Discovery of Correlated Electron Molecular Orbital Materials using Graph Representations Md. Rajbanul Akhond, Alexandru B. Georgescu Abstract Correlated electron molecular orbital (CEMO) materials host emergent electronic states built from molecular orbitals localized over clusters of transition metal ions, yet have historically been discovered sporadically and generally been treated as isolated case studies. Here we establish CEMO materials as a systematically discoverable class and introduce a graph-based framework to identify, classify, and organize transition-metal cluster motifs in inorganic solids. Starting from crystal structures in the Materials Project, we construct transition metal connectivity graphs, extract cluster motifs using a bond-cutting algorithm, and determine cluster point groups, effective cluster sublattice dimensionality, and translational symmetry. Applying this approach in a high-throughput screen of 34,548 compounds yields 5,306 cluster-containing materials, including 2,627 stable or metastable compounds with isolated clusters and 984 materials featuring mixed-metal clusters. The resulting dataset reveals symmetry and element-dependent trends in cluster formation. By integrating cluster classification with flat-band lattice topology and battery-relevant information, we provide further relevant information to multiple scientific communities. The accompanying open dataset, Cluster Finder software, and interactive web platform enable systematic exploration of cluster-driven electronic phenomena and establish a general pathway for discovering correlated quantum materials and functional materials with cluster-based or extended metal-metal bonding in inorganic solids. \mciteErrorOnUnknownfalse † † email: mdakhond@iu.edu † † email: georgesc@iu.edu † † affiliation: Department of Chemistry, 800 East Kirkwood Avenue, Indiana University, Bloomington, Indiana 47405, United States 1 Introduction Correlated electron molecular orbital materials (CEMO) - or correlated electron quantum cluster materials - are solid state, inorganic materials where emergent correlated electron states are obtained from molecular orbitals extended on clusters of transition metal ions, as opposed to atomic d- and f-orbitals. These have also been refered to as ’orbital molecules’ in solids Attfield 2015 . Extended molecular orbitals in atomic clusters within solid-state inorganic materials resemble molecular orbitals in organic and inorganic chemistry, forming extended states across multiple ions. These localized molecular orbitals form flatbands, however they do not arise from hopping frustration; rather from electron localization. As a result, they tend to be in proximity to a Mott transition, and have been investigated as potential spin liquids. In molecular chemistry, molecular orbitals emerge from the overlap of atomic orbitals, creating bonding and antibonding states that govern stability, and the resulting electronic properties. In cluster materials, the symmetry and electron filling of the molecular orbital states plays a key role in cluster formation and materials’ stability, as well as a ’just right’ electron correlation strength Kumari et al. 2025 ; strong electron correlations may lead to local d-orbital states (charge order), while delocalized orbitals may lead to longer-range metal-metal bonding motifs (e.g. linear chains). As the clusters host electrons in molecular-like orbitals, they offer platforms to investigate novel quantum states and phenomena arising from electron correlations and spatial confinement. Their unusual geometry allows for a variety of applications across chemistry, physics, energy, electronics, and quantum information science. The study of cluster-based quantum effects, exemplified by 2D trimer materials like \text{Nb}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{Br}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} , reveals spin S=1/2 characteristics on \text{Nb}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}} triangles (making it a spin-liquid candidate material); 2D polar modes perpendicular to the plane have led to these materials being used as the first field-free Josephson diode for quantum computing applications. Wu et al. 2022 . \text{Mo}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} based cluster magnets Nikolaev et al. 2021 such as, \text{Li}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{InMo}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} shows a conventional magnetic ordering Gall et al. 2013 while, \text{LiZn}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{Mo}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} forms collective molecular magnets Mourigal et al. 2014 and \text{Li}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{ScMo}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} has been studied as a potential spin-liquid Haraguchi et al. 2015 , depending on the nature of interaction and disorder. Another promising material from this class is \text{LixScMo}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} - which can be doped electrochemically, and as a result, the magnetic state can be tuned. Wyckoff et al. 2023 Another family of materials with Mo clusters are the Chevrel phases, which contain Mo_{3n} clusters, of which the most well known have the general formula \text{MxMo}{\vphantom{\text{X}}}_{\smash[t]{\text{6}}}\text{X}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} , where X is a chalcogen like S or Se, and M is an intercalated metal. These phases not only show promise as electrode materials for monovalent and multivalent ion batteries Mei et al. 2017 but also as possible superconducting phases Banerjee et al. 2024 ; Marini et al. 2021 with some even showing reentrant superconductivity Gosławska and Matlak 1998 . Figure 1: Flow chart of how the Cluster Finder identifies, classifies and determines different geometrical properties of the clusters in materials. Another material family is that of lacunar spinels, which form tetrahedral tetramers. They are a family of compounds with the general formula \text{GaM}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{X}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} , where M represents transition metals and X denotes chalcogens. These materials may display relevant properties, including Jahn-Teller induced ferroelectricity Xu and Xiang 2015 , multiferroicity Singh et al. 2014 ; Ruff et al. 2015 , large negative magnetoresistance Dorolti et al. 2010 ; their molecular orbital-derived bands can be measured optically Takegami et al. 2025 . Figure 2: The ”Cluster Materials Explorer” web-app where the dataset generated by Cluster Finder can be searched by formula, materials project id, elements, cluster type, and filtered by materials and cluster properties (Top). It allows for the visualization of selected materials and show details about specific materials, including flatband and battery properties (Bottom). Figure 3: a,b,c Simple eigenbasis for a particle on a triangle (trimer) model system using s-like isomorphic orbital basis, showing the relative phases of the s-orbitals forming the molecular orbitals for the bonding A Molecular orbital, and two degenerate T Molecular orbitals. d: DFT results showing a Molecular Orbital on a Nb_{3} clusters in triangular trimer material \text{Nb}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{Cl}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} , as obtained by plotting an isosurface of a wavefunction at the \Gamma point (Nb: blue, Cl: green). e) This behavior is associated with a breathing distortion that stabilizes triangular motifs leading to an f) polar structural mode Kelly et al. 2019 . This molecular orbital (MO) picture is required to understand correlated electron materials with transition metal clusters, as exemplified in Nb 3 X 8 (X=Cl, Br, I) Kumari et al. 2025 ; Grytsiuk et al. 2024 ; Aretz et al. 2025 , Na 2 Ti 3 Cl 8 Kelly et al. 2019 , and GaV 4 S 8 Kim et al. 2020 , where it builds upon and moves beyond atomic descriptions by capturing intracluster hybridization and correlations that drive cluster stability, a novel polar mode-induced multiferroicity, and Mott gap opening respectively. For cluster stability and formation, the MO picture reveals that optimal electron filling (e.g., 6-8 electrons per trimer in Nb 3 X 8 ) driven by symmetry (occupying bonding MOs while avoiding antibonding ones), with intermediate correlation strength balancing delocalization and localization to prevent over-clustering or charge order are needed Kumari et al. 2025 . MO-driven distortions can activate polar modes Kelly et al. 2019 , leading to coupled electric polarization and magnetism, where noncentrosymmetric phases emerge without traditional pseudo-Jahn-Teller d 0 Bersuker 2013 or lone-pair mechanisms, enabling magnetoelectric effects in kagome layers (Figure 3 ). Finally, the MO picture explains Mott insulation through narrow, flat bands from cluster-localized states, where single-site methods fail but cluster MO-DMFT naturally opens gaps Aretz et al. 2025 , and leading to potential spin-liquid phases. Another similar material from these Lacunar spinel phases is \text{GaTa}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{Se}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} which may display topological superconductivity induced by pressure Park et al. 2020 . Despite their wide range of properties, these materials have generally been discovered sporadically. In this paper, we will discuss how to systematically discover, indentify and classify correlated electron molecular orbital quantum materials, and related materials with strong metal-metal bonding. The resulting dataset alongside with the ”Cluster Finder” python library and the ”Cluster Explorer” web application offer platforms to analize and explore materials displaying this type of structural and electronic pattern. In addition, we add easy-to-use tools to visualize their potential flatband topology, and battery properties as obtained from high throughput databses, as discussed below. 2 Cluster Motifs Figure 4: Different type of cluster motifs discovered by the Cluster Finder code with cluster point group classification. Attfield Attfield 2015 gave a catalog of materials which contains these extended electronic orbital states, and labeled MOs as ’orbital molecules’, while Khomskii Khomskii and Streltsov 2021 later expanded them as ’Molecules in Solids’. However, the space of materials with cluster motifs which can sustain these extended MOs or orbital molecules is vast and underexplored. In this work we use the CEMO term to avoid confusion with molecular/non-periodic cluster systems, but still reference their electronic building blocks as molecular orbitals using standard chemistry terminology. Using our Cluster Finder code we have found a diverse collection of transition metal compounds with cluster motifs, which can be categorized by size and by geometry (Table S.1 ). For example, clusters of size three (trimers) can be present in perfectly equilateral triangles ( D_{3h} ), isosceles triangles ( C_{2v} ) or linear trimers ( D_{\infty h} )(Figure 4 ). Finally, these cluster motifs are spatially distributed according to a diverse set of translational and space group symmetries. The arrangement of the clusters in the solid can reduce the effective dimensionality of these compounds (e.g., Ir trimers in \text{Ba}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{Ir}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{10}}} show effective 1D chain behavior; Nb triangular trimers in \text{Nb}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{Cl}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} are arranged in a 2D triangular lattice Figure S.2 ). To test the validity of our method, we found that it successfully identifies known quantum cluster materials including dimer material \text{K}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{Co}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{(}\text{SeO}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{)}\text{}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}} Zhong et al. 2020 , linear trimers in \text{Ba}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{Ir}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{10}}} Cao et al. 2020 , tetramers in Mott insulating \text{Al}\text{(}\text{VS}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{)}\text{}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}} Kim et al. 2020 among many others (Figure 4 ). More importantly, we have uncovered several previously unrecognized candidate materials. One notable example is \text{Lu}\text{(}\text{NbCl}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{)}\text{}{\vphantom{\text{X}}}_{\smash[t]{\text{6}}} , where we identify a \text{Nb}{\vphantom{\text{X}}}_{\smash[t]{\text{6}}} cluster with an intra-cluster Nb-Nb distance of 2.96\,\text{\AA} and a total magnetic moment of 1.0\,\mu_{\mathrm{B}} per formula unit. Cluster symmetry analysis reveals an octahedral ( O_{h} ) point group embedded within an R\bar{3}m rhombohedral lattice, forming two-dimensional layers with triangular packing (Figure S.3 ). This compound has been experimentally synthesized Ihmaïne et al. 1988 , though its cluster-based electronic properties have not yet been explored. All the cluster motifs we have discussed so far are mainly isolated. However, due to the large number of materials that may display electronic properties arising from short metal-metal bonds in a solid, we also classified other type of materials - these may not all display CEMO properties as the clusters are no longer isolated. This can be seen in Figure 5 . Beyond isolated clusters, we have identified extended cluster motifs, where the cluster motif in one cell is connected to the next cell. We typically observe this kind of cluster when one or more atoms from the cluster motif lies near the cell boundary. Another type of connectivity is shared, where as we make a supercell the cluster motif itself changes in terms of size and geometry and these generally occurs when the atoms are at the cell boundary. Figure 5: Different types of cluster motif sharing. An example cluster of size four with T_{d} cluster point group symmetry is shown, where \text{Al}\text{(}\text{MoS}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{)}\text{}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}} contains isolated Mo cluster motifs, \text{Al}\text{(}\text{CuO}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{)}\text{}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}} contains Cu cluster motifs extending in one direction through bond sharing and \text{TiNi}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{GeO}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} contains coalescence of Ni cluster motifs by corner (atom) sharing. To make the materials we have identified available and easy to visualize, as well as their cluster motifs, we have built an online interface, ”Cluster Materials Explorer” (Figure 2 ). This includes information on symmetry-allowed physical properties, including the cluster point group, metal-metal bond lengths, dimensionality, as well as enantiomorphism, optical, polarity, and piezoelectricity (Figure S.4 ). We have added possible flatband and battery information according to the crystal net catalog Neves et al. 2024 and Materials Project battery dataset Jain et al. 2013 , respectively. We mention that the flatbands in the crystal net catalog use a different methodology and focuses on flatbands arising from frustration, rather than cluster molecular orbital localization. All the materials can be viewed in an interactive way in the interface and each material is linked to its Materials Project (MP) and/or ICSD structure if available. The methodology used to identify, analyze and classify these materials will be discussed below. 3 High-throughput Summary To identify if a material contains clusters, the ”Cluster Finder” code first retrieves the crystal structure from Materials Project, builds a connectivity matrix for the TM sites and builds a graph from it. Graph representations not only captures this connectivity and provide appropiate features to model this kind of system, but also provide an opportunity to connect with other material properties (e.g. sublattice based flat bands, ion migration paths in battery materials) which can be represented in the same way (Figure S.5 ). From the graph, using distance criteria (for example, the distances between the TM sites have to be <3\,\text{\AA} ) and splitting algorithms, we identify unique clusters. To classify the clusters, we determine their point group. Then, to understand how the clusters relate to each other through translations, we determine a centroid position for each unique cluster. Using the centroid positions as lattice points, we build a sub-lattice and determine its space group. Moreover, we determine the dimensionality of the spatial distribution of the centroids using Principal Component Analysis (Figure 1 ). More detail on this process will be discussed in the method section. A high-throughput search was performed using our custom ”Cluster Finder” code on a subset of Materials Project, on materials containing not more than four unique elements, with a total magnetization in the range of 0.01-10 \mu B . The total number of analyzed compounds were 34,548. Among them, the total number of compounds with possible clusters (after removing missing data and duplicates) was 5,306. For aditional analysis we further downselect materials based on the stability criterion that the material is within 0.1eV/atom from the convex Hull energy (to account for errors in DFT and the possibility of synthesizing metastable materials). The total count of stable/metastable compounds with only isolated cluster were 2,627. We also found a total of 984 mixed element clusters. Additionally, the total number of battery materials containing clusters was 1,590 (470 compounds with battery data from materials project). Among the stable isolated clusters, the Co column of TMs were found to form more clusters, with Co forming the most. A deviation from this trend can be seen for the element Hf. However, this may be less of a reflection of the tendency of the elements themselves, and more of a reflection of the compounds on materials project. Most transition metals show a strong preference for D_{\infty h} symmetry (linear with inversion center), with decreasing numbers in lower symmetries including C_{s} (mirror plane only), C_{2v} (two-fold rotation) and C_{\infty v} (linear without inversion, common for mixed element linear clusters). Co not only shows a higher likelihood to form clusters but it also shows the highest variety of clusters in terms of identified cluster point groups. Mo shows the highest count in D_{3d} , slightly higher than D_{\infty h} . Mo is another ion that displays a wide range of clusters, which can be attributed to the high number of Mo_{3n} structures in Chevrel phases. These phases are built around Mo_{3n} , ranging from Mo_{3} to much larger units like Mo_{12} or Mo_{18} in extended or shared frameworks. These clusters exhibit D_{3d} point group symmetry due to a rhombohedral distortion of the ideal octahedron (which would be O_{h} ). Moreover, highly symmetric point groups such as O_{h} and T_{d} can be seen for some specific TMs: Nb, Mo and Co, Fe respectively (Figure 11 c). 4 Mixed-Metal Clusters The cluster finder identified a large subset of materials with mixed-metal clusters - a particularly understudied direction. We can also think of these mixed elements clusters in terms of doping e.g. \text{Zr}{\vphantom{\text{X}}}_{\smash[t]{\text{6}}}\text{CoI}{\vphantom{\text{X}}}_{\smash[t]{\text{14}}} can be thought as if a Zr octahedra has been doped by Co at the cube center (Figure 6 a). This doping strategy has been used in \text{Ba}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{RuMn}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{10}}} , where by doping the trimers the material has led to a polar material, resulting in second harmonic generation Skaggs et al. 2024 . We also found materials with extended metal-metal bonding, however with mixed metal types, e.g. \text{Ca}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{CoRhO}{\vphantom{\text{X}}}_{\smash[t]{\text{6}}} (Figure 6 b) which may hold quasi-one-dimensional Ising spin chains built from strongly coupled Rh-Co-Rh trimers of face-sharing octahedral RhO 6 and trigonal-prismatic CoO 6 polyhedra. Moreover, previous first-principles calculations and neutron diffraction establish that this chain geometry promotes substantial metal-metal overlap and a robust, highly anisotropic ferromagnetic coupling within the chains, while antiferromagnetic coupling between the chains arranged on a triangular lattice drives a partially disordered antiferromagnetic state below T_{1}\approx 90 K with an additional low-temperature freezing near T_{2}\approx 30 K Paulose et al. 2008 ; Eyert et al. 2008 ; Hardy et al. 2003 . While this is not a correlated electron molecular orbital material, it does highlight the relevance of the other types of correlated electron materials we’ve identified via short metal-metal bonds. Figure 6: Examples of mixed clusters, (a) \text{Zr}{\vphantom{\text{X}}}_{\smash[t]{\text{6}}}\text{CoI}{\vphantom{\text{X}}}_{\smash[t]{\text{14}}} showing hexamer cluster of Zr with a Co inside the polyhedra and (b) \text{Ca}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{CoRhO}{\vphantom{\text{X}}}_{\smash[t]{\text{6}}} showing a spin chain with a Rh-Co-Rh building block. 5 Clusters and Flatband Materials There is significant overlap between CEMO materials, and materials that provide appropiate crystal structure motifs for flat band formation arising from band topology. We have added flat band information to the cluster materials dataset, indentified according to the crystal net tight binding (TB) method Neves et al. 2024 . This method identifies non-trivial flat bands in materials via a high-throughput workflow: extracting elemental sublattices from the Materials Project database, constructing nearest-neighbor tight-binding models with Hamiltonian H=-t\sum_{\langle i,j\rangle}(c_{i}^{\dagger}c_{j}+c_{j}^{\dagger}c_{i}) , eliminating trivial isolated bands, computing bandstructures on a k-point grid to detect flat bands with dispersion \leq 0.1\% , classifying lattices topologically, and filtering for robust candidates based on structural criteria, including atom count and bond distances, yielding thousands of unique flat band nets. Some of the possible stable cluster materials flagged as flatband materials according to the method are shown in Table 1 , showing a wide varity of cluster materials containing different number and type of flat band(s). Among them we have calculated tight binding band structures using the appropiate TM sublattice for dimer Sc 5 NCl 8 , trimer Co 3 (SnS) 2 and tetramer V 4 GaS 8 (Figure 7 ). From the TB band structure for Sc 5 NCl 8 , we can clearly see two distinct flat bands, one at the fermi level and another above it (Figure 7 a). DFT with spin-orbit coupling (SOC) flags it as a promising, but so far largely unexplored, platform with quasi-2D flat bands hosting weakly dispersive bands in proximity to the Fermi level, rooted in its layered Sc–N/Cl framework and possible nontrivial topology. Regnault et al. 2022 In contrast, Co 3 (SnS) 2 , a kagome ferromagnet and Weyl semimetal, exhibits an experimentally established, spin-orbit entangled flat band intersecting the Fermi level, as revealed by both our trivial TB band structure (Figure 7 b) and spectroscopic probes along with many-body analyses; this band is intimately linked to its large anomalous Hall response and unconventional orbital magnetism, thereby providing a paradigmatic realization of flat-band-enhanced topological transport in a correlated metal. Xu et al. 2020 ; Yin et al. 2019 . Finally, in V 4 GaS 8 , lacunar spinel chemistry stabilizes V 4 -cluster molecular orbitals that form an ultra-narrow, effectively flat half-filled band, where reduced kinetic energy, Jahn-Teller distortions and strong on-site interactions cooperatively drive a molecular Mott insulating ground state intertwined with multiferroicity, These examples highlights cluster-based flat-band architectures as a robust design principle for strongly correlated and multifunctional materials. Table 1: Top stable cluster materials with possible flatbands according to crytal net tight-binding model with relevant cluster and flat band properties. FB sub-lattice indicates transition metal sub-lattice responsible for the flat bands; No. FBs shows number of possible flat bands and FB lattice dimensionality indicates the dimensionality of the cluster sub-lattice. Formula Cluster sizes Cluster space group Cluster point groups Cluster dimensionality Min. avg. distance FB sublattice No. FBs FB lattice dimensionality Co 5 SbO 8 3 R \bar{3} m D_{3h} 2 2.94 Co 1 2 Co 3 (SnS) 2 3 R \bar{3} m D_{3h} 2 2.68 Co 1 2 Cu 3 H 2 (OF 2 ) 2 2 P2/m D_{\infty h} 3 2.98 Cu 2 1, 1 Ce 2 B 2 Ir 5 7, 3 Cm D_{3d} , D_{3h} 2 2.73 Ir 3 3 Li 7 (Mo 3 Se 4 ) 4 6, 6 Cm D_{2h} , O_{h} 3 2.66 Mo 4 1, 2 Ba 2 Nb 15 O 32 6 R \bar{3} m D_{3d} 2 2.84 Nb 1 2 Sn 2 Rh 3 S 2 3 R \bar{3} m D_{3h} 2 2.85 Rh 1 2 Sc 7 NiBr 12 4 R \bar{3} m C_{3v} 3 2.44 Sc 4 1 Ta 3 Al 2 CoC 2, 2, 2 R3m C_{\infty v} 2 2.98 Co 2 3 Sm 3 Ti 3 (SeO 4 ) 2 2, 2 P2 1 /m D_{\infty h} 2 2.84 Ti 2 1 Li 2 V 3 SnO 8 3, 3 Cmcm C_{2v} 2 2.92 V 2 2, 2 V 4 GaS 8 4 Fm \bar{3} m T_{d} 3 2.92 V 2 3 Zr 9 S 2 2, 2 I4/mmm D_{\infty h} 3 2.91 Zr 4 3 Figure 7: Tight-binding band structure and density of states (DOS) according to crystal net model for different types of cluster materials, (a) Sc dimer in \text{Sc}{\vphantom{\text{X}}}_{\smash[t]{\text{5}}}\text{NCl}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} showing two flatbands, one at the Fermi level and the other above it, (b) \text{Co}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{(}\text{SnS}\text{)}\text{}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}} with a Co-trimer showing one flat band above the fermi level and the same can be seen for the V-tetrameter in \text{V}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}}\text{GaS}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} . 6 Clusters in Battery Materials To indentify battery materials within our dataset we searched for common conducting ions in cathode materials e.g. Li, Na, K, Mg, Zn, Ca, Al and also added battery cathode information from MP battery explorer for the resulting 470 available compounds. Some of the example cluster battery materials are shown in Table 2 . Traditional battery materials suffer limitations as a result of volume changes Hu et al. 2025 ; Sun et al. 2024 and Jahn-Teller distortions Li et al. 2016 ; Ning et al. 2017 ; Ning et al. 2017 and non-optimal ion-migration pathways Wang et al. 2025 . The varied geometries of transition metal cluster materials, and their different electron-crystal structure interplay may lead to more effective battery materials. Separately, electrochemical doping may allow for the control of the correlated electron properties of this class of materials, as for example, in Mo 3 -based kagome networks, including \text{LixScMo}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{O}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} , where Li doping controls the spin state of the material Wyckoff et al. 2023 . Our work and methodology led to the identificatin of a wide variety of materials with potential battery applications. Table 2: Top stable cluster materials with relevant battery properties. Formula Cluster sizes Cluster space group Cluster point groups Cluster dimensionality Min. avg. distance Battery formula Average voltage (V) Gravimetric capacity (mAh/g) Ba 2 Na(CuO 2 ) 3 3, 3 Fmmm C_{2v} 2 2.55 Na 0-1 Ba 2 (CuO 2 ) 3 0.218 45.87 Li 7 (Mo 3 Se 4 ) 4 6, 6 Cm D_{2h} , O_{h} 3 2.66 Li 0-1.75 Mo 3 Se 4 2.047 76.16 Al(VS 2 ) 4 4 Fm \bar{3} m T_{d} 3 2.90 Al 0-0.25 VS 2 0.133 165.01 Li 4 MnNi 3 O 8 3 R \bar{3} m D_{3h} 2 2.92 Li 2-4 MnNi 3 O 8 3.333 138.59 Ba 2 Li(CuO 2 ) 3 6 Fmmm D_{2h} 2 2.65 Li 0-1 Ba 2 (CuO 2 ) 3 2.390 47.17 Li 4 Co 3 NiO 8 3 R \bar{3} m D_{3h} 2 2.86 Li 0-4 Co 3 NiO 8 3.335 274.01 Li 8 Ti 15 O 32 3 R \bar{3} m D_{3h} 0 2.93 Li 4-8 Ti 15 O 32 2.207 83.40 NaNiO 2 2, 2 P2/m D_{\infty h} 2 2.85 Na 0-1 NiO 2 1.073 235.76 Ba 8 LiNb 7 O 24 2 P6/mmm D_{\infty h} 2 2.60 Li 0-1 Ba 8 Nb 7 O 24 1.450 12.52 Ca 3 CoRhO 6 3 Pmmm D_{\infty h} 2 2.74 Ca 1.5-3 CoRhO 6 1.045 212.67 Li 4 Cr 3 CoO 8 3 R \bar{3} m D_{3h} 2 2.96 Li 2-4 Cr 3 CoO 8 3.088 144.61 Zn 4 Cu 5 (TeO 6 ) 3 4 C2/m C_{2h} 3 2.88 Zn 1.5-4 Cu 5 (TeO 6 ) 3 -1.841 107.19 7 Methods 7.1 Cluster Identification Transition Metal Cluster Analysis Algorithm Step 1: Create Connectivity Matrix Step 2: Graph Representation Step 3: Cluster Identification Step 4: Analyze and Split Clusters Connectivity Matrix C : C_{ij}=\begin{cases}1&\text{if }d(s_{i},s_{j})<\text{cutoff}\\ 0&\text{otherwise}\end{cases} Graph G=(V,E) : V=\{0,1,...,n-1\} , E=\{(i,j)|C_{ij}=1\} Clusters: Connected components in G where size \geq CLUSTER SIZE Avg Distance: \bar{d}=\frac{1}{|E|}\sum_{(i,j)\in E}d(s_{i},s_{j}) Split if: size > CLUSTER SIZE+1 and \bar{d}_{\text{parent}}-\bar{d}_{\text{sub}}>\text{threshold} Split Algorithm Centroid Calculation: 1. \vec{f}_{i}=\text{lattice}^{-1}\cdot\vec{r}_{i} 2. \vec{f}_{\text{avg}}=\frac{1}{n}\sum_{i=1}^{n}\vec{f}_{i} 3. \vec{c}=\text{lattice}\cdot\vec{f}_{\text{avg}} Figure 8: Mathematical flow and logic of transition metal cluster identification and analysis algorithm To identify potential CEMO materials we have developed a custom algorithm as summarized in the diagram (Figure 8 ). First, we begin by selecting compounds based on a predefined chemical composition criterion - for example, in the case of our high throughput search, compounds containing transition metals (TMs) such as Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Y, Zr, Nb, Mo, Tc, Ru, Rh, Hf, Ta, W, Ir and anions such as B, C, N, O, P, S, Cl, Se, Br, Te, I with at most four distinct types of atoms with a total magnetization in the range of 0.01-10 \mu B . We then apply intial distance cutoff of 3\,\text{\AA} between TM sites to construct the connectivity matrix. Utilizing this matrix we build the initial parent graph where the TM sites form the nodes and distances between them form the edges. Using this initial graph we get the initial cluster guess and calculate the inital average distance at this point. Then, utilizing a moving cutoff scheme we cut the longest egde and calculate the average distances of the new child graphs (sub-clusters). If the new average metal-metal distance is lower than the initial (parent) average distance then we keep the split otherwise we keep the initial cluster (Figure 9 ). The splitting process stops when the average bond length stabilizes, indicating the resulting graph accurately represents a physically meaningful cluster without unnecessarily fragmenting chemically coherent groups. Finally, we calculate the center of point mass (centroid) of the clusters and use it to make sure the splitted clusters are well separated. The smallest cluster size is two and biggest is eight. Other properties and parameters of the cluster indentification process can be found in the configuration file used in our code. To indentify cluster connectivity we again utilize the average cluster distance, however, in terms of extending the unit cell in this case. So, we build a 2*2*2 supercell and compare how the cluster average distance chages compared to unit cell. If supercell has larger cluster with same average distance then it is flagged as extended cluster and if the average distance is different then it is flagged as shared cluster, otherwise the cluster is flagged as isolated and we go through the remaining steps as before. Following the initial identification of promising candidates based on these signatures, clusters within the identified materials are further characterized and their symmetry and lattice properties are analyzed in detail. Stage 1: Initial Cluster Nb Nb Nb Nb Nb Nb Nb 2.8 2.7 2.7 2.8 3.8 2.9 2.8 2.8 Avg. Nb-Nb bond: 3.3 Å Stage 2: Cut Longest Bonds Nb Nb Nb Nb Nb Nb Nb 2.8 2.7 2.7 2.8 3.8 2.9 2.8 2.8 Cut Cluster 1 (2.75 Å) Cluster 2 (2.83 Å) Stage 3: Subclusters after splitting Nb Nb Nb Nb 2.8 2.7 2.7 2.8 Nb Nb Nb 2.9 2.8 2.8 Subcluster 1 Subcluster 2 Figure 9: Visualization of the cluster splitting process, as exemplified on Nb ions. The algorithm starts with a cluster containing longer bonds (red dashed lines), progressively cuts them, identifying them as inter-cluster bonds, until the average bond lengths for the graphs stabilizes, identifying two clusters with short average Nb-Nb bond lengths. 7.2 Symmetry and Dimensionality Analysis Once TM clusters are identified, their internal symmetry is determined by calculating their point group symmetry centered on the centroid of the cluster pseudo-molecule. This local symmetry feature can capture potential couplings between electronic, magnetic, and structural degrees of freedom. After that, we utilize the centroid points to construct a sublattice and determine the space group of that sublattice which gives the translational relationship between the clusters in the 3D space. To build a proper sublattice, we label each centroid points in terms of its uniqueness. We utilize cluster properties e.g. element type, cluster size, point group, average distance and orientation (calculated using moment of inertia tensor) to determine the uniqueness. If any of these properties for two clusters are different we give them different/unique labels otherwise, we give them the same label. The extended sublattice structure formed by these clusters is further characterized by assigning each cluster unique spatial coordinates and using Principal Component Analysis (PCA) to determine the effective dimensionality—differentiating among three-dimensional frameworks, two-dimensional layers, one-dimensional chains, and zero-dimensional molecular like like solids. Further discussion on the PCA method can be found in the supporting information. coords n\times 3 \begin{bmatrix}x_{1}&y_{1}&z_{1}\\ \vdots&\vdots&\vdots\\ x_{n}&y_{n}&z_{n}\end{bmatrix} centered n\times 3 \begin{bmatrix}x_{1}-\mu_{x}&y_{1}-\mu_{y}&z_{1}-\mu_{z}\\ \vdots&\vdots&\vdots\\ x_{n}-\mu_{x}&y_{n}-\mu_{y}&z_{n}-\mu_{z}\end{bmatrix} SVD U:n\times 3 s:[s_{0},s_{1},s_{2}] V^{T}:3\times 3 s_normalized [1,\,s_{1}/s_{0},\,s_{2}/s_{0}] Classify “0D”, “1D”, “2D” or “3D” Subtract \mu SVD: U\,\mathrm{diag}(s)\,V^{T} Normalize s Thresholds Figure 10: SVD-based dimensionality classification process For our dimentional analysis (Figure 10 ), first of all, we extract 3D coordinates (x,y,z) of n centroid points from a large supercell constructed from our conventional cluster sublattice unit cell. So, we have n\times 3 , where n is the number of sites (centroids), and 3 corresponds to the spatial dimensions which can be representated as, \text{coords}=\begin{bmatrix}x_{1}&y_{1}&z_{1}\\ x_{2}&y_{2}&z_{2}\\ \vdots&\vdots&\vdots\\ x_{n}&y_{n}&z_{n}\end{bmatrix} (1) Now, we compute the mean of each column (\mu_{x},\mu_{y},\mu_{z}) and subtracts it from each coordinate to center the data around the origin, a prerequisite for PCA. \mu=\left[\frac{\sum x_{i}}{n},\frac{\sum y_{i}}{n},\frac{\sum z_{i}}{n}\right],\quad\text{size }1\times 3. Through broadcasting we get, \text{centered}=\text{coords}-\frac{1}{n}\mathbf{1}_{n}\mu, where \mathbf{1}_{n} is an n\times 1 vector of ones. Now, we perform SVD as, \text{centered}=U\Sigma V^{T} Here, V ( 3\times 3 ) (columns of V^{T} transposed) gives the principal directions in 3D space and \Sigma (size 3 vector) is eigenvalues for those directions. Then, we normalize the Singular Values as, \text{s\_normalized}=\left[\frac{\sigma_{0}}{\sigma_{\max}},\frac{\sigma_{1}}{\sigma_{\max}},\frac{\sigma_{2}}{\sigma_{\max}}\right],\quad\text{where }\sigma_{\max}=\sigma_{0}. Thus, \text{s\_normalized}=\left[1,\frac{\sigma_{1}}{\sigma_{0}},\frac{\sigma_{2}}{\sigma_{0}}\right],\quad\text{for }n\geq 3. This normalization facilitates comparison using thresholds, focusing on relative significance. The dimensionality is determined using the following criteria: Dimensionality Condition 0D s[1]<0.2 and s[2]<0.2 1D 1-s[1]\leq 0.02 and s[2]<0.5 2D \frac{s[2]}{s[1]}<0.8 3D \frac{s[2]}{s[1]}\geq 0.8 Table 3: Conditions for Dimensionality Classification Based on Normalized Singular Values Figure 11: Summary of the results of our high-throughput search using the Cluster Finder, (a) normalized likelihood of forming clusters by different transition metals in stable materials contained within Materials Project, (b) stability distribution of all the identified cluster materials in terms of energy above the hull with dark green indicating stable, light green metastable, yellow slightly unstable and red unstable cluster materials and (c) correlation heatmap showing the normalized likelihood of forming stable clusters with specific point groups by a transition metal. 8 Discussion Our work serves as a framework to classify and identify materials displaying metal clusters and extended metal-metal bonding in solids, identifying potential correlated electron molecular orbital (CEMO) materials, identifying their symmetry-related properties. We also connect these materials to other tools to analyze their potential flatband and battery properties. We have provided our database in an easy to use, interactive website. Our work can serve as a building block for the systematic study and discovery of this wide class of materials. 9 Code and Data availability The code to indentify, classify and analyze clusters is available at Cluster Finder GitHub repo with full CLI support. The dataset containing flat band and battery materials can be found here: 10.24435/materialscloud:76-n2 . To explore and download the full dataset visit: Cluster Explorer . 10