GNoME: Scaling Materials Discovery from Ten-Thousand-Scale Enumeration to Million-Scale Convex Hulls
New1. Overview
GNoME is a materials discovery system from Google DeepMind, published in November 2023 as "Scaling deep learning for materials discovery" in Nature, volume 624, pages 80-85. The problem it targets is concrete: the space of candidate crystals is far too large to enumerate, and deciding whether a structure sits on the energy convex hull requires first-principles calculation. Bracketed by these two costs, materials discovery had been confined to tens of thousands of entries.
The paper's answer is to place a graph neural network in the middle as a filter, forming a closed loop. Candidates come from two generation pipelines, the model predicts energy and stability, only structures that pass a prediction threshold go to DFT, and the DFT results flow back into the training set for the next model round. The paper describes this as a data flywheel in which model-driven filtration and DFT jointly improve predictions. Each turn of the flywheel widens the searchable chemical space.
Three numbers in the results are often conflated and need to be kept apart. The candidate pool exceeds 10^9 structures across the active learning campaign. The models identify 2.2 million structures below the current convex hull. The updated convex hull adds 381,000 new entries, bringing the total of stable crystals to 421,000 against a prior baseline of 48,000. These correspond to how many were searched, how many were hit, and how many were retained.
Two independent validation lines exist. Against experiment, 736 structures from the aggregated ICSD experimental entries match structures obtained independently through GNoME, with six examples shown in the paper. Against a higher-fidelity functional, a recomputation with r2SCAN finds that 84 percent of discovered binary and ternary crystals retain negative phase-separation energies.
The accompanying artifacts are a substantive part of the work. The official repository google-deepmind/materials_discovery was published on GitHub on the same day as the paper and provides a stable-materials summary CSV, compressed CIF directories organized by id, composition, and reduced formula, r2SCAN recomputation results, model definitions for the static graph network and the NequIP potential, and several Colab notebooks. Data is hosted in the public bucket gs://gdm_materials_discovery and mirrored through a Materials Project page under permanent link 10.17188/2009989.
For simulation engineers the significance goes beyond the count of new materials. GNoME pushes the pattern of using learned models to compress the search space and first-principles calculation to arbitrate to a scale not seen before, and it opens a second line of work: a general-purpose machine-learning interatomic potential trained on relaxation trajectories and used for zero-shot ionic conductivity estimates. That line runs parallel to the foundation potential route and to the general model lineage catalogued in this site's physics foundation models collection.
2. How It Works
The technical skeleton has three parts: candidate generation, model filtering with DFT verification, and the active learning feedback loop.
Candidate generation follows two complementary pipelines. The structural pipeline starts from known crystals and applies substitution, but the authors deliberately avoid simple complete replacement. They adjust ionic substitution probabilities to favor exploration and introduce symmetry-aware partial substitutions, a framework that supports incomplete replacement and so avoids the combinatorial explosion of full substitution. The paper uses substitution in double perovskites of composition A2B2X6 as the motivating case, where strict chemical constraints exclude partially occupied structures. Across the active learning campaign the structural pipeline produces more than 10^9 candidates. The compositional pipeline ignores structure entirely and takes reduced chemical formulas, applying relaxed oxidation-state balancing constraints for initial filtering since strict balancing would miss compositions such as Li15Si4, then initializing 100 random structures per surviving composition for ab initio random structure search.
Filtering is performed by two models. The structural model is a static graph neural network performing message passing on atomic graphs. The authors note that aggregate projections are shallow multilayer perceptrons with swish nonlinearities, and that for structural models it is important to normalize messages from edges to nodes by the average adjacency of atoms across the entire dataset. The compositional model takes reduced formulas as input. Both models output an energy and a threshold chosen from relative stability via decomposition energy, and structures passing the filter proceed downstream.
Uncertainty quantification uses deep ensembles together with volume-based test-time augmentation to filter the billion-scale candidate pool. Structures are then clustered, polymorphs are ranked, and only these are evaluated with DFT.
DFT evaluation uses VASP with the PBE functional and PAW potentials. The authors state the settings are consistent with Materials Project workflows as encoded in pymatgen and atomate, including a Hubbard U parameter applied to a subset of transition metals, a 520 eV plane-wave cutoff, and magnetization settings. This step serves two roles at once: it arbitrates model predictions and it supplies training data.
The active learning flywheel is therefore self-consistent. The more valuable the filtered structures, the more accurate the DFT results, the faster the training set grows, and the better the next round of filtering. The paper reports that model accuracy improves as a power law with data volume, consistent with scaling behavior observed in other deep learning domains.
The accuracy trajectory is verifiable. Initial models were trained on a 2018 Materials Project snapshot of roughly 69,000 materials, where the improved networks reached a mean absolute error of 21 meV per atom against a prior benchmark of 28 meV per atom on the same task. At full scale, energy prediction reached 11 meV per atom, the structural pipeline pushed the hit rate for stable predictions above 80 percent, and the compositional pipeline alone reached 33 hits per 100 trials against 1 percent in previous work.
One model capability deserves attention. The paper reports accurate prediction for structures with five or more unique elements even though such compositions were omitted from training, which the authors describe as emergent out-of-distribution generalization and which bears directly on exploring high-order chemical space.
On the relaxation trajectories, the paper additionally trains a general-purpose machine-learning interatomic potential. The data come from first-principles energies and forces computed during ionic relaxation, with unprecedented diversity and scale. The architecture is NequIP, implemented in JAX. Reported training details include a 94-element one-hot embedding, self-connections, division by 26 after each convolution for internal normalization, the Adam optimizer at a learning rate of 2e-3, and a batch size of 32, where each batch draws 16 structures from all frames of all relaxations and 16 from the first relaxation step only. The authors report that oversampling first-step structures substantially improved downstream performance. Robustness experiments cover As24Ca24Li24, Ba8Li16Se32Si8, K24Li16P24Sn8, and Li32S24Si4, with models trained on increasing training set sizes from 400 K ab initio molecular dynamics trajectories and evaluated at both 400 K and 1000 K.
3. Scope and Assumptions
GNoME conclusions rest on explicit premises, and reading the numbers without them leads to wrong conclusions.
First, the stability criterion is a PBE-level decomposition energy. Recomputation with r2SCAN shows 84 percent of discovered binary and ternary crystals retain negative phase separation, which is evidence of robustness rather than a guarantee. The dataset documentation states plainly that stability metrics change with the choice of functional, so not all released materials remain stable by that metric. Community work on universal interatomic potential benchmarks and on chemistry-resolved disagreement between Materials Project-compatible PBE(+U) and r2SCAN verdicts repeatedly raises the same point. The 736 experimental matches concern structure identity rather than a full experimental recheck of every stability verdict.
Second, the convex hull is a snapshot. The documentation states that the provided materials update the convex hull of a snapshotted version of Materials Project and similar databases, so the set may not remain stable as more crystals are discovered. The dataset uses a convex-hull threshold of 5e-5 eV, while the August 2024 revision widened the release to materials within 1 meV per atom of the hull, raising the count from 381,000 to more than 520,000.
Third, energy conventions must be aligned manually. Reconstructing the complete convex hull needs energies from Materials Project, OQMD, and WBM, and the pseudopotential conventions differ across those archives. The dataset supplies per-atom pseudopotential corrections for Ga, Ge, Li, Mg, and Na to bring Materials Project entries onto the GNoME convention. The paper notes that GNoME displaces at least 5,000 stable materials from Materials Project and OQMD. Mixing energies from different sources without correction introduces systematic bias.
Fourth, synthesizability is not a model criterion. GNoME optimizes thermodynamic stability along a single dimension. Kinetic accessibility, synthetic route economics, element supply, and supply-chain constraints are absent from the model output. The paper itself frames the final-hull entries as a candidate pool for further technological screening rather than a validated synthesis-ready list.
Fifth, the potential branch targets bulk solids. Its robustness experiments cover a limited set of ionic compounds with an evaluation ceiling of 1000 K, and the demonstrated application is a zero-shot ionic conductivity estimate. Extrapolation to surfaces, defects, melts, or complex interfaces is outside the validated range.
Sixth, data licensing constrains use in practice. The dataset is released under CC BY-NC 4.0, meaning non-commercial use only, while code and notebooks are Apache-2.0. The repository disclaimer also states the data is for theoretical modeling only and is not intended for or approved for any medical or clinical use.
4. Getting Started
The most efficient entry point is the official dataset descriptor DATASET.md, which documents downloading, fields, convex-hull reconstruction, and version history in more practical detail than the paper. Two download routes are provided. The command line uses gcloud storage cp --recursive gs://gdm_materials_discovery/ data/, and the Python side supplies scripts/download_data_wget.py and scripts/download_data_cloud.py, with the latter preferable when the Google Cloud CLI is already authorized. Both accept a --data_dir flag and default to a data directory in the current working directory.
The bucket layout is regular. Under gnome_data there are stable_materials_summary.csv, holding compositions and DFT raw energies along with formation and decomposition energies; stable_materials_r2scan.csv with the r2SCAN recomputation results; and three compressed structure directories by_composition.zip, by_id.zip, and by_reduced_formula.zip. The auxiliary_gnome_data folder holds supporting data for a2c structure prediction, and external_data holds the Materials Project snapshot summary and the external materials summary.
Because of the convex-hull organization, only one structure per composition is needed, so the recommended workflow is to locate the composition in the summary CSV first and then extract the matching structure from the compressed archive. The visualization notebook demonstrates extracting individual files from the compressed ZIP rather than unpacking everything.
The field design is worth learning before writing scripts. Beyond composition, MaterialId, reduced formula, and element set, the summary provides site count, volume, density, point group, space group and space group number, crystal system, uncorrected energy, MP2020-corrected energy, formation energy per atom, decomposition energy per atom, several variants of the decomposition energy including relative to the downloaded Materials Project hull, to all entries, to all entries except the current one, to Materials Project only, and to Materials Project plus OQMD including recomputations. There is also Dimensionality Cheon, a dimensionality predicted by the 2017 method of Cheon et al., Bandgap where available, and Is Train indicating membership in the associated machine learning training set. Formation energies are computed against a fixed set of elemental references from Materials Project and are not lowered if a lower-energy elemental structure was found; the references are listed separately in elemental_references.json.
Rebuilding the complete convex hull requires merging in Materials Project, OQMD, and WBM energies alongside the GNoME summary. The dataset provides external_materials_summary.csv with convex-hull entries and energies for element compositions matching those external archives, which simplifies the join.
The model directory provides definitions for the two model families used in the paper. GNoME is the static graph neural network that was the predominant model behind new materials discovery, a simple message-passing architecture optimized by training on a 2018 Materials Project snapshot and reaching 21 meV per atom. NequIP corresponds to the architecture created by Batzner et al. in 2022 and was used to train the interatomic potentials described in the paper, provided as a JAX implementation with basic configuration parameters. The README lists reference structures and search paths, model training notebooks and configs, and additional properties such as band structure as upcoming items that are not yet complete.
For a reader who only needs to judge the stability of one system, the minimal path is to download stable_materials_summary.csv, filter candidates with the decomposition energy fields, and pull the structure by id from by_id.zip for further first-principles or machine-learning potential calculation.
5. Comparison with Alternatives
Placed alongside related work, the most direct references are the same-issue Nature paper and the later model lineage.
The companion article in the same Nature issue, "An autonomous laboratory for the accelerated synthesis of inorganic materials" at volume 624, pages 86-91, represents a different route: it moves the verification end of materials discovery from computation into the laboratory and actually synthesizes candidates. The two are complementary, since GNoME widens the search space while the autonomous laboratory carries out the empirical test. The difference shows most clearly in failure modes, because a GNoME failure is a stable structure that cannot be made, while an autonomous laboratory failure is a correct structure outside the synthesis window.
Methodologically, the GNoME static graph neural network belongs to the CGCNN lineage. The crystal graph convolutional neural network proposed by Xie and Grossman in 2018 is the origin of that lineage, and the paper anchors the 28 meV per atom prior benchmark to it. GNoME's contribution is pushing such structural models onto a billion-scale candidate pool and using active learning with scaling laws to reach 11 meV per atom.
Against the foundation potential direction, the GNoME route to potentials is a byproduct of relaxation trajectories rather than a program of pretraining on general chemical space followed by fine-tuning. The GNoME potential covers bulk solids and sits within the general interatomic potential family, but it was not iterated as an independent product line. Community work combining GNoME data with more recent potentials has since appeared, including a study that screens novel cathode materials from the Energy-GNoME database using MACE force fields together with DFT.
Against the Materials Project route, the relationship is not one of replacement. Materials Project provides a high-quality archive of structures and properties built on experiment and computation, and the historical base of stable crystal entries near 48,000 comes mainly from the accumulated work of MP, OQMD, and related efforts. GNoME uses snapshots of those archives as training starting points and hull references, then expands outward with models and active learning. The paper states that GNoME displaces at least 5,000 stable materials from Materials Project and OQMD, so the two partially take over each other's entries.
On the parallel track of generative materials design, the 2025 Nature paper "A generative model for inorganic materials design" represents another class of method that generates structures or compositions directly and then hands them to first-principles verification. GNoME generates candidates by search and substitution, with structural diversity coming from SAPS and random structure search rather than from a learned generative distribution. That difference sets the strengths and limits: GNoME is strong at systematic scanning within the neighborhood already covered by chemical intuition and weaker at structures that break away from it.
6. Benchmarks and Evidence
Verifiable quantitative results concentrate on model accuracy and filtering efficiency.
On accuracy, the initial structural model trained on the 2018 Materials Project snapshot of roughly 69,000 materials reached a mean absolute error of 21 meV per atom against a prior benchmark of 28 meV per atom, and the final model reached 11 meV per atom, with accuracy improving as a power law of data volume.
On filtering efficiency, the structural pipeline raised the hit rate for stable predictions above 80 percent, and the compositional pipeline alone reached 33 hits per 100 trials against 1 percent previously, with more than 10^9 candidates generated and only a fraction reaching DFT evaluation.
On structural diversity, prototype analysis with XtalFinder found that the discovered stable crystals correspond to more than 45,500 new prototypes, a 5.6-fold increase over the 8,000 of Materials Project, a growth the authors argue could not arise from full substitution or prototype enumeration alone.
On functional robustness, r2SCAN recomputation showed that 84 percent of discovered binary and ternary crystals retain negative phase-separation energies.
On experimental validation, 736 structures aggregated in the ICSD match structures obtained independently through GNoME, with six examples shown in the paper.
On applied outcomes, the paper reports 528 promising lithium-ion conductors found under a holistic screening approach, a 25-fold increase over the original study, and for the Li/Mn transition-metal oxide family considered as a replacement for LiCoO2, 15 additional stable candidates relative to Materials Project against the original nine.
On the potential branch, the general potential trained on relaxation trajectories was used for zero-shot ionic conductivity estimates, connecting structure discovery to transport property prediction. Several downstream databases and screening pipelines have since grown around that link. Energy-GNoME is a living database of materials selected for energy applications, published in Energy and AI in December 2025 and also available as arXiv:2411.10125. A study screening novel cathode materials using MACE force fields and DFT appeared in Journal of Physics: Materials in October 2026 and as arXiv:2511.22504. A computed thermoelectric feature database covering 50,992 GNoME materials was posted to ChemRxiv with a v1 in August 2026 and a v2 in September 2026. Further work searching for thermodynamically stable ambient-pressure superconducting hydrides in the GNoME database and an experimental investigation of the magnetic properties of MnFeCo4Si2 discovered by GNoME indicate that discovered entries have reached experimental verification.
Community re-analysis also surfaces disagreement, with a study from October 2026 specifically examining chemistry-resolved disagreement between Materials Project-compatible PBE(+U) and r2SCAN stability verdicts in the GNoME database. Such follow-up work cautions against treating GNoME entries as fully confirmed stable phases.
7. Role in Physics-AI Workflows
GNoME's place in a physics AI workflow can be understood on three levels, each corresponding to a different way of using it.
The first level is as a data source. Its convex-hull dataset is among the largest public collections for bulk crystal stability, covering broad chemical space with 381,000 new stable entries and more than 520,000 with the extended release, alongside structure files, energies, several decomposition energy conventions, dimensionality, and band gap. For work on crystal property prediction, structure generation, or preliminary assessment of defect and doping feasibility, the data can serve directly as training material or an evaluation benchmark.
The second level is as a methodological pattern. The loop of generation, learned filtering, first-principles verification, and feedback into training transfers to any problem where evaluation cost is high and the candidate space is enormous. Typical migrations include electrolyte and electrode interface screening for batteries, alloy composition and phase stability design, catalyst surface adsorption configuration search, and layered structure exploration in composites. The layered material and solid-electrolyte candidate screening reported in the paper, along with the 528 lithium-ion conductors, are direct products of this pattern.
The third level is as a source of potential training data. The energies and forces contained in relaxation trajectories are the raw material for training general interatomic potentials, and the paper trains a NequIP potential on this data for zero-shot ionic conductivity prediction, connecting structure discovery to property simulation. The practical value for engineering teams is that once a target system has stable structures and their relaxation trajectories, a potential for that system can be obtained in the same way, enabling long-timescale molecular dynamics.
Concrete CAE and simulation uses include several classes. Virtual screening can come first, filtering out thermodynamically implausible compositions before expensive multiphysics simulation to shrink the geometry and meshing effort that follows. System plausibility checks can query decomposition energy for a structure produced by another workflow to decide quickly whether it deserves compute resources. And the data can serve as a starting point for potential fine-tuning when the target system lies inside the chemical space GNoME covers.
When results enter formal technical decisions, it is worth presenting PBE and r2SCAN outcomes side by side, labeling experimentally realized and unrealized entries separately, and distinguishing entries within 1 meV per atom of the hull from those strictly on it. That presentation keeps the boundary of the method visible and lets readers judge the evidence strength behind each claim.
8. Limitations and Failure Modes
The failure modes of GNoME are as clear as its success factors, and listing them is more useful than a general caveat.
The first is stability misjudgment driven by functional dependence. The decomposition energy criterion is computed at the PBE level, and the r2SCAN recomputation finding 84 percent of binary and ternary structures still stable implies that roughly a sixth of cases are sensitive to the functional. For transition metal oxides, lanthanide-containing systems, or strongly correlated compounds, PBE(+U) and meta-GGA functionals can give systematically different stability rankings, and community work already quantifies this. Using GNoME hull distances directly in such systems is a known high-risk practice.
The second is the timeliness of the snapshot hull. The data correspond to a specific point in time for the Materials Project, OQMD, and WBM snapshots. As new crystals are discovered, some entries on an old hull lose their status as most stable. In the other direction, the 381,000 entries added by GNoME themselves get corrected by later calculations. Treating a GNoME decomposition energy as an absolute value accumulates risk over time.
The third is inconsistent pseudopotentials and conventions. Materials Project, OQMD, and GNoME use different pseudopotentials, most visibly for Ga, Ge, Li, Mg, and Na, and the dataset ships explicit correction values for exactly these. Concatenating energies without applying them places spurious points on the hull.
The fourth is the synthesizability gap. Thermodynamic stability and experimental synthesizability are separated by reaction pathways, kinetic windows, and process conditions. The 736 experimentally realized entries are a very small fraction of the 381,000 new stable entries. A substantial share of the remainder may be hard to prepare because of kinetic barriers or synthesis economics. Handing a convex-hull entry list to a synthesis team as a candidate material list mistakes a screening result for a validated conclusion.
The fifth is weaker stability for the extended entries. The August 2024 revision released all materials within 1 meV per atom of the hull, taking the count above 520,000. Those entries are by definition not on the hull, so their characterization as stable follows from a looser threshold and differs in kind from strict hull entries. The two should not be mixed in one candidate pool.
The sixth is extrapolation risk for the potential. The GNoME potential is trained on PBE-level relaxation trajectories over a limited set of ionic compounds with robustness experiments up to 1000 K. Using it on surfaces, defects, melts, polarizable systems, or heavy-element compositions is extrapolation, with performance unassessed, and structural problems arise for systems needing charge self-consistency or long-range electrostatics.
The seventh is that the dimensionality prediction is heuristic. The Dimensionality Cheon field comes from a 2017 method, supports layered material screening, and is an auxiliary signal rather than a substitute for direct structural analysis.
The eighth is a licensing boundary on usage. The dataset is CC BY-NC 4.0 and commercial use is restricted, while the code is Apache-2.0. An engineering team intending to introduce the data into a commercial project must first confirm the authorization path.
The ninth is physics GNoME does not cover. Magnetism, elasticity, mechanics, thermal transport coefficients, phase transition kinetics, and interfacial phenomena are all outside scope. A complete materials assessment is covered only in its thermodynamic stability segment.
9. Version and Source Notes
Several conventions should be checked one by one before citing this work, since mismatched numbers are common.
The three scale figures are precise. The candidate pool is more than 10^9 generated and filtered structures across active learning. The structural hit count is 2.2 million structures below the current convex hull. The new stable entries number 381,000, bringing stable crystals to 421,000 against a prior baseline of 48,000. A statement that 2.2 million new crystal structures were predicted needs correction, since 2.2 million is the total below the hull while only 381,000 are newly added stable entries, a different order of magnitude.
The count of experimentally synthesized entries needs correction from hundreds to 736, which is the figure given in both the abstract and the body of the paper.
The dataset size differs by version. The paper published 381,000 new stable entries. The official README states that as of August 2024 the dataset was expanded to release all materials within 1 meV per atom of the hull, for a total above 520,000. When citing dataset size, note the version date.
The version history is explicit. The initial release was on November 29, 2023. On December 1, 2023 the paper filters were re-introduced to remove unphysical energies and two missing columns, Dimensionality Cheon and Is Train, were added. On August 21, 2024 the threshold was adjusted to 1 meV per atom off the hull, significantly increasing released crystals, the require_bound version of MP corrections was adopted, structures were re-relaxed where VASP had not updated lattice coordinates, convex-hull entries from MP and OQMD recomputations were provided, colabs for access were added, and file naming in by_id.zip was fixed. On November 21, 2024 formation energies were made consistent and the elemental references and pseudopotential corrections used in the GNoME paper were supplied.
Two threshold numbers apply to the hull. The dataset documentation states that 5e-5 eV is the threshold for determining whether a material is on the convex hull, due to numerical precision and simulation error, while the release expansion uses the looser 1 meV per atom criterion. These answer two different questions and conflating them inflates the stable count.
The repository state is fixed. The google-deepmind/materials_discovery repository was created on November 28, 2023, the day before publication, with default branch main, an Apache-2.0 license, roughly 1,251 stars and 197 forks at retrieval, and no releases or tags. The README states that this is a research project and not an official Google product.
Publication and license details are as follows. The paper was received on May 8, 2023, accepted on October 10, and published on November 29, with authors at Google DeepMind and Google Research, and is open access under CC BY 4.0. The dataset also has a Materials Project publication channel under permanent link DOI 10.17188/2009989.
The potential branch has its own positioning. The GNoME potential is trained on relaxation trajectories using the NequIP architecture, which is a different product form from the general foundation potential route. When assessing its value as a foundation model, first fix the comparison in terms of data provenance and training objective.
10. References
- GNoME official repository and dataset descriptor. https://github.com/google-deepmind/materials_discovery
- Merchant A., Batzner S., Schoenholz S.S., Aykol M., Cheon G., Cubuk E.D. Scaling deep learning for materials discovery. Nature, 624, 80-85, 2023. DOI: https://doi.org/10.1038/s41586-023-06735-9
- GNoME dataset release page and permanent link at Materials Project. https://materialsproject.org/gnome , DOI: https://doi.org/10.17188/2009989
- Public cloud bucket layout for the GNoME dataset. gs://gdm_materials_discovery
- Szymanski N.J. et al. An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature, 624, 86-91, 2023. DOI: https://doi.org/10.1038/s41586-023-06734-w
- Xie Y., Grossman J.C. Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties. Physical Review Letters, 120, 145301, 2018. DOI: https://doi.org/10.1103/PhysRevLett.120.145301
- Batzner S. et al. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13, 2453, 2022. DOI: https://doi.org/10.1038/s41467-022-29939-5
- Furness J.W., Kaplan A.D., Ning J., Perdew J.P., Sun J. Accurate and numerically efficient r2SCAN meta-generalized gradient approximation. The Journal of Physical Chemistry Letters, 11, 8208-8215, 2020. DOI: https://doi.org/10.1021/acs.jpclett.0c02405
- Jain A. et al. Commentary: The Materials Project: a materials genome approach to accelerating materials innovation. APL Materials, 1, 011002, 2013. DOI: https://doi.org/10.1063/1.4812323
- Kingsbury R. et al. Performance comparison of r2SCAN and SCAN metaGGA density functionals for solid materials via an automated, high-throughput computational workflow. Physical Review Materials, 6, 013801, 2022. DOI: https://doi.org/10.1103/PhysRevMaterials.6.013801
- Energy-GNoME: A living database of selected materials for energy applications. Energy and AI, 2025. DOI: https://doi.org/10.1016/j.egyai.2025.100605
- Screening novel cathode materials from the Energy-GNoME database using foundation machine learning force fields and DFT. Journal of Physics: Materials, 2026. DOI: https://doi.org/10.1088/2515-7639/aeae74
- A computed thermoelectric feature database for 50,992 GNoME materials. ChemRxiv, v1 2026-08-25, v2 2026-09-08. DOI: https://doi.org/10.26434/chemrxiv.15007873/v1
- Search for thermodynamically stable ambient-pressure superconducting hydrides in GNoME database. arXiv:2508.19781, 2025. DOI: https://doi.org/10.48550/arXiv.2508.19781
- Experimental investigation of magnetic properties of MnFeCo4Si2 discovered by GNoME. arXiv:2603.22748, 2026. DOI: https://doi.org/10.48550/arXiv.2603.22748
- Chemistry-resolved disagreement between Materials Project-compatible PBE(+U) and r2SCAN stability verdicts in the GNoME database. arXiv:2610.33101, 2026. DOI: https://doi.org/10.48550/arXiv.2610.33101
- Universal machine learning interatomic potentials are ready for phonons. npj Computational Materials, 11, 2025. DOI: https://doi.org/10.1038/s41524-025-01650-1
- Benchmarking universal machine learning interatomic potentials on elemental systems. npj Computational Materials, 12, 2026. DOI: https://doi.org/10.1038/s41524-026-02251-2
- Hicks D. et al. AFLOW-XtalFinder: a reliable choice to identify crystalline prototypes. npj Computational Materials, 7, 30, 2021. DOI: https://doi.org/10.1038/s41524-020-00483-4
- Chen Y. et al. A generative model for inorganic materials design. Nature, 639, 624-632, 2025. DOI: https://doi.org/10.1038/s41586-025-08628-5
- Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery. Chemistry of Materials, 2024. DOI: https://doi.org/10.1021/acs.chemmater.4c00643