. Hargus, E. Hermes, P. C. Jennings, P. B. Jensen, J. Kermode, J. R. Kitchin, E. Kolsbjerg, J. Kubal, K. Kaasbjerg, S. Lysgaard, J. B. Maronsson, T. Maxson, T. Olsen, L. Pastewka, A. T. Peterson, C. Rostgaard, J. Schiøtz, O. Schütt, M. Strange, K. S. Thygesen, T. Vegge, L. Vilhelmsen, M. Walter, Z. Zeng, and K. W. Jacobsen, The Atomistic Simulation Environment — A Python library for working with atoms, J. Phys. Condens. Matter 29, 273002 (2017).DOI
  79. [80]A. Thorn, J. Rojas-Nazar, S. Hajinazar, S. E. Baltazar, and A. N. Kolmogorov, Toward ab Initio Ground States of Gold Clusters via Neural Network Modeling, J. Phys. Chem. C 123, 30088 (2019).
  80. [81]A. N. Kolmogorov, S. Shah, E. R. Margine, A. F. Bialon, T. Hammerschmidt, and R. Drautz, New Superconducting and Semiconducting Fe-B Compounds Predicted with an Ab Initio Evolutionary Search, Phys. Rev. Lett. 105, 217003 (2010).DOI
  81. [82]R. Fasoulis, G. Paliouras, and L. E. Kavraki, RankMHC: Learning to Rank Class-I Peptide-MHC Structural Models, Journal of Chemical Information and Modeling 64, 8729–8742 (2024).DOI
  82. [83]J. Qu, Y. R. Xie, K. M. Ciesielski, C. E. Porter, E. S. Toberer, and E. Ertekin, Leveraging language representation for materials exploration and discovery, npj Comput. Mater. 10, 10.1038/s41524-024-01231-8 (2024).DOI
  83. [84]D. Gochitashvili, M. Meyers, C. Wang, and A. N. Kolmogorov, Improving structure search with hyperspatial optimization and TETRIS seeding, Phys. Chem. Chem. Phys. 27, 25636 (2025).arxiv.org/abs/2507.13791
  84. [85]A. Van Der Geest and A. Kolmogorov, Stability of 41 metal–boron systems at 0GPa and 30GPa from first principles, Calphad 46, 184 (2014).
  85. [86]Y. Zhang, J. W. Furness, B. Xiao, and J. Sun, Subtlety of TiO2 phase stability: Reliability of the density functional theory predictions and persistence of the self-interaction error, J. Chem. Phys. 150, 014105 (2019).DOI
  86. [87]T.-T. Pham and D.-L. Nguyen, First-principles prediction of superconductivity in MgB2C3, Phys. Rev. B 107, 134502 (2023).
  87. [88]U. Häussermann and S. Simak, Origin of the c/a variation in hexagonal close-packed divalent metals, Phys. Rev. B 64, 245114 (2001).
  88. [89]U. Wedig, H. Nuss, J. Nuss, M. Jansen, D. Andrae, B. Paulus, A. Kirfel, and W. Weyrich, Electronic origin of the structural anomalies of zinc and cadmium, Z. Anorg. Allg. Chem. 639, 2036 (2013).DOI
  89. [90]K. Takemura, The zinc story under high pressure, J. Miner. Mater. Character. Eng. 7, 354 (2019).DOI
  90. [91]A. F. Bialon, T. Hammerschmidt, R. Drautz, S. Shah, E. R. Margine, and A. N. Kolmogorov, Possible routes for synthesis of new boron-rich Fe–B and Fe1−xCrxB4 compounds, Appl. Phys. Lett. 98, 081901 (2011).DOI
  91. [92]H. Gou, N. Dubrovinskaia, E. Bykova, A. A. Tsirlin, D. Kasinathan, A. Richter, M. Merlini, M. Hanfland, A. M. Abakumov, D. Batuk, and L. Dubrovinsky, Discovery of a superhard iron tetraboride superconductor, Phys. Rev. Lett. 111, 157002 (2013).DOI
  92. [93]H. Niu, J. Wang, X.-Q. Chen, D. Li, Y. Li, P. Lazar, R. Podloucky, and A. N. Kolmogorov, Structure, bonding, and possible superhardness of CrB4, Phys. Rev. B 85, 144116 (2012).DOI
  93. [94]A. Knappschneider, C. Litterscheid, D. Dzivenko, J. A. Kurzman, R. Seshadri, N. Wagner, J. Beck, R. Riedel, and B. Albert, Possible superhardness of CrB4, Inorg. Chem. 52, 540 (2013).DOI
  94. [95]A. Knappschneider, C. Litterscheid, N. C. George, J. Brgoch, N. Wagner, J. Beck, J. A. Kurzman, R. Seshadri, and B. Albert, Peierls-distorted monoclinic MnB4 with a Mn–Mn bond, Angew. Chem. Int. Ed. 53, 1684 (2014).
  95. [96]H. Gou, A. A. Tsirlin, E. Bykova, A. M. Abakumov, G. Van Tendeloo, A. Richter, S. V. Ovsyannikov, A. V. Kurnosov, D. M. Trots, Z. Konôpková, and L. Dubrovinsky, Peierls distortion, magnetism, and high hardness of manganese tetraboride, Phys. Rev. B 83, 064108 (2014).DOI
  96. [97]Z. Xiang, Y. Zhang, Q. Lu, Q. Li, Y. Li, T. Huang, Y. Zhu, Y. Ye, J. Sun, and H. Wen, Superconductivity up to 14.2 K in MnB4 under pressure, Advanced Mate- F. E. Wang, M. A. Mitchell, R. A. Sutula, J. R. Holden, and L. H. Bennett, Crystal structure study of a new compound Li₃B₄, J. Less-Common Met. 61, 237 (1978).
  97. [99]M. Wörle and R. Nesper, Infinite, linear, unbranched borynide chains in LiBₓ—Isoelectronic to polyyne and polycum ulene, Angew. Chem. 112, 2439 (2000).
  98. [100]A. N. Kolmogorov and S. Curtarolo, Theoretical study of metal borides stability, Phys. Rev. B 74, 224507 (2006).DOI
  99. [101]A. N. Kolmogorov, R. Drautz, and D. G. Pettifor, Ab initio modeling of Li-B-H boron-chain alloys for hydrogen storage applications, Phys. Rev. B 76, 184102 (2007).DOI
  100. [102]A. Hermann, A. Suarez-Alcubilla, I. G. Gurtubay, L.-M. Yang, A. Bergara, N. W. Ashcroft, and R. Hoffmann, LiB and its boron-deficient variants under pres- sure, Phys. Rev. B 86, 144110 (2012).DOI
  101. [103]A. N. Kolmogorov, S. Hajinazar, C. Angyal, V. L. Kuznetsov, and A. P. Jephcoat, Synthesis of a predicted layered LiB via cold compression, Phys. Rev. B 92, 144110 (2015).DOI
  102. [104]G. P. Kafle, C. R. Tomasetti, I. I. Mazin, A. N. Kolmogorov, and E. R. Margine, Ab initio study of Li-Mg-B superconductors, Phys. Rev. Mater. 6, 084801 (2022).DOI
  103. [105]A. N. Kolmogorov and S. Curtarolo, Prediction of different crystal structure phases in metal borides: A lithium monoboride analog to MgB₂, Phys. Rev. B 73, 180501 (2006).DOI
  104. [106]J. Towns, T. Cockerill, M. Dahan, I. Foster, K. Gaither, A. Grimshaw, V. Hazelwood, S. Lathrop, D. Lifka, G. D. Peterson, R. Roskies, J. R. Scott, and N. Wilkins-Diehr, XSEDE: Accelerating scientific discovery, Comput. Sci. & Eng. 16, 62 (2014).DOI

Paper details

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

views · 10

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