Materials Project Battery Explorer

• Materials Project Battery Explorer is a specialized interface and data model cataloging DFT-calculated intercalation electrode properties such as capacities, voltages, and reaction energetics.
• It supports diverse workflows including composition-only screening, voltage surrogate models, and agentic design to identify promising battery materials.
• Recent expansions integrate structural generation, mobility analysis, and community contributions to address challenges like dataset imbalance and extrapolation to new chemistries.

Materials Project Battery Explorer is the battery-specific exploration layer of the Materials Project (MP) ecosystem: a web- and API-accessible view of intercalation and insertion electrode data derived from first-principles calculations, together with the broader battery endpoints that expose reaction energetics, average voltages, capacities, and structure metadata. In one recent benchmark, the term is used in a narrower but operationally important sense, referring to the public electrode database behind the legacy Materials Project Battery Explorer tool, containing 5,574 candidate electrode materials evaluated as intercalation hosts by density functional theory (DFT) (Wu et al., 8 Mar 2026). Across the MP platform, Battery Explorer has also functioned as one of the application-specific explorers layered on top of the core database and API, and its data have become a substrate for voltage surrogates, composition-only screening, multivalent-cathode discovery, and agentic design workflows (Huck et al., 2015).

1. Meaning, scope, and data lineage

Battery Explorer is best understood as both an interface and a data model. In the narrow dataset-centric usage, each entry corresponds to an intercalation host and stores the Materials Project identifier, charged and discharged compositions, the working ion, gravimetric capacity, volumetric capacity, and average intercalation voltage versus the working-ion metal (Wu et al., 8 Mar 2026). In the broader MP workflow sense, the same battery endpoints supply DFT-calculated average voltages, capacities, reaction energies, and voltage profiles for existing reaction entries, and multiple studies access those endpoints directly through pymatgen rather than through the graphical interface alone (Hossain et al., 17 Mar 2025).

Different studies have used different snapshots of the same underlying MP battery infrastructure, which makes dataset versioning central to any interpretation of reported metrics. One early extraction used 4,250 data instances for 3,580 intercalation based electrode materials and, after removing inconsistencies and repetitions, retained 3,977 instances (Joshi et al., 2019). A later voltage-prediction study extracted 4,351 battery entries from MP v2022.10.28 (Hossain et al., 17 Mar 2025). The 2026 composition-only benchmark used the 5,574-entry Battery Explorer dataset and emphasized explicitly that the benchmark question was what can be done from composition alone with the Battery Explorer data (Wu et al., 8 Mar 2026).

A common misconception is that Battery Explorer is only a lithium-ion cathode browser. The source material is broader: working ions in the 5,574-entry benchmark include Li, Mg, Na, K, Ca, Zn, Al, and others, with Li at about 43.6% and Mg at about 25.6% of the distribution (Wu et al., 8 Mar 2026). At the same time, the data remain centered on intercalation or insertion reactions rather than on the full space of battery phenomena, which is why later extensions add alloying, hybrid cathodes, multivalent mobility, and empirical cycle-life layers rather than treating the stock explorer as exhaustive.

2. Electrochemical observables and reaction representation

The Battery Explorer data model is anchored in a compact set of electrochemical observables. In the 5,574-entry dataset, the principal targets are gravimetric capacity $D_g$, volumetric capacity $D_v$, and average voltage $\bar{V}$, with capacities defined as

$$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$

where $n$ is the total number of electrons transferred per unit cell, $e$ is the elementary charge, $W_u$ is the mass of the unit cell, and $V_u$ is the volume of the unit cell (Wu et al., 8 Mar 2026). These definitions make clear that Battery Explorer is not merely cataloguing compositions: it is encoding reaction stoichiometry and intensive performance metrics derived from endpoint structures.

Average intercalation voltage follows the standard thermodynamic construction used throughout MP battery workflows. In the general intercalation setting, the average voltage is written as

$$\bar{V} = -\frac{\Delta G}{nF} \approx -\frac{\Delta E}{nF},$$

with $\Delta G$ or, in practice, the DFT total-energy difference $D_v$0, referenced to the working-ion metal (Joshi et al., 2019). For sodium layered oxides, one study writes this explicitly as

$D_v$1

which is the same average-voltage formalism specialized to the MP battery reaction representation (Hossain et al., 17 Mar 2025).

How one chooses the reaction endpoint representation matters methodologically. In the composition-only benchmark, discharged composition is always used as input because it contains all species participating in the electrochemical reaction, whereas charged compositions may omit the working ion (Wu et al., 8 Mar 2026). This distinction is consequential: it separates the endpoint-derived quantities that Battery Explorer stores from the surrogate representations used to predict them. It also clarifies why some workflows remain composition-only while others rely on full structural data or explicit charged/discharged pairs.

3. Access patterns and software ecology

Programmatic access has been as important as the web interface in defining what Battery Explorer is in practice. MP battery data are routinely retrieved through pymatgen, the Materials Project API, or MPRester, and downstream workflows use StructureMatcher, phase-diagram construction, and grand-potential formalisms to reinterpret MP entries as screening spaces for new chemistries (Biby et al., 2024). The result is an ecosystem in which Battery Explorer functions less as a static catalogue than as a reaction- and structure-indexed backend for derived workflows.

The MPContribs framework provides the architectural precedent for extending MP explorers, including battery-specific ones, with user-contributed datasets. MPContribs combines a flexible MPFile text format, RESTful API integration, and a display framework capable of rendering hierarchical metadata, tables, and interactive plots on top of MongoDB and Django; within that framework, the Lithium Battery Explorer is identified as an application with battery-specific search criteria (Huck et al., 2015). This makes Battery Explorer part of a larger pattern in MP: a core materials database overlaid with domain-specific explorers and, when needed, additional contributed or derived data layers.

Agentic workflows have begun to treat the MP battery endpoints as a tool layer inside broader reasoning systems. In ChatBattery, the Materials Project API is used for exact-formula existence checks during hypothesis generation, while ICSD similarity retrieval and MACE-MP energy evaluations handle feasibility refinement, ranking, and validation (Liu et al., 21 Jul 2025). This suggests a shift in Battery Explorer’s functional role: from a repository of known DFT reactions toward a decision substrate inside closed-loop design systems.

4. Statistical learning on Battery Explorer data

The most systematic recent benchmark treats Battery Explorer as a composition-only regression problem with three targets—$D_v$2, $D_v$3, and $D_v$4—and three model families: MODNet with 273 matminer-derived descriptors, CrabNet with mat2vec-based elemental embeddings and fractional encoding, and a random-forest regressor over 21 Magpie features (Wu et al., 8 Mar 2026). The benchmark uses unstratified 5-fold cross-validation as the main test, then adds leave-one-cluster-out (LOCO) and cluster-stratified folds, together with t-SNE and DBSCAN on MODNet features to expose latent chemistry structure.

For the full 5,574-entry dataset, CrabNet is the best-performing composition-only model across all three targets.

Target	CrabNet, unstratified 5-fold CV	CrabNet, LOCO SMAE
Gravimetric capacity $D_v$5	MAE 24.73, SMAE 0.284, $D_v$6	0.551
Volumetric capacity $D_v$7	MAE 94.31, SMAE 0.295, $D_v$8	0.530
Average voltage $D_v$9	MAE 1.087 V, SMAE 0.474, $\bar{V}$0	0.603

These results establish two points simultaneously. First, composition-only models can extract substantial signal from Battery Explorer data, especially for capacities. Second, out-of-cluster generalization is much harder than random-split interpolation: LOCO errors rise substantially, and voltage remains intrinsically harder to learn from composition alone. The same study shows that on a $\bar{V}$1-filtered subset CrabNet improves to MAE 18.13 for $\bar{V}$2, 77.81 for $\bar{V}$3, and 0.653 V for voltage, with $\bar{V}$4 values of 0.724, 0.722, and 0.660, respectively; underrepresented ions such as Al, Rb, and Cs show much larger errors than Li and Mg, and bootstrap-style subsampling shows monotonically decreasing error with increasing dataset fraction (Wu et al., 8 Mar 2026).

Voltage-specific surrogates built directly on MP battery entries push the predictive regime further. A PyTorch DNN trained on 4,351 MP battery entries with 262 descriptors—232 from XenonPy and the remainder from Matminer and Battery Explorer of MP—reports 10-fold cross-validation performance of $\bar{V}$5 and MAE $\bar{V}$6 V, while an 80/20 split yields $\bar{V}$7 and MAE$\bar{V}$8 V (Hossain et al., 17 Mar 2025). That model is explicitly positioned as a learned surrogate for the DFT voltage data already exposed through MP and is then used to screen hypothetical Na layered oxides.

An earlier generation of MP-trained voltage models used DNN, SVR, and KRR on 3,977 cleaned battery instances. On the holdout set, the reported MAEs are 0.43 V for DNN, 0.40 V for SVR, and 0.39 V for KRR, and the associated web-accessible tool is described as being able, within a minute, to estimate the voltage of any bulk electrode material for a number of metal ions; that same workflow was then used to propose nearly 5,000 candidate electrode materials for Na- and K-ion batteries (Joshi et al., 2019).

Taken together, these benchmarks show that Battery Explorer has become a canonical supervised-learning substrate for at least three distinct problem classes: composition-only capacity screening, structure- and descriptor-based voltage surrogates, and transfer screening into hypothetical or underrepresented chemistries. They also show the central limitation: interpolation on MP chemistry is much easier than extrapolation to new ion families or new structural clusters.

5. Expansion into virtual, hybrid, and generative design spaces

Battery Explorer began as an explorer of existing intercalation reactions, but recent work increasingly uses MP battery data as a springboard into spaces that the stock explorer does not enumerate directly. One route is structural generation. AIRSS-based cathode exploration rediscovers known LiCoO$\bar{V}$9 and LiFePO$$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$0 polymorphs, maps the Li–Cu–F phase space with more than 15,000 relaxed structures, and proposes a family of $$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$1 transition-metal oxalates with candidate gravimetric energy densities exceeding about 900 Wh kg$$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$2; for the best layered Fe oxalate polymorph, the reported Li migration barrier is about 0.31 eV (Lu et al., 2021). This is not Battery Explorer in the strict sense, but it supplies exactly the missing upstream ingredient: new structures and metastable polymorphs for later insertion into a Battery Explorer-like workflow.

A second route is large-scale virtual databasing against experimentally relevant targets rather than only theoretical ones. A generalized informatics framework for Li spinels generates 125,000 distinct $$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$3 compositions, predicts capacity from a graph-based manifold-learning model trained on experimental data, and reports $$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$4 for the graph-based model versus $$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$5 when experimental capacity is regressed against theoretical capacity from Materials Project (Broderick et al., 2022). This directly addresses a recurrent limitation of Battery Explorer-style screening: theoretical capacity is necessary but not sufficient as a proxy for experimentally realized performance.

A third route extends beyond intercalation to architectures that combine multiple electrochemical mechanisms. An MP-based inverse-design framework for hybrid cathode materials uses MP structures, pymatgen.StructureMatcher, convex-hull construction, and grand-potential phase diagrams to classify intercalation versus conversion phases and screen intercalation–conversion pairs. In a case study targeting gravimetric energy density above NMC333, it identifies $$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$6-$$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$7 with an average gravimetric energy density of 1,424 Wh/kg on a lithiated cathode basis, compared with NMC333’s theoretical maximum of 1,028 Wh/kg, while also emphasizing small intercalation-host volume change, sulfur immobilization, and interface stability criteria (Biby et al., 2024).

A fourth route addresses reaction classes that Battery Explorer does not natively emphasize. Alloy-anode screening built on MP and AFLOW trains CGCNN energy models and then recomputes alloying potentials and capacities across Li, Na, K, Zn, Mg, Ca, and Al systems, identifying approximately 120 low potential and high specific capacity alloy anodes (Shi et al., 2024). Here the MP role is not to provide ready-made battery entries but to supply the structural and energetic substrate from which a battery-like explorer for alloying reactions can be constructed.

Generative and agentic methods push this expansion further. ChatBattery treats MP primarily as an existence and novelty filter inside an expert-guided multi-agent loop for cathode design. Starting from NMC811, it generates 100 candidate compositions, reduces them to 89 unique compositions, ranks them through charge neutrality, preparation complexity, and voltage heuristics, and experimentally validates three final cathodes with practical capacity improvements of 28.8%, 25.2%, and 18.5% over NMC811 (Liu et al., 21 Jul 2025). Battery Explorer remains central in that workflow, but as a constraint source rather than as the entire search space.

6. Mobility, community data, and persistent limitations

One of the sharpest limits of the classic Battery Explorer paradigm is that equilibrium reaction thermodynamics do not determine whether ions can move on practical timescales. Geometry-augmented diffusion screening addresses this gap. For multivalent Mg cathodes, an automatic path finder based on Voronoi tessellations, Chebyshev centers, and dumbbell sites extracts transition-state-like geometries from relaxed supercells before any NEB calculation, and a 16-material, 40-path DFT+NEB study reports migration barriers from 0.23 eV to 1.42 eV together with a claimed speed-up of 3–4 orders of magnitude relative to brute-force NEB (Bölle et al., 2021). The methodological significance is broader than Mg: it shows how a Battery Explorer can acquire a mobility layer through geometry-derived descriptors and selective NEB refinement.

The same logic now extends to Ca batteries at scale. A 2026 workflow screens 52,945 non-Ca MP structures by calibrating a Ca-compatible Voronoi polyhedral volume window from 4,350 Ca-containing structures, then down-selects by charge neutrality, exclusion of other mobile cations, thermodynamic (meta)stability, average voltage, and Ca migration barriers predicted by MACE, Orb-v3, and a transfer-learned ALIGNN model. From that pipeline, 37 promising Ca cathode candidates emerge, with $$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$8 and $$D_g = \frac{n e}{W_u}, \qquad D_v = \frac{n e}{V_u},$$9 highlighted as low-barrier materials and $n$0, $n$1, $n$2, and $n$3 identified as candidates with thermodynamically stable charged states (Tekliye et al., 27 May 2026). This makes explicit what earlier Battery Explorer studies left implicit: for multivalent systems, geometry, voltage, stability, and migration barriers must be co-screened.

Community-data infrastructure points toward an additional expansion beyond equilibrium DFT quantities. MPContribs provides a mechanism for storing contributed datasets alongside the core MP database through MPFile, RESTful API submission, and a display framework for metadata, tables, and interactive plots (Huck et al., 2015). A plausible implication is that Battery Explorer can absorb not only reaction thermodynamics but also contributed voltage–capacity curves, cycling datasets, and spectroscopy. An external prototype of such an empirical layer already exists in the lithium-metal domain: the ABC plus MatGD workflow mines 6,444 papers, builds a database of 8,074 cells with both structured materials information and cycle-resolved performance, and trains models for initial capacity, capacity at 100/200/300 cycles, and stability classification (Lee et al., 2024).

The main limitations of Battery Explorer, as currently used in the literature, are therefore well defined rather than ambiguous. Dataset imbalance by working ion materially affects model performance; underrepresented ions such as Al, Rb, and Cs have much higher errors than Li and Mg. Voltage is harder to learn than capacity from composition alone. Random-split metrics can overstate performance relative to leave-one-cluster-out testing on new chemistry families. And many practically decisive quantities—migration barriers, interfacial compatibility, electrolyte stability windows, long-term cycle life, and synthesis constraints—are not intrinsic to the core explorer and must be layered on through additional models or data sources (Wu et al., 8 Mar 2026). The modern conception of Materials Project Battery Explorer is consequently broader than the legacy interface: it is a modular battery-discovery substrate in which MP reaction thermodynamics serve as the backbone, while mobility, virtual structure generation, empirical performance data, and agentic design are increasingly treated as first-class extensions rather than optional afterthoughts.